AI · Live reference

The frontier AI models, right now

The frontier AI models, as of September 8, 2026: Anthropic's Claude Fable 5.1 (the Mythos-class flagship; released September 1, 2026, generally available on every platform on day one with no preview stage — Claude Opus 5, shipped July 24, 2026, leads the Opus tier below it), OpenAI's GPT-6 Astra (released September 3, 2026 to a limited set of organizations, generally available by September 8 — the GPT-5.6 Sol / Terra / Luna tiers stay served below it), Google's Gemini 3.8 Flash (Stable GA September 2, 2026), xAI's Grok 4.6 (released August 12, 2026), Meta's Muse Spark 1.3 (released September 2, 2026), DeepSeek's DeepSeek-V4-Pro (public preview April 24, 2026; generally available August 13, 2026), Mistral's Mistral Medium 3.5 (released April 28, 2026), and Alibaba's Qwen3.8-Max (released August 3, 2026). The detail below covers each family in full — every release, every API string, every context window.

Last verified: September 8, 2026 · refreshed daily. 8 families · 3 US public · 2 US private · 2 Chinese · 1 EU.

Filter by lab status

Frontier AI model families table

Family
ChatGPT
GPT · o-series
OpenAI US private
GPT-6 Astra
Sep 3, 2026

What is the newest ChatGPT or GPT model?

OpenAI's flagship line, from GPT-1 (June 2018) through GPT-6 Astra, released September 3, 2026. OpenAI's launch post described Astra as “rolling out today to a limited set of organizations,” with the API and the Plus / Pro / Business / Enterprise plans following “over the coming days”; that gate has since come off