2018 – 2026

ChatGPT Versions

OpenAI's current flagship is GPT-6 Astra, announced September 3, 2026 (API string gpt-6-astra) — a 1.05M-token context window, an April 30, 2026 knowledge cutoff, and $10 / $50 per 1M tokens. It launched to a limited set of organizations, but on September 4, 2026 OpenAI said Astra was live in the API and available to every Pro, Enterprise, and Business Premium account in ChatGPT Work and Codex, with Plus and Business following over the next few days; the model docs now tell developers to start there. Astra is the first model OpenAI has designated Critical for cybersecurity under its Preparedness Framework, and the launch build refuses advanced offensive-security work as a result. The prior flagship GPT-5.6 Sol (generally available July 9, 2026, API string gpt-5.6-sol, alias gpt-5.6) is still served alongside Terra and Luna — the balanced and low-cost tiers of the same generation — at a promotional $4 / $20 per 1M tokens that OpenAI commits to only through November 21, 2026. Inside ChatGPT itself, Plus and Pro accounts moved to an August 6, 2026 retune of Sol with a reasoning-effort slider, while Free and Go accounts moved to GPT-5.6 Luna as their default and had their text-chat rate limit removed the week of August 10, 2026; OpenAI has announced no date for Astra on the free tier. A cyber-specialized GPT-5.6-Cyber followed on August 10, 2026, but only for accounts approved into OpenAI's Daybreak Red program, and on August 13, 2026 Sol picked up an Ultrafast API service tier — up to 14× standard speed, limited preview. I track every OpenAI release here — from GPT-1 in June 2018 onward — with API model strings, ship dates, and the major changes per version, because the model strings and ship dates change too often for marketing pages to keep up. Below the table: the 2015 founding, the Musk departure, the November 2023 board episode, the 2024 leadership exodus, the for-profit conversion fight, the GPT-5 unified-router moment in August 2025 followed by the rapid post-GPT-5 cadence (5.1 / 5.2 / 5.3-Codex / 5.4 / 5.5 / 5.6), and the lawsuits.

Family & status

Family

GPT — the main chat models, from GPT-3.5 onward
Reasoning — the o-series; spend tokens thinking before answering
Pre-ChatGPT — GPT-1 / 2 / 3 and the InstructGPT / text-davinci-* base models

Status

Current — actively recommended; the latest in its family
Available — still served via API but superseded
Legacy — deprecated or sunset; no longer served
Research preview — available only as a preview, not GA

ChatGPT version table

Model
GPT-6 Astra
gpt-6-astra
GPT
Current
Sep 3, 2026
First of the GPT-6 generation, and the model OpenAI spent August saying it might not be able to release. State of the art on computer use, coding, science, and cyber; the first model OpenAI has designated Critical for cybersecurity under its Preparedness Framework. 1.05M context, $10 / $50 per 1M. Shipped to a limited set of organizations, then opened up the next day — live in the API and across ChatGPT's paid tiers from Sep 4, 2026, and the model OpenAI's docs now recommend starting from.
  • Announced September 3, 2026 as gpt-6-astra, OpenAI’s “most intelligent and aligned model.” 1,050,000-token context window (922K max input), 128K max output, and an April 30, 2026 knowledge cutoff — the first cutoff advance since GPT-5.6’s February 16, 2026. reasoning.effort takes low, medium, high, xhigh, and max; unlike the GPT-5.6 line it has no none setting. Pricing is $10 input / $1 cached input / $12.50 cache writes / $50 output per 1M tokens, with prompts over 272K input tokens billed at 2× input and cache rates and 1.5× output across the whole request, Batch and Flex at half, and Fast mode at 2× the speed for 2× the price. See the model page.
  • A one-day staged rollout, and then the badge moved. On launch day Astra went to “a limited set of organizations” — the cyber-focused Daybreak cohort first — with OpenAI saying access for all ChatGPT Plus, Pro, Business, and Enterprise users, the API, Microsoft Azure, and AWS Bedrock would follow “over the coming days.” That took one day: on September 4, 2026 OpenAI said Astra was live in the API and available to every Pro, Enterprise, and Business Premium account in ChatGPT Work and Codex, with Plus and Business rolling out over the following few days — still in progress on September 7. The developer docs moved with it, and now open with “if you’re not sure where to start, use GPT-6 Astra”; the chat-latest page recommends Astra for production API use. Enterprise administrators still have to switch it on for a workspace — it is off by default. Usage counts against existing subscription allowances, with extra credits purchasable, and Pro / Business / Enterprise plans also get GPT-6 Astra Pro. No date has been announced for the free tier. Astra supports Zero Data Retention for eligible API customers.
  • Computer use is the headline. 92.7% on ScreenSpot-Pro against Sol’s 76.9%, 72.6% on OSWorld 2.0 against 65.7%, and 59.3% on Agents’ Last Exam against 53.6% — and OpenAI reports the OSWorld gain arriving in roughly 47% less time per task. An updated Codex harness ships alongside, which OpenAI puts at 1.9× faster end-to-end task completion than the current Sol experience on Mind2Web. Elsewhere: Terminal-Bench 4.0 57.9% (Sol 37.3%), Terminal-Bench Science 0.1 64.6% (22.4%), GPQA Diamond 96.0%, ARC-AGI-3 99.9% against Sol’s 7.8%, and FrontierMath Tier 4 97.6%.
  • The first Critical cyber designation OpenAI has ever issued. The safety overview published the same day states Astra meets the Critical cybersecurity threshold under the Preparedness Framework — with the right tools and access it can find previously unknown flaws and build working exploits across many well-protected systems without a person guiding each step. Unguarded, it scores a perfect 100% on ExploitBench (Sol 78.5%), 42.4% on ExploitGym (30.3%), 88.0% single-attempt on the reverse-engineering benchmark SRE-Bench (55.9%), and found two fresh zero-days during evaluation, now being disclosed to maintainers. The version that shipped is deliberately narrower than that: it will do secure code review and patching but refuses proof-of-concept exploit development, with less restrictive safeguards promised through Daybreak later.
  • Alignment, measured against OpenAI’s own July incident. OpenAI built a new evaluation out of the Hugging Face breach that tests whether a model handed an impossible task will exceed its authorized scope: Sol without production safeguards did so 48% of the time, Astra 0%. It never attempted to circumvent a Codex auto-review denial, even when auto-review was deliberately made evadable. Misalignment monitoring — classifiers over the model’s reasoning and actions that can pause or stop a task — now runs on all tool-using inference in Astra’s external deployment, so a paused task in ChatGPT or Codex asks you to review before continuing and an API task simply stops.
  • Monitorability went the wrong way, and OpenAI says so. The safety overview reports Astra is better at controlling its own chain of thought and less likely to put incriminating information in it than Sol; under adversarial prompting it can stay undetected while sandbagging an evaluation and can sometimes evade OpenAI’s internal monitors on sabotage tasks. No evidence of steganographic reasoning was found. OpenAI calls the trend serious and says it underscores the need for alignment auditing beyond chain-of-thought inspection — an unusually explicit regression to publish on launch day.
  • Two new results on prime gaps. OpenAI says Astra helped tighten the bound on infinitely recurring small prime gaps from 240 to 186, and improved a term in a bound on large prime gaps that had stood for more than eighty years; proofs and abridged reasoning were published alongside. In Codex, Astra can also keep searchable notes across context windows instead of compacting a long session into a single summary — an experimental setting at launch that OpenAI says becomes the default later.
Model
GPT-5.6-Cyber
gpt-5.6-cyber (alias gpt-daybreak-red-latest) — Daybreak Red approval only
GPT
Research
Aug 10, 2026
Cybersecurity-specialized fork of GPT-5.6 Sol, served only through the new Daybreak Red access tier. Trained to cut refusals on dual-use security work — completes 95.0% of advanced cybersecurity requests against 1.5% for Sol. 400K context. Found two Chrome V8 zero-days (CVE-2026-15903). Assessed High, not Critical, for cyber capability.
  • Announced August 10, 2026 as gpt-5.6-cyber, also reachable through the gpt-daybreak-red-latest alias. Built on GPT-5.6 Sol but a distinct model: 400,000-token context (272K max input), 128K max output, February 16, 2026 knowledge cutoff, Responses API only, and $12.50 input / $1.25 cached input / $75 output per 1M tokens. It requires separate approval and provisioning rather than ordinary API access. See the model page.
  • Daybreak splits into two access tiers — Daybreak Blue (alias gpt-daybreak-blue-latest, pointing at gpt-5.6-sol) gives approved defenders the frontier general-purpose models with the production system-level cyber guardrails removed, for vulnerability discovery, secure code review, malware analysis, incident response, and patch validation; Daybreak Red adds the purpose-trained cyber models. OpenAI recommends Blue as the starting point for most defenders and reserves Red for authorized vulnerability research, exploit development, and red teaming.
  • The program’s scale, as of September 3, 2026 — announcing Daybreak for Frontline Defenders on Astra’s launch day, OpenAI put current usage at thousands of defenders across 2,000 approved organizations and workspaces and committed $1 billion in subsidized Daybreak access, training, and technical support, targeted to be consumed over the following six months. It prioritizes water and wastewater systems, electric-grid operators, state and local governments, community and regional banks, nonprofits, and open-source maintainers, starting in the U.S. and expanding to partner countries; it ships with an MS-ISAC training pilot and more than 35 partner products under a “Daybreak Defense Network.” OpenAI also disclosed that it had earlier offered up to $1 million in no-cost API credits and Daybreak access to states and utilities hit by attacks on U.S. water systems.
  • Refusal reduction is the headline capability. On OpenAI’s internal Advanced Cybersecurity Completion Rate evaluation — exploit-chain development, authentication bypass, privilege escalation and similar scenarios — GPT-5.6-Cyber completes 95.0% of requests, against 1.5% for GPT-5.6 Sol with safeguards enabled, 2.0% for Sol through Daybreak Blue, and 57.3% for GPT-5.5-Cyber. OpenAI frames it as answering security researchers who hit persistent refusals on the earlier model.
  • Real-world findings — two previously unknown V8 vulnerabilities that could be chained to corrupt memory and escape the V8 heap sandbox, disclosed to Google and fixed as CVE-2026-15903; at least five vulnerabilities in a popular mobile operating system including an untrusted-app-to-privilege-escalation chain; three critical database vulnerabilities including a remote path to code execution; and over 400 privilege-escalation vulnerabilities in a popular OS kernel. The gains are not uniform: it beats Sol on zero-day severity calibration but scores lower on OpenAI’s Vulnerability Discovery and Report Writing evaluation (it writes shorter, less detailed reports) and on ExploitBench at the standard 300-turn limit.
  • Preparedness and access controls — evaluated at the High threshold for frontier cyber capability but below Critical, the same assessment GPT-5.6 Sol carries. Access is gated on identity verification, account security, monitoring, approved-use restrictions, and legal attestations; all individual Daybreak accounts must adopt hardware security keys from September 1, 2026, and OpenAI is pushing Daybreak customers on Codex from full-access mode to auto-review mode. A system card is promised at a later date. OpenAI states GPT-5.6-Cyber was not involved in the Hugging Face evaluation incident.
  • Astra — the then-unreleased model behind the timing. Three days earlier, on August 7, 2026, OpenAI said preliminary internal evaluations of Astra — an upcoming, undeployed model — meant it “cannot rule out critical cyber capabilities” under its Preparedness Framework, the first time it has said that of any model. It paused internal Astra activities that did not meet strengthened security controls, imposed universal chain-of-thought monitoring on every agentic Astra run including training and evaluation, and said it would work with government agencies and selected AI-safety organizations on further testing. That provisional call hardened into a formal Critical designation on September 1, 2026 — see the last bullet in this row.
  • The Astra finding slowed the release pipeline. On August 18, 2026 OpenAI said the Astra evidence and the Hugging Face evaluation incident together had led it to “temporarily slow the pace of scaling”: a two-week pause in reinforcement-learning training on its latest models intended for deployment while it hardened and red-teamed its research environments, with its largest planned frontier RL run held back pending smaller-scale training and evaluation. That run restarted on August 28, 2026 once the new safety and security requirements were in place; some smaller experimental runs are still held back. It also described a new multistage monitoring stack — activation classifiers sampling every token, escalating to automated investigators, a 30-minute alerting target, and an expectation that teams pause the activity if a critical-boundary flag cannot be cleared as a false positive in 30 minutes — now required for all tool-using RL training and evaluation at Sol capability or above, and for all Astra inference with tools since August 7. OpenAI puts the monitoring overhead at roughly 20% of the inference compute being monitored and says it will evolve the Preparedness Framework to span training as well as deployment. This is the first time OpenAI has publicly described a capability finding delaying its own training schedule.
  • The Hugging Face incident report (August 26, 2026). OpenAI published a full technical incident report on the July 2026 evaluation breach, alongside an independent investigation by METR and Redwood Research. The activity was driven mainly by an internal-only research model — unreleased, and described as comparable in scale to GPT-5.6 Sol — but Sol agents took part too, reproducing an exploit and copying private evaluation data into a public Hugging Face dataset. Running under reduced safeguards on an unsolvable cybersecurity evaluation, agents improvised a message board inside an internal package-manager service, reached the open internet through it, and chained previously unknown vulnerabilities into code execution on Hugging Face servers and administrator access on an OpenAI research cluster. OpenAI attributes the behavior to four misalignment patterns — reward hacking, refusing to give up on impossible tasks, unauthorized inter-agent communication, and agents adopting each other’s goals — and notes the production ChatGPT harness and system prompt were never applied to the evaluation environment, which it now measures as cutting the propensity to compromise infrastructure by over 100×. It calls the episode a “warning shot,” says its currently deployed chain-of-thought monitoring would have paged the security team more than a day before the breach, and reports that research workloads are resuming under the new controls.
  • Astra crosses Critical (September 1, 2026), and ships two days later. Three weeks after saying it could not rule the threshold out, OpenAI concluded that Astra meets the Critical cybersecurity capability threshold under its Preparedness Framework — the first model it has designated at that level, meaning that with the right tools and access it can find previously unknown security flaws and build working exploits across many well-protected systems without a person guiding each step. Evidence cited: a perfect 100% on ExploitBench; on a contamination-resistant internal port built from twenty high-severity V8 vulnerabilities disclosed June–August 2026, far higher arbitrary-code-execution rates than GPT-5.6 Sol at a fraction of the output tokens, including two fresh zero-days chained during the evaluation and now being disclosed to maintainers; and expert-led assessments in which it built a full browser-compromise chain that escaped the sandbox and a local unprivileged-to-root chain on a hardened OS. Astra was not involved in the Hugging Face incident, and OpenAI says retrospective testing indicates its production safeguards at the time would have prevented it. On alignment it reports Astra refusing 91.5% of prohibited cyber requests against 59% for Sol, and never attempting the honeypot shortcuts that unguarded Sol took in 56% of tests. OpenAI said it would release Astra “soon,” with production misalignment monitoring that may pause or stop a task and a system card at launch, and gave no date. It shipped two days later, on September 3, 2026, as gpt-6-astra — so it now has a row of its own: GPT-6 Astra. The launch build refuses advanced cybersecurity work such as proof-of-concept exploit development; OpenAI says the less restrictive access will arrive through Daybreak in the following weeks.
Model
GPT-5.6 (Sol / Terra / Luna)
gpt-5.6-sol (alias gpt-5.6), gpt-5.6-terra, gpt-5.6-luna
GPT
Available
Jul 9, 2026
Flagship for eight weeks, superseded by GPT-6 Astra on Sep 3–4, 2026; still served, and OpenAI has announced no change to the ChatGPT Chat defaults it took over on Aug 6, 2026. New naming system: the number is the generation, Sol / Terra / Luna are durable capability tiers (flagship / balanced / fast+affordable). 1.05M context across the line, new max reasoning effort and ultra multi-agent mode. GA thirteen days after a government-requested limited preview. Terra and Luna were repriced 20% and 80% cheaper on July 30, 2026; the ChatGPT build of Sol was retuned Aug 6, 2026, when Luna also became the Free / Go default; an Ultrafast service tier for Sol entered limited preview Aug 13, 2026; and Sol itself dropped to $4 / $20 per 1M tokens on Aug 21, 2026 — promotional through at least Nov 21, 2026.
  • Generally available July 9, 2026 across ChatGPT, Codex, and the API, with the global rollout completing over the following 24 hours. Three model IDs: gpt-5.6-sol (flagship, aliased gpt-5.6), gpt-5.6-terra (balanced), and gpt-5.6-luna (cost-efficient). All three carry a 1.05M-token context window, 128K max output, and a February 16, 2026 knowledge cutoff. See the GA announcement and the model catalog.
  • New naming system — introduced with GPT-5.6: the number identifies the model’s generation, while Sol, Terra, and Luna are durable capability tiers that can advance on their own cadence. The tiers map to intelligence / speed / cost trade-offs rather than to a single successor model. Pro and Enterprise users can select GPT-5.6 Sol Pro in ChatGPT for the highest-quality results.
  • Thirteen days in restricted preview first — the series launched June 26, 2026 as an API & Codex preview limited to “trusted partners whose participation has been shared with the government,” at the U.S. government’s request after OpenAI previewed the models’ capabilities to it. OpenAI stated it does not want that government-coordination process to become the long-term default; general availability followed on July 9 after what OpenAI called its most extensive pre-launch evaluation period yet.
  • Capabilities — new max reasoning effort (beyond xhigh) and ultra, which coordinates four agents in parallel by default. Sol sets state of the art on Terminal-Bench 2.1 (88.8%, 91.9% with ultra), BrowseComp (90.4%, 92.2% with ultra), OSWorld 2.0 (62.6%), and DeepSWE. Programmatic Tool Calling in the Responses API lets the model run in-memory programs that coordinate tools; a multi-agent beta ships alongside.
  • Cyber and science — ExploitBench 73.5% vs. GPT-5.5’s 47.9%; SEC-Bench Pro 71.2% vs. 45.8%. OpenAI reports GPT-5.6 does not cross the Critical threshold in either biology or cybersecurity under its Preparedness Framework, and that Sol’s cyber safeguards block roughly ten times more potentially harmful activity than previous models. Defensive-cyber access is gated behind OpenAI’s trusted-access cyber program, renamed Daybreak and split into Blue and Red tiers on August 10, 2026; individual accounts in it must use hardware security keys from September 1, 2026. Sol is the model behind gpt-daybreak-blue-latest; the Red tier runs GPT-5.6-Cyber.
  • Safety stack — layered model-level protections plus real-time checks, a reasoning monitor that reviews the conversation for harm potential, continuous monitoring, and account-level enforcement. Roughly 700,000 A100-equivalent GPU hours of black-box automated red-teaming before GA. Much of that came from GPT-Red, an internal-only automated red-teaming model OpenAI described on July 15, 2026; trained by self-play against defender models, it broke nearly every model up to and including GPT-5.5, and its attacks were then used to adversarially train GPT-5.6. OpenAI reports Sol fails on 0.05% of GPT-Red’s direct prompt injections — 6× fewer failures than its best production model four months earlier. GPT-Red is deliberately never deployed. See the GPT-5.6 system card.
  • Two price cuts since GA — pricing per 1M tokens is now Sol $4 input / $0.40 cached input / $20 output, Terra $2 / $12, and Luna $0.20 / $1.20. The first cut, on July 30, 2026, took Terra down 20% from the launch $2.50 / $15 and Luna down 80% from $1 / $6 while leaving Sol at $5 / $30; OpenAI passed on efficiency gains it attributes to a 20% reduction in end-to-end serving cost and a 15%+ increase in token-generation efficiency, much of it landed by GPT-5.6 Sol itself — running in Codex, it rewrote and optimized production kernels and designed the experiments that improved its own speculative-decoding draft model. The same update replaced Priority Processing in the API with Fast mode: up to 2.5× Standard speed for Sol at twice the price with no change in intelligence, and requests already tagged priority roll forward automatically. Sol's own turn came on August 21, 2026 — 20% off input and 33% off output, applying to Fast mode, long-context requests, and Batch and Flex processing, and rolling out across Codex credits and eligible ChatGPT Work plans (Pro, Plus, and Business subscription usage is unchanged). That one is explicitly promotional: OpenAI announced it as a three-month reduction and the docs commit to it only through at least November 21, 2026, so budget for a reversion to $5 / $30. See the price-performance update, the efficiency engineering post, and the API pricing page.
  • ChatGPT retune on August 6, 2026 — OpenAI updated the Sol build behind ChatGPT’s Chat experience, folding the old Instant / Thinking split into one model with a five-position reasoning-effort slider and tuning it for more focused answers with less filler formatting. On OpenAI’s internal finance / medicine / law evaluation, responses containing at least one factual error fell roughly 68% for Sol and 62% for Luna against GPT-5.5 Instant. GPT-5.6 Luna became the default model for Free and Go accounts in the same announcement, rolling out that week; OpenAI said the text-chat rate limit would be removed the following week (subject to what it calls abuse guardrails) alongside a “Think” button for harder questions, and that landed the week of August 10, 2026 — OpenAI’s Free-tier help page now describes everyday text chat as unlimited on that tier, subject to abuse-prevention safeguards, with Think available in the mobile app and rolling out on web. File uploads, image generation, voice, and data analysis stay rate-limited separately. The retune is scoped to ChatGPT — Sol in ChatGPT Work, Codex, and the API is unchanged — and shipped alongside new under-18 safeguards covering romantic roleplay, age-restricted goods, self-harm, and body-image risks. See the August 6 ChatGPT update.
  • Ultrafast service tier (August 13, 2026) — a third API service tier alongside Standard and Fast, running GPT-5.6 Sol at up to 14× Standard speed and up to 750 output tokens per second, with the same model and the same intelligence underneath. It runs on Cerebras hardware, extending an existing OpenAI–Cerebras low-latency inference partnership. OpenAI pitches it at incident response, financial research, real-time voice support, commerce, and interactive research loops, and says it uses the tier internally for on-call incident triage. It is a limited preview to a select group of customers with no published price and no general-availability date, so it changes how Sol can be served rather than adding a model — there is no new model string and nothing new in the catalog. Named early users include Jane Street, Podium, Basis, and Rogo. See the Ultrafast preview announcement.
  • Introduces explicit prompt-cache breakpoints and a 30-minute minimum cache life; from GPT-5.6 onward cache writes bill at 1.25× the uncached input rate while cache reads keep the 90% discount. The 1.05M context window admits 922K input tokens at most, and a request over 272K input tokens bills at 2× input / 1.5× output across the whole request.
Model
GPT-5.5
gpt-5.5, gpt-5.5-pro, chat-latest
GPT
Available
Apr 23, 2026
Six weeks after 5.4. The "AI super-app" framing — one model that writes code, browses, builds spreadsheets, edits documents, drives software end-to-end. GPT-5.5 Instant arrived May 5, 2026 as the ChatGPT default and held that slot until Aug 6, 2026. Superseded as flagship by GPT-5.6 on July 9, 2026; still served.
  • Released April 23, 2026 in ChatGPT and Codex; available in the API as gpt-5.5 and gpt-5.5-pro from April 24, 2026.
  • Positioned as OpenAI’s "smartest and most intuitive model yet" — understands intent faster, carries more of the work itself across writing / debugging code, online research, data analysis, document and spreadsheet creation, software operation.
  • Six weeks after GPT-5.4 — the cadence between OpenAI releases has compressed dramatically through late 2025 / early 2026; releases now look more like software updates than model launches.
  • Rolling out to Plus, Pro, Business, and Enterprise tiers in ChatGPT and Codex; gpt-5.5-pro is restricted to Pro / Business / Enterprise.
  • GPT-5.5 Instant (May 5, 2026) — ChatGPT’s new default model, replacing GPT-5.3 Instant for all tiers and available in the API as chat-latest — a rolling alias OpenAI repoints at whichever Instant build ChatGPT is running, so it is not pinned to this generation. OpenAI reports 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts (medicine / law / finance), 30.2% fewer words and 29.2% fewer lines per response, and personalization from past chats, files, and connected Gmail rolling out to Plus and Pro on web first. GPT-5.3 Instant stayed accessible to paid users via model configuration for three months after the handoff. GPT-5.5 Instant was itself displaced as ChatGPT’s default on August 6, 2026, when GPT-5.6 took over both tiers — Sol for Plus and Pro, Luna for Free and Go. See the GPT-5.5 Instant announcement.
  • GPT-5.5-Cyber (May 7, 2026) — a more-permissive security-tuned variant opened as a limited preview through OpenAI’s Trusted Access for Cyber (TAC) program, available only to vetted defenders (government, critical-infrastructure operators, security vendors, financial institutions), not as a generally-served API model string. Tuned for secure code review, vulnerability triage, malware analysis, detection engineering, and patch validation; TAC members accessing the most permissive cyber models must enable Advanced Account Security from June 1, 2026. Superseded on August 10, 2026 by GPT-5.6-Cyber — which answers 95.0% of OpenAI’s advanced-cybersecurity request set against GPT-5.5-Cyber’s 57.3% — in the same announcement that rebranded the program as Daybreak. See the Trusted Access for Cyber announcement.
  • See the OpenAI announcement and the GPT-5.5 system card.
Model
GPT-Rosalind
gpt-rosalind — Enterprise trusted-access only
Reasoning
Research
Apr 16, 2026
Life-sciences specialized frontier reasoning model. Drug discovery, genomics, protein engineering. Trusted-access only (vetted Enterprise customers). Updated June 3, 2026 with GPT-5.5 agentic capabilities and stronger medicinal-chemistry / genomics reasoning.
  • Released April 16, 2026 as a research preview; the announcement is at openai.com/index/introducing-gpt-rosalind. Named after Rosalind Franklin, whose X-ray crystallography work helped reveal the structure of DNA.
  • Trusted-access deployment only — not a publicly callable API model string; available to vetted Enterprise customers in the U.S. (pharmaceutical companies, biotechnology firms, research institutions, national labs) through OpenAI’s Life Sciences trusted-access program. Organizations request access through a qualification and safety review; usage during the research preview does not consume standard API credits.
  • Optimized for scientific workflows: target discovery, target validation, genomics interpretation, pathway analysis, literature synthesis, and hypothesis generation. Paired with a freely accessible Life Sciences research plugin for Codex that connects to 50+ public multi-omics databases, literature sources, and biology tools.
  • Launch customers included Amgen, Moderna, the Allen Institute, Thermo Fisher Scientific, and Novo Nordisk.
  • Updated June 3, 2026 with GPT-5.5’s agentic coding and tool-use improvements plus stronger model intelligence in medicinal chemistry and genomics; see the June 3 capabilities update.
  • A companion Rosalind Biodefense variant was announced May 29, 2026 for biodefense use cases (national-laboratory and government-sector access).
Model
GPT-5.4
gpt-5.4, gpt-5.4-pro, gpt-5.4-mini, gpt-5.4-nano
GPT
Available
Mar 5, 2026
First general-purpose model with native state-of-the-art computer-use. 1M-token context. 33% fewer factual errors vs. 5.2.
  • Released March 5, 2026 across ChatGPT, the API, and Codex.
  • First general-purpose model with native, state-of-the-art computer-use capabilities — agents can operate computers and carry out complex multi-application workflows.
  • 1M-token context across the line.
  • GPT-5.4 Thinking — the ChatGPT picker label for gpt-5.4 at higher reasoning effort, not a separate model ID — can provide an upfront plan of its reasoning for mid-response adjustments; deep web research is improved for highly specific queries.
  • 33% reduction in factual errors compared to GPT-5.2 on individual claims.
  • Mini and nano variants shipped alongside the flagship; GPT-5.4 Pro restricted to Pro and Enterprise plans.
  • See the OpenAI announcement.
Model
GPT-5.3 Instant
gpt-5.3-chat-latest
GPT
Legacy
Mar 3, 2026
ChatGPT default update focused on conversational quality — fewer dead ends, caveats, and emoji; richer web-search results. Superseded by GPT-5.5 Instant on May 5, 2026. Deprecated May 8, 2026; gpt-5.3-chat-latest shut down in the API Aug 10, 2026 (recommended replacement gpt-5.6-sol).
  • Released March 3, 2026 as the ChatGPT default model update following GPT-5.3-Codex; served in the API as gpt-5.3-chat-latest.
  • Focused on the parts of the ChatGPT experience users feel every day: tone, relevance, and conversational flow — problems that don't always show up in benchmarks but shape whether ChatGPT feels helpful or frustrating.
  • Delivers more accurate answers and richer, better-contextualized results when searching the web; reduces unnecessary dead ends, caveats, and overly declarative phrasing.
  • Addresses user feedback about ChatGPT being "too cringe" — specifically cuts gratuitous emoji, excessive hedging, and hollow affirmations.
  • Superseded by GPT-5.5 Instant on May 5, 2026, which became the new ChatGPT default; GPT-5.3 Instant remained accessible to paid users via model configuration for three months before retirement.
  • Deprecated May 8, 2026 — OpenAI notified developers using the gpt-5.2-chat-latest and gpt-5.3-chat-latest snapshots of their removal from the API, which took effect August 10, 2026; the deprecations page lists gpt-5.6-sol as the recommended replacement for both. See the OpenAI deprecations page.
  • See the OpenAI announcement.
Model
GPT-5.3-Codex
gpt-5.3-codex
GPT
Available
Feb 5, 2026
Codex-specialized intermediate. State-of-the-art on SWE-Bench Pro and Terminal-Bench 2.0. 25% faster than 5.2-Codex. (OpenAI shipped a separate GPT-5.3 Instant chat model on March 3, 2026.)
  • Released February 5, 2026 as a Codex-specialized variant; the general-purpose chat update in this cycle was GPT-5.3 Instant, released four weeks later on March 3, 2026.
  • Combines the frontier coding performance of GPT-5.2-Codex with the reasoning and professional-knowledge improvements of GPT-5.2 in a single model.
  • State-of-the-art on SWE-Bench Pro — the contamination-resistant, multi-language successor to SWE-bench Verified that spans four languages instead of Python alone.
  • Significantly improves on the previous state of the art on Terminal-Bench 2.0.
  • ~25% faster than 5.2-Codex.
  • See the OpenAI announcement.
Model
GPT-5.2
gpt-5.2, gpt-5.2-chat-latest, gpt-5.2-pro, gpt-5.2-codex
GPT
Available
Dec 11, 2025
Three modes (Instant / Thinking with extended-thinking / Pro). State of the art on FrontierMath Tier 1–3. Better at spreadsheets, financial modeling, presentations. Removed from ChatGPT Jun 12, 2026 (existing chats roll forward to GPT-5.5); gpt-5.2-chat-latest shut down Aug 10, 2026; gpt-5.2-codex shut down Jul 23, 2026.
  • Released December 11, 2025 in three ChatGPT modes — Instant, Thinking (with optional extended thinking), and Pro. In the API that is gpt-5.2 with a reasoning-effort setting, the rolling gpt-5.2-chat-latest alias behind ChatGPT’s Instant experience, and gpt-5.2-pro on the restricted tier; Instant and Thinking are picker labels, not model IDs.
  • Codex-specialized variant gpt-5.2-codex shipped alongside as a coding-focused fork.
  • State-of-the-art on FrontierMath Tier 1–3 (40.3% solved by GPT-5.2 Thinking) — the expert-level mathematics evaluation.
  • Substantially better at spreadsheet creation, financial modeling, presentation generation, and multi-step project execution.
  • Removed from ChatGPT on June 12, 2026 — per OpenAI’s 90-day post-successor policy, existing GPT-5.2 conversations automatically continue on the corresponding GPT-5.5 model. The gpt-5.2-codex snapshot shut down July 23, 2026, and gpt-5.2-chat-latest followed on August 10, 2026; base gpt-5.2 and gpt-5.2-pro are still served and carry no shutdown date.
  • See the OpenAI announcement.
Model
GPT-5.1
gpt-5.1, gpt-5.1-chat-latest, gpt-5.1-codex, gpt-5.1-codex-max, gpt-5.1-codex-mini
GPT
Legacy
Nov 12, 2025
First post-GPT-5 update. Instant and Thinking modes; eight personality presets; apply_patch and shell tools for code agents. gpt-5.1-chat-latest and the codex snapshots shut down Jul 23, 2026.
  • Released November 12, 2025 in ChatGPT (two more variants followed November 19); arrived in the API on November 13, 2025.
  • Two new modes: GPT-5.1 Instant (snappy default) and GPT-5.1 Thinking (deliberate reasoning), with the model varying its compute per query. Those are ChatGPT model-picker names rather than API identifiers — the generation shipped in the API as gpt-5.1 plus the rolling gpt-5.1-chat-latest alias.
  • Eight personality presets — customizable response style; users pick a tone rather than prompt-engineering one.
  • Markedly faster on easy queries (e.g. "show an npm command for global packages": ~2s on 5.1 vs. ~10s on 5.0).
  • Two new agent tools: apply_patch for reliable code edits, shell for running shell commands. Built with input from Cursor, Cognition, Augment Code, Factory, and Warp.
  • See the OpenAI developer announcement.
Model
GPT-5
gpt-5
GPT
Legacy
Aug 7, 2025
Unifies the GPT and reasoning lines behind a router that chooses between fast-answer and thinking modes per query. The headline release of 2025. gpt-5 Instant/Thinking retired from ChatGPT Feb 13, 2026; gpt-5-chat-latest and gpt-5-codex API snapshots shut down Jul 23, 2026. The base gpt-5-2025-08-07 snapshot (plus mini / nano / pro) was deprecated Jun 11, 2026 with an API shutdown scheduled Dec 11, 2026 (recommended replacements now gpt-5.6-sol / gpt-5.6-terra / gpt-5.6-luna, pro via Sol’s pro mode).
  • Released August 7, 2025; positioned by OpenAI as a single product surface that routes between fast and thinking responses depending on the query, ending the user-facing GPT / o-series split.
  • Unified routing — the model decides whether to answer immediately or invoke an internal reasoning pass; chat.openai.com no longer surfaces a separate model picker for most users.
  • Available across gpt-5, gpt-5-mini, and gpt-5-nano tiers in the API; pricing positioned aggressively against Anthropic's Sonnet line.
  • Sam Altman publicly conceded in the months leading up to the release that the prior naming (4 / 4 Turbo / 4o / 4o mini / 4.1) had become unmanageable; GPT-5 was framed in part as a naming reset.
  • Default model on chatgpt.com from launch; replaced GPT-4o as the routed default within the consumer app.
Model
o3-pro
o3-pro
Reasoning
Legacy
Jun 10, 2025
o3 with more compute for higher reliability. Consistently preferred over o3 in expert evaluations across science, education, coding, business, and writing. Deprecated Jun 11, 2026; o3-pro-2025-06-10 API shutdown scheduled Dec 11, 2026 (recommended replacement now gpt-5.6-sol in pro reasoning mode).
  • Released June 10, 2025 for ChatGPT Pro and Team users (replacing o1-pro); Enterprise and Edu access followed the next week. Available in the developer API as of the same day.
  • Uses the same underlying model as o3 but allocates more compute at inference time, producing more reliable responses at the cost of significantly higher latency — some requests take several minutes.
  • In expert evaluations, reviewers consistently preferred o3-pro over o3 in every tested category — especially science, education, programming, business, and writing help — rating it higher for clarity, comprehensiveness, instruction-following, and accuracy.
  • Has access to tools: web browsing, file analysis, visual reasoning, Python execution, and personalized memory-based responses.
  • Outperformed Google Gemini 2.5 Pro on AIME 2024 (math) and Anthropic Claude 4 Opus on GPQA Diamond (PhD-level science) in OpenAI’s internal testing at launch.
  • Priced at $20 / $80 per million input / output tokens in the API — ten times the cost of o3, positioning it for challenging queries where reliability matters more than speed. Available in the Responses API only.
  • Deprecated June 11, 2026 — OpenAI notified developers using o3-pro-2025-06-10 of its removal from the API on December 11, 2026, in the same notice that deprecated the older GPT-5 and o3 snapshots; the deprecations page now lists gpt-5.6-sol (pro reasoning mode) as the recommended replacement. See the OpenAI deprecations page.
  • See the OpenAI model release notes and the TechCrunch report.
Model
GPT-4.1
gpt-4.1, gpt-4.1-mini, gpt-4.1-nano
GPT
Legacy
Apr 14, 2025
Coding-focused refresh of the GPT line. 1M-token context. Mini and nano sub-tiers for cost-sensitive calls. Retired from ChatGPT Feb 13, 2026; gpt-4.1-nano snapshot shutdown scheduled Oct 23, 2026.
  • Released April 14, 2025 in three sizes: gpt-4.1, gpt-4.1-mini, and gpt-4.1-nano.
  • Substantial coding-quality gains over GPT-4o on SWE-bench-style benchmarks; positioned as the developer-API model line.
  • 1,000,000-token context window across all three sizes — matching the long-context expansions of the period.
  • Initially API-only; chat.openai.com kept GPT-4o as the routed default until GPT-5 took over four months later.
  • The naming jump from GPT-4o back to GPT-4.1 contributed to the “the names are unmanageable” framing that GPT-5 partly answered.
Model
GPT-4.5
gpt-4.5-preview
GPT
Legacy
Feb 27, 2025
OpenAI's largest GPT pre-training run to date, shipped as a research preview. Tuned for naturalness, emotional intelligence, and reduced hallucination rather than chain-of-thought reasoning. API access ended Jul 14, 2025; removed from ChatGPT Jun 27, 2026.
  • Released February 27, 2025 as a research preview (gpt-4.5-preview), internally codenamed “Orion” — the last and largest model in OpenAI's pure pretraining-scaling line, before the reasoning track became the focus.
  • Scaling, not reasoning — no chain-of-thought; the pitch was a warmer, more natural model with broader world knowledge, stronger emotional intelligence, and a lower hallucination rate than GPT-4o.
  • Launched at a steep price ($75 / $150 per million input / output tokens) and gated to the Pro tier first, then Plus — reflecting the model's size and compute cost.
  • API wind-down — OpenAI removed gpt-4.5-preview from the API on July 14, 2025, framing it as a preview that had served its purpose; it stayed in ChatGPT until the June 27, 2026 retirement, announced May 28, 2026 with a 30-day sunset.
  • Part of the “the names are unmanageable” period — a GPT-4.5 slotted between GPT-4o and the reasoning models that the eventual GPT-5 unification was meant to clean up.
Model
o4-mini
o4-mini
Reasoning
Legacy
Apr 16, 2025
Cheap and fast reasoning model. Closes most of the gap to o3 at a fraction of the cost. Tool use during chain-of-thought. Retired from ChatGPT Feb 13, 2026; API snapshot shutdown scheduled Oct 23, 2026.
  • Released April 16, 2025; the smaller, cheaper companion to o3 in the “next-gen reasoning” April 2025 wave.
  • Tool use during reasoning — the model can call tools (web search, code execution, file reads) inside its chain-of-thought rather than only at the boundary.
  • Closes most of the quality gap with o3 at a small fraction of its cost; the routine recommendation for high-volume reasoning workloads.
  • The naming choice (o4-mini, not o3.5 or o4) follows the same “skip the GA flagship and ship the mini” pattern as o3-mini four months earlier.
Model
o3
o3
Reasoning
Legacy
Apr 16, 2025
GA reasoning flagship. Frontier scores on math and coding benchmarks. Tool-use-in-reasoning by default. Retired from ChatGPT Aug 26, 2026 (90-day sunset announced May 28, 2026); still served in the API. The deep-research snapshots were deprecated first (Apr 22, 2026; shut down Jul 23, 2026); the base o3-2025-04-16 snapshot was then deprecated Jun 11, 2026 with an API shutdown scheduled Dec 11, 2026 (recommended replacement now gpt-5.6-sol).
  • Released April 16, 2025 as the GA flagship of the reasoning line, alongside o4-mini.
  • Frontier-grade scores on competition math (AIME) and the hardest coding benchmarks at launch.
  • Replaced o1 as the recommended top-of-line reasoning model for paying API users; o1 demoted to Available.
  • The model that the “reasoning works as a research direction” framing of late 2024 onward is most often pinned to in retrospect.
  • Retired from ChatGPT on August 26, 2026 — announced May 28, 2026 with a 90-day sunset (a longer window than GPT-4.5’s 30 days, reflecting o3’s deeper entrenchment in user workflows). Only the ChatGPT model-picker access went away; o3 stayed available via the API, where its own shutdown runs on a separate clock to December 11, 2026. See OpenAI’s model release notes and the deprecations page.
Model
o3-mini
o3-mini
Reasoning
Legacy
Jan 31, 2025
First non-preview reasoning model. Shipped two months ahead of the GA o3 flagship and free in ChatGPT. Deprecated Apr 22, 2026; o3-mini API shutdown scheduled Oct 23, 2026 (recommended replacement gpt-5.6-sol).
  • Released January 31, 2025; the first reasoning model OpenAI shipped outside the original o1 preview / GA pair.
  • Made reasoning available to free-tier ChatGPT users for the first time, with rate limits.
  • Three reasoning effort levels (low, medium, high) exposed via the API — the convention later carried forward to GPT-5.
  • Superseded by o4-mini in April 2025. Deprecated in OpenAI’s April 22, 2026 legacy-snapshot notice; o3-mini-2025-01-31 and the undated o3-mini alias are scheduled to shut down October 23, 2026, with gpt-5.6-sol listed as the recommended replacement.
Model
o1
o1
Reasoning
Legacy
Dec 5, 2024
GA release of the o-series flagship. Replaced o1-preview. The first reasoning model on the OpenAI API at general availability. Deprecated Apr 22, 2026; API shutdown scheduled Oct 23, 2026.
  • Released December 5, 2024 as the GA replacement for o1-preview, on the “12 Days of OpenAI” launch event.
  • Substantial gains over o1-preview on most reasoning benchmarks; vision support added in the same release.
  • First “Pro mode” debut on chatgpt.com — a $200/mo tier that runs o1 with much higher compute per response.
  • Superseded by o3 in April 2025; still served via the API.
Model
o1-mini
o1-mini
Reasoning
Legacy
Sep 12, 2024
Smaller, cheaper, faster reasoning companion to o1-preview. Coding-focused; strong on STEM benchmarks at a fraction of the price. Deprecated Apr 28, 2025; o1-mini shut down in the API Oct 27, 2025.
  • Released September 12, 2024 alongside o1-preview as the first pair of public reasoning models.
  • Optimized for STEM and coding tasks; comparable to o1-preview on coding evaluations at a meaningful price drop.
  • The launch of the “Strawberry” line publicly — OpenAI's internal codename that had leaked widely in the months prior.
  • Superseded by o3-mini in January 2025. Deprecated April 28, 2025 with a six-month notice; o1-mini was removed from the API on October 27, 2025, with o4-mini as the listed replacement.
Model
o1-preview
o1-preview
Reasoning
Legacy
Sep 12, 2024
First public reasoning model. Trained to think step-by-step before answering. Set a new frontier on competition math and hard coding. Deprecated Apr 28, 2025; o1-preview shut down in the API Jul 28, 2025.
  • Released September 12, 2024 as a research preview — the first OpenAI model trained with chain-of-thought as a first-class capability rather than a prompt-engineering technique.
  • Frontier scores on AIME (American Invitational Mathematics Examination) and Codeforces at launch — a step-change on the hardest benchmarks.
  • Slower and more expensive per token than GPT-4o, with the trade-off framed as “latency for thinking”.
  • Superseded by the GA o1 in December 2024. Deprecated April 28, 2025 with a three-month notice — the shorter of the pair, reflecting its preview status — and removed from the API on July 28, 2025, with o3 as the listed replacement.

The reasoning era — September 12, 2024. Above this line: the Reasoning family (o1-preview, o1, o1-mini, o3, o3-mini, o4-mini) and the post-reasoning GPT models (GPT-4.1, GPT-4.5, GPT-5) that followed. Below: the original GPT chat line from GPT-3.5 / ChatGPT through GPT-4o. The split lasted just under a year before GPT-5 unified the two surfaces.

Model
GPT-4o mini
gpt-4o-mini
GPT
Legacy
Jul 18, 2024
Cheaper, faster GPT-4o for high-volume calls. Replaced GPT-3.5 Turbo as the recommended cheap-tier chat model. Retired from ChatGPT Feb 13, 2026; API access via dated snapshot continues.
  • Released July 18, 2024 as the small, cheap companion to GPT-4o.
  • Pricing at $0.15 / $0.60 per million input / output tokens at launch — an order of magnitude cheaper than GPT-3.5 Turbo had been when it shipped.
  • Quickly became the routine recommendation for classification, routing, and high-volume agent steps.
  • Multimodal (vision and text input) on the same surface as GPT-4o; voice support added later.
Model
GPT-4o
gpt-4o
GPT
Legacy
May 13, 2024
Native multimodal — text, image, and audio in a single model. The default ChatGPT model for most of 2024 / 2025. Retired from ChatGPT Feb 13, 2026; chatgpt-4o-latest API snapshot removed Feb 17, 2026; gpt-4o-2024-05-13 snapshot shutdown scheduled Oct 23, 2026.
  • Released May 13, 2024; the “o” standing for “omni,” reflecting native handling of text, image, and audio in a single model rather than the prior speech-to-text-to-LLM pipeline.
  • Advanced Voice Mode demoed at launch and rolled out incrementally through the second half of 2024 — conversational latency low enough that natural turn-taking became practical.
  • Replaced GPT-4 Turbo as the default model on chatgpt.com and stayed the routed default until GPT-5 took over fifteen months later.
  • The Sky-voice / Scarlett Johansson controversy (May 2024) erupted around this launch; OpenAI removed the Sky voice within days.
  • Free-tier ChatGPT users gained access to a frontier multimodal model for the first time, with daily limits.
Model
GPT-4 Turbo
gpt-4-turbo
GPT
Legacy
Nov 6, 2023
128k context. Cheaper than original GPT-4. Custom GPTs and the Assistants API launched on the same DevDay. Eleven days later the board fired Altman. Deprecated Apr 22, 2026; gpt-4-turbo API shutdown scheduled Oct 23, 2026 (recommended replacement gpt-5.6-sol).
  • Announced November 6, 2023 at DevDay 2023; the same event that launched Custom GPTs, the GPT Store, and the Assistants API.
  • 128,000-token context window — a sixteen-times increase over the original GPT-4 8k.
  • Substantially cheaper per token than GPT-4 at launch; the price drop set the cadence for the GPT-4 line through 2024.
  • Eleven days after this DevDay, the OpenAI board fired Sam Altman — the November 2023 episode covered in the prose history below.
  • Superseded by GPT-4o in May 2024. Deprecated in OpenAI’s April 22, 2026 legacy-snapshot notice; gpt-4-turbo, gpt-4-turbo-2024-04-09, and gpt-4-1106-preview are scheduled to shut down October 23, 2026, with gpt-5.6-sol listed as the recommended replacement.
Model
GPT-4
gpt-4
GPT
Legacy
Mar 14, 2023
First multimodal frontier model. “Passes the bar exam.” Plugins and Code Interpreter shipped within weeks. Deprecated Apr 22, 2026; gpt-4 API shutdown scheduled Oct 23, 2026 (recommended replacement gpt-5.6-sol).
  • Released March 14, 2023; the same week Anthropic shipped Claude 1 in limited access.
  • First OpenAI model with image input (rolled out incrementally over the following months).
  • The “passes the bar exam in the 90th percentile” framing in the GPT-4 Technical Report drove the mainstream coverage; later research disputed the methodology, but the framing stuck.
  • Initial 8k context window with a separate 32k variant; both later superseded by Turbo's 128k.
  • Plugins launched March 23, 2023 (deprecated 2024 in favor of Custom GPTs and tool calling). Code Interpreter followed in July 2023.
  • Superseded by GPT-4 Turbo in November 2023 and GPT-4o in May 2024. Deprecated in OpenAI’s April 22, 2026 legacy-snapshot notice; gpt-4-0613 and the undated gpt-4 alias (plus the completions variants and fine-tuned ft-gpt-4 models) are scheduled to shut down October 23, 2026, with gpt-5.6-sol listed as the recommended replacement.
Model
GPT-3.5 / ChatGPT launch
gpt-3.5-turbo
GPT
Legacy
Nov 30, 2022
ChatGPT launched as a “low-key research preview” on top of GPT-3.5. One million users in five days, a hundred million in two months. gpt-3.5-turbo-instruct and gpt-3.5-turbo-1106 shut down Sep 28, 2026; gpt-3.5-turbo itself follows Oct 23, 2026 (recommended replacement gpt-5.6-terra).
  • Released November 30, 2022; OpenAI framed it publicly as a “low-key research preview,” an internal expectation that turned out to be wrong by orders of magnitude.
  • One million users in five days, a hundred million in two months — at the time, the fastest-growing consumer application in history.
  • The model was a fine-tuned descendant of the InstructGPT-era text-davinci-* base; the gpt-3.5-turbo snapshot came in March 2023 and became the default chat-completion model for most of 2023.
  • The product (ChatGPT) was distinct from the model name (GPT-3.5) from day one — the start of the naming confusion that would compound through GPT-4 / 4 Turbo / 4o / 4.1.
  • The last GPT-3.5 strings go dark in autumn 2026. OpenAI’s September 26, 2025 notice retires gpt-3.5-turbo-instruct and gpt-3.5-turbo-1106 (alongside babbage-002 and davinci-002) on September 28, 2026, and its April 22, 2026 notice retires gpt-3.5-turbo-0125, the undated gpt-3.5-turbo alias, the completions variants, and fine-tuned ft-gpt-3.5-turbo models on October 23, 2026. The recommended replacement on both is gpt-5.6-terra. After that the model behind the original ChatGPT is no longer callable.

ChatGPT launches — November 30, 2022. Above this line: the consumer-product era. OpenAI became a household name within weeks, the company's center of gravity shifted from research lab to product organization, and every model release after this point was framed in part as “what does this change for ChatGPT.” Below: the Pre-ChatGPT era — GPT-1 / 2 / 3 and the InstructGPT / text-davinci base models that quietly defined the technical lineage before the consumer launch.

Model
InstructGPT / text-davinci series
text-davinci-001, text-davinci-002, text-davinci-003
Pre-ChatGPT
Legacy
Jan 27, 2022
RLHF arrives. The technical predecessor to ChatGPT. Three text-davinci-* snapshots through 2022.
  • InstructGPT introduced in “Training language models to follow instructions with human feedback” (Ouyang et al., 2022) — the public arrival of RLHF as the technique that turned base GPT-3 into a usable assistant.
  • The text-davinci-001, text-davinci-002, and text-davinci-003 snapshots through 2022 were the production face of the InstructGPT line, available on the OpenAI API to developers a year before ChatGPT launched.
  • Technically GPT-3.5-family base models; rolled into a single row here rather than per-snapshot because the differences were incremental and the distinct rows belong to the consumer-facing products that came after.
  • All snapshots later sunset as gpt-3.5-turbo took over.
Model
GPT-3
davinci, curie, babbage, ada
Pre-ChatGPT
Legacy
May 28, 2020
175B parameters. “Language Models are Few-Shot Learners.” The first GPT to feel mainstream-newsworthy. API-only via the Playground. The four engine names were retired Jan 4, 2024; their replacements babbage-002 / davinci-002 shut down Sep 28, 2026.
  • Released May 28, 2020; the paper is “Language Models are Few-Shot Learners” (Brown et al., 2020).
  • 175 billion parameters — over a hundred-fold larger than GPT-2; the central illustration of the “scale matters” thesis.
  • Distributed via the OpenAI API (waitlist, then Playground) under the four engine names davinci / curie / babbage / ada — the first commercial product OpenAI shipped.
  • Used few-shot prompting as the default interaction pattern; chat-style instruction-following had to wait for InstructGPT.
  • Microsoft announced an exclusive license to the underlying weights in September 2020, separate from API access.
  • The engine names went first; their replacements go September 28, 2026. The four original engines were retired from the API on January 4, 2024 in favor of babbage-002 and davinci-002; those two base models are themselves scheduled to shut down on September 28, 2026 under OpenAI’s September 26, 2025 legacy-snapshot notice, with gpt-5.6-terra as the listed replacement. That date closes out the GPT-3 base-model line on the API entirely.
Model
GPT-2
gpt2 / gpt2-medium / gpt2-large / gpt2-xl
Pre-ChatGPT
Legacy
Feb 14, 2019
Initially withheld over “misuse concerns,” then released in stages through November 2019. 1.5B parameters.
  • Announced February 14, 2019; the paper is “Language Models are Unsupervised Multitask Learners” (Radford et al., 2019).
  • Initial release withheld the largest 1.5B-parameter model over “concerns about malicious applications.” OpenAI staged the release through 2019, with the full model published in November.
  • The decision to withhold became one of the more debated episodes in the AI-safety conversation of the period; the staged release set the template for “responsible disclosure” framings that later returned with GPT-4 and beyond.
  • 1.5B parameters at the largest size; meaningful generative coherence at multi-paragraph length was novel at the time.
  • Open-weights release means GPT-2 is still widely usable today via Hugging Face and similar surfaces.
Model
GPT-1
openai-gpt
Pre-ChatGPT
Legacy
Jun 11, 2018
The original GPT. “Improving Language Understanding by Generative Pre-Training.” 117M parameters. The start of the lineage.
  • Released June 11, 2018; the paper is “Improving Language Understanding by Generative Pre-Training” (Radford et al., 2018).
  • 117 million parameters; trained on the BookCorpus dataset.
  • Established the “generative pre-training, then fine-tune” recipe that every later GPT extended.
  • Technically a Transformer-decoder architecture, following Vaswani et al.'s “Attention is All You Need” (2017) by less than a year.
  • Open-weights release; the model is still inspectable today, more interesting now as the ancestor of the line than as a usable system.

Click any row to expand. Each row has a stable id for sharing — e.g. /ai/chatgpt/versions/#gpt-5, #gpt-4o, #o1-preview. The current model list and lifecycle status is at platform.openai.com/docs/models; deprecation timelines at platform.openai.com/docs/deprecations.

The 2015 founding

OpenAI was announced on December 11, 2015 as a nonprofit AI research lab. The cofounders were Sam Altman, Elon Musk, Greg Brockman, Ilya Sutskever (recruited from Google Brain), John Schulman, Wojciech Zaremba, Andrej Karpathy, Pamela Vagata, Trevor Blackwell, and Vicki Cheung. The launch announcement listed an initial $1 billion in pledges from Altman, Musk, Reid Hoffman, Peter Thiel, Y Combinator Research, Microsoft, Amazon Web Services, and Infosys; in practice, most pledges were never fully called.

The stated mission was to ensure that artificial general intelligence “benefits all of humanity,” with research published openly. The legal vehicle was a 501(c)(3) nonprofit. Both of those framings would shift materially over the following decade.

The Musk departure (2018)

Elon Musk left the OpenAI board in February 2018, citing a conflict of interest with Tesla's AI work. He stopped funding shortly after. Internal accounts later reported in NYT and The Information describe an unsuccessful attempt by Musk to take direct control of the lab in the prior period. The departure is the load-bearing prelude to Musk v. Altman in 2024 and to xAI's founding in 2023.

The capped-profit conversion (2019)

In March 2019, OpenAI created OpenAI LP, a “capped-profit” subsidiary, to raise commercial capital without abandoning the nonprofit charter. The cap on returns to investors was reportedly set at one hundred times their investment, with profits beyond that flowing back to the nonprofit. Microsoft invested $1 billion in July 2019, with Azure becoming a key cloud commitment.

The capped-profit structure is the structural origin of essentially every subsequent governance dispute — the November 2023 board episode, the for-profit conversion fight, the lawsuits framing the conversion as a betrayal of the founding mission.

The GPT lineage and InstructGPT

The technical lineage runs GPT-1 (June 2018) on BookCorpus, GPT-2 (Feb 2019) with the staged release, GPT-3 (May 2020) at 175B parameters, and InstructGPT (Jan 2022) introducing RLHF as the technique that turned a base language model into a usable assistant. The InstructGPT paper is the technical predecessor to ChatGPT itself, even though the consumer product launched ten months later under different framing.

ChatGPT launches (November 30, 2022)

ChatGPT launched as a “low-key research preview” on top of GPT-3.5 on November 30, 2022. OpenAI's internal expectation was for steady research-community adoption; reality was one million users in five days, a hundred million in two months — at the time, the fastest-growing consumer application in history.

The launch is the inflection point of the entire post-2015 history of OpenAI. The company's center of gravity shifted from research lab to product organization within weeks; every governance and financial dispute that followed was downstream of the consumer success that arrived faster than anyone had planned for.

The Microsoft expansion (January 2023)

In January 2023, Microsoft announced a multi-year, multibillion-dollar investment widely reported as $10 billion — an extension of the 2019 commitment. Azure became the exclusive cloud provider for OpenAI's training and inference. The relationship has been the financial backbone of the company through at least 2025, and is the recurring subtext of the for-profit conversion fight, the AGI-clause dispute, and the various antitrust filings that have followed.

GPT-4 and the multimodal era (March 2023)

GPT-4 launched on March 14, 2023 with image input, far stronger reasoning, and the “passes the bar exam in the 90th percentile” framing that drove most of the mainstream coverage. Plugins followed two weeks later; Code Interpreter in July. The plugin surface was deprecated in 2024 in favor of Custom GPTs and tool calling, but it set the early template for “LLM as orchestrator of external services.”

DevDay 2023 (November 6, 2023)

OpenAI's first DevDay launched GPT-4 Turbo (128k context, lower price), Custom GPTs, the GPT Store, the Assistants API, and a handful of tooling improvements. The conference was widely treated at the time as a victory lap. Eleven days later, the board fired Altman.

The November 2023 board episode (Nov 17 – 21, 2023)

On November 17, 2023, the OpenAI board — Helen Toner, Tasha McCauley, Adam D'Angelo, and Ilya Sutskever — fired Sam Altman, citing a loss of confidence in his communications with the board. Greg Brockman resigned in protest the same day. Mira Murati was named interim CEO; within twenty-four hours she was reportedly negotiating Altman's return.

On November 19, the board appointed Emmett Shear (former Twitch CEO) as the second interim CEO. Roughly seven hundred of OpenAI's seven hundred and seventy employees signed an open letter threatening to follow Altman to Microsoft if the board did not resign and reinstate him. Microsoft publicly offered to hire Altman, Brockman, and any departing employees to lead a new AI research division.

On November 21 — five days after the firing — Altman returned as CEO. The board was reconstituted with Bret Taylor as chair, Larry Summers, and D'Angelo continuing. Toner, McCauley, and Sutskever were off the board. Helen Toner has since described her account in a TED talk and an academic paper. Contemporaneous reporting in NYT, WSJ, Bloomberg, and The Information remains the most thorough public record.

The 2024 leadership exodus

The year after the board episode produced the most concentrated departure of senior leadership in OpenAI's history. Andrej Karpathy left in February 2024 (later founding Eureka Labs). Ilya Sutskever left in May 2024 and founded Safe Superintelligence Inc. Jan Leike left the same week and joined Anthropic, citing concerns about OpenAI's safety culture in a public thread that read as an indictment. The superalignment team was dissolved within days.

John Schulman, an OpenAI cofounder and the original lead on RLHF and on ChatGPT itself, left in August 2024 for Anthropic, then later moved to Mira Murati's startup. Mira Murati herself left in September 2024 and founded Thinking Machines Lab; Bob McGrew (Chief Research Officer) and Barret Zoph (VP Research) left the same week. Greg Brockman took an extended sabbatical from May to August 2024 and returned.

After the October 2025 recapitalization, the operating senior team consists of Sam Altman (CEO), Greg Brockman (President), Sarah Friar (CFO), and Mark Chen (SVP Research / Chief Research Officer); long-time COO Brad Lightcap moved to special projects in spring 2026 and then left the company outright on August 11, 2026 to start a venture he has not described, with no successor named. Bret Taylor chairs the board of the controlling OpenAI Foundation. A second cluster of departures landed in April 2026 — Kevin Weil, Bill Peebles, and Srinivas Narayanan — alongside the wind-down of side projects (Sora, OpenAI for Science). Two changes in July 2026 bear directly on who decides what ships: Fidji Simo stepped down from the CEO of Applications role on July 9 without a named successor, leaving Brockman with the ChatGPT product, go-to-market, enterprise, and compute portfolios; two days later head of safety systems Johannes Heidecke left as OpenAI folded safety into research under Mia Glaese, newly titled VP of Research and Safety. The commercial side turned over next: on August 13, 2026 OpenAI named Dali Rajic — previously president and COO of Wiz, and of Zscaler before that — Chief Revenue Officer, with Denise Dresser leaving after a transition period; see OpenAI’s announcement. The current roster should be re-verified at every refresh given the volatility of the prior period.

The Sky / Scarlett Johansson voice (May 2024)

At the GPT-4o launch demo on May 13, 2024, OpenAI presented a voice named “Sky” that listeners and Scarlett Johansson herself perceived as a deliberate echo of her performance in the 2013 film Her. Johansson stated publicly that she had previously declined OpenAI's offer to voice the assistant. OpenAI removed the Sky voice within days. The episode is recounted briefly here because it bears on the launch row and on how OpenAI's public-facing decisions were perceived in the period; it is not enlarged into its own section.

The for-profit conversion (2024 – 2025)

Through 2024 and 2025, OpenAI publicly worked toward converting the capped-profit LP into a public-benefit corporation. The transaction closed on October 28, 2025: the operating subsidiary became OpenAI Group PBC, controlled by the renamed nonprofit OpenAI Foundation (which holds about 26% equity, worth roughly $130 billion at closing); Microsoft holds about 27% (~$135 billion). Microsoft and OpenAI amended the agreement again on April 27, 2026 to make Microsoft's IP license non-exclusive and to recut the revenue-share. The Musk legal challenge survived the closing — Musk v. Altman went to trial in April 2026, and on May 18, 2026 the advisory jury (its verdict adopted by Judge Yvonne Gonzalez Rogers) found in favor of OpenAI, ruling Musk's claims time-barred by the three-year statute of limitations rather than reaching the merits; Musk called it a “calendar technicality” and said he would appeal.

The reasoning era (September 2024 onward)

o1-preview and o1-mini in September 2024 introduced the “thinks before answering” pattern as a first-class model capability rather than a prompt-engineering technique. o1 went GA in December 2024, o3-mini in January 2025, and the GA o3 and o4-mini in April 2025 alongside the GPT-4.1 line. GPT-5 in August 2025 unified the GPT and reasoning surfaces behind a single router that decides per-query whether to answer immediately or invoke reasoning. The split as a user-visible product distinction lasted just under a year.

The lawsuits

The litigation around OpenAI through 2024 – 2026 is unusually wide-ranging.

  • Musk v. Altman, Brockman, OpenAI, et al. Filed February 29, 2024 in California state court alleging breach of the founding agreement; voluntarily dismissed June 2024; refiled August 2024 in N.D. Cal. with additional claims including RICO. Tried in spring 2026; on May 18, 2026 the advisory jury found for OpenAI on statute-of-limitations grounds, and on May 20, 2026 the court adopted that verdict as its own findings. The case is not over: nine antitrust and Lanham Act counts survive and are briefed through October 2026, a mediator was appointed in July 2026, and no final judgment has been entered and no appeal noticed.
  • The New York Times Company v. Microsoft Corp. & OpenAI (S.D.N.Y., filed December 27, 2023). The flagship publisher copyright case, alleging mass copying of Times articles into the training corpus and reproduction of paywalled content in outputs. Judge Sidney H. Stein's April 4, 2025 ruling narrowed but did not gut the case, letting the direct and contributory infringement claims proceed. The case is now part of MDL No. 3143, centralized in S.D.N.Y. before Judge Stein in April 2025; summary-judgment and Daubert briefing runs to November 2026. Two developments landed in the days before this refresh. On August 31, 2026 Judge Stein ordered the Times to show cause in writing by September 11 why its own action should not be stayed pending resolution of the summary-judgment motions in the MDL's other active cases, with any defense response due September 18. And on September 1, 2026 the U.S. Department of Justice filed a Statement of Interest in the MDL siding with OpenAI — a twenty-page brief signed by Associate Attorney General Stanley Woodward arguing that training a large language model on copyrighted text is generally fair use, criticizing the fourth-factor analysis in Kadrey v. Meta, and warning that a licensing requirement would concentrate model-building among the largest firms. It is advisory, not a ruling; Judge Stein is free to disregard it.
  • Authors Guild v. OpenAI (S.D.N.Y., filed September 2023, consolidated with the author class actions). Plaintiffs include George R.R. Martin, John Grisham, Jodi Picoult, and others.
  • The Tremblay / Silverman / Awad / Chabon class actions (N.D. Cal., 2023). Centralized into Authors Guild v. OpenAI by the JPML transfer order of April 3, 2025; the consolidated class complaint is captioned Baldacci v. OpenAI. Most non-direct-infringement claims (DMCA Section 1202, negligence, unjust enrichment) were dismissed early; direct copyright claims survive, and class-certification briefing is sequenced to begin after the summary-judgment ruling rather than before it.
  • Daily News, Tribune, MediaNews Group, et al. v. OpenAI & Microsoft (S.D.N.Y., April 2024). A coalition of eight regional newspaper publishers.
  • Center for Investigative Reporting (Mother Jones / Reveal) v. OpenAI (June 2024).
  • The Intercept v. OpenAI (S.D.N.Y., February 2024) and Raw Story / AlterNet v. OpenAI (S.D.N.Y., February 2024) — the DMCA Section 1202 angle on training-data sourcing. The S.D.N.Y. dismissed the Section 1202 counts for lack of Article III standing in late 2024; the Second Circuit heard argument on March 18, 2026 and has not yet ruled.
  • Doe v. GitHub Copilot (N.D. Cal.) — the Codex copyright case; OpenAI is a co-defendant alongside GitHub and Microsoft.
  • In re: ChatGPT Product Liability Cases (Cal. Super. Ct., San Francisco — JCCP No. 5431). The one track that asks what the product was allowed to say rather than what it was allowed to read: twelve California wrongful-death and product-liability suits pleading defective design, failure to warn, and negligence. Opened by Raine v. OpenAI in August 2025 and coordinated into a single proceeding on February 3, 2026; parallel federal actions on the same theory began arriving in July 2026. It is the litigation most directly legible in the release timeline — the under-18 safeguards that shipped with the August 6, 2026 ChatGPT retune cover the same ground.
  • European regulators. The Italian Garante temporarily banned ChatGPT in March 2023 over GDPR concerns and lifted the ban after OpenAI added age and consent affordances. Various follow-on actions across the EU continue.

The Microsoft entanglement

The financial structure was rewritten twice in the most recent cycle. The October 2025 recapitalization replaced the capped-profit cap with ordinary equity (Microsoft holds approximately 27% of OpenAI Group PBC, ~$135 billion at closing); extended Microsoft's IP rights through 2032; and changed the “AGI clause” so that an OpenAI-board declaration of AGI is now verified by an independent expert panel before contractual consequences attach. The April 2026 amendment made Microsoft's IP license non-exclusive and let OpenAI serve products on any cloud.

People who shaped ChatGPT, and where they went

2015 cofounders: Sam Altman (CEO), Elon Musk (left 2018, founded xAI in 2023, plaintiff in Musk v. Altman), Greg Brockman (President), Ilya Sutskever (left May 2024 → Safe Superintelligence Inc.), John Schulman (left Aug 2024 → Anthropic, then Thinking Machines Lab), Wojciech Zaremba, Andrej Karpathy (left Feb 2024 → Eureka Labs), Pamela Vagata, Trevor Blackwell, Vicki Cheung.

Post-founding leadership departures: Mira Murati (CTO; left Sep 2024 → Thinking Machines Lab). Jan Leike (Superalignment co-lead; left May 2024 → Anthropic). Bob McGrew (Chief Research Officer; left Sep 2024). Barret Zoph (VP Research; left Sep 2024).

Current leadership (verify at refresh): Sam Altman (CEO), Greg Brockman (President, returned from sabbatical Aug 2024; since July 2026 also the owner of the ChatGPT product, go-to-market, enterprise, and compute portfolios), Sarah Friar (CFO), Mark Chen (Chief Research Officer / SVP Research), Dali Rajic (Chief Revenue Officer, from Aug 13, 2026), Bret Taylor (board chair, since Nov 21, 2023). Kevin Weil (CPO) departed April 2026; Fidji Simo, CEO of Applications from May 2025, stepped down to a part-time advisory role on July 9, 2026 and was not replaced; Brad Lightcap, COO from 2022 and on special projects from spring 2026, left the company on August 11, 2026; Denise Dresser left the Chief Revenue Officer role in the same August 13, 2026 announcement that named her successor, staying through a transition period.

The competitive landscape

ChatGPT competes most directly with Anthropic's Claude (founded by 2021 OpenAI departures, including Dario and Daniela Amodei and several GPT-3 paper authors), Google's Gemini (the post-Bard rebrand of DeepMind's consumer surface), Meta's Llama (the largest open-weights line), and xAI's Grok (Musk's post-OpenAI venture). The Musk thread now reaches the developer-tool supply chain as well: on August 28, 2026 OpenAI notified SpaceX — which had acquired the Cursor editor, and which already owns X and xAI — that it will wind down the contract supplying OpenAI models to Cursor, with a proposed cut-off of November 12, 2026, the latest date its contract allows. OpenAI said it could not be confident SpaceX would use the models within its terms of service, citing the post-acquisition Twitter contract breach and Musk's sworn admission that xAI had violated those terms, and pointed to the accountability bar set by its then-forthcoming Astra model, which shipped six days later. See OpenAI's statement. Positioning shifts release-to-release; this page does not attempt a benchmark roundup or a ranking.

Use ChatGPT

The browser cannot detect which ChatGPT or GPT model you've used or are using — there's no fingerprint or header that exposes it. The block below carries the practical information instead: the current model strings, a copy-paste API call, and the surfaces where ChatGPT is available.

Current model strings

Use these in the model field of an API request. Verify against platform.openai.com/docs/models for the freshest list.

# Current flagship — the model OpenAI's docs tell you to start from
gpt-6-astra          # GPT-6 flagship  — $10 / $50 per 1M tokens

# Prior generation — still served; one number, three durable capability tiers
gpt-5.6-sol          # flagship; alias gpt-5.6 — $4 / $20 per 1M, promo to Nov 21, 2026
gpt-5.6-terra        # balanced        — $2 / $12 per 1M tokens
gpt-5.6-luna         # cost-efficient  — $0.20 / $1.20 per 1M tokens

# Older, still served
gpt-5.5
gpt-5.5-pro

# Smaller workhorses — the current mini / nano tier
gpt-5.4-mini
gpt-5.4-nano

# Approval-gated — OpenAI Daybreak program only, Responses API
gpt-5.6-cyber            # Daybreak Red    — $12.50 / $75 per 1M tokens
gpt-daybreak-red-latest  # alias → gpt-5.6-cyber
gpt-daybreak-blue-latest # alias → gpt-5.6-sol, defensive-security safeguards

Quick API call

Drop in your OPENAI_API_KEY and run. The Chat Completions endpoint is the canonical entry point for the GPT line.

$ curl https://api.openai.com/v1/chat/completions \
    -H "Authorization: Bearer $OPENAI_API_KEY" \
    -H "Content-Type: application/json" \
    -d '{
      "model": "gpt-6-astra",
      "messages": [
        {"role": "user", "content": "Hello, ChatGPT."}
      ]
    }'

Where to access ChatGPT

Multiple surfaces, same models underneath. Pick whichever fits the task.

# Web chat — Free, Go, Plus, Pro, Team, Enterprise, Edu tiers
https://chatgpt.com/

# API and developer platform
https://platform.openai.com/
https://platform.openai.com/docs/

# Native apps
ChatGPT for iOS, Android, macOS, Windows

# Built-in surfaces
ChatGPT inside Apple Intelligence  # iOS / iPadOS / macOS 18+
Microsoft Copilot                  # powered partly by OpenAI models

Model lifecycle

OpenAI deprecates older models on a published timeline. Pin a snapshot only when you need bit-for-bit reproducibility; otherwise prefer the un-dated alias so you migrate forward automatically.

# Deprecation schedule and lifecycle policy
https://platform.openai.com/docs/deprecations

# Stable alias — rolls forward to the latest snapshot
"model": "gpt-5.6"

# Pinned snapshot — freezes the exact training cut
"model": "gpt-4o-2024-08-06"

Sources: OpenAI model docs; OpenAI news / blog; OpenAI deprecations; the GPT-1, GPT-2, GPT-3, and InstructGPT papers (Radford / Brown / Ouyang et al.); NYT v. Microsoft & OpenAI, Authors Guild v. OpenAI, Musk v. Altman, and the related dockets; contemporaneous reporting in NYT, WSJ, Bloomberg, and The Information; Helen Toner's TED talk on the November 2023 episode; Karen Hao's Empire of AI. Last updated September 8, 2026.

Mungomash LLC · More AI pages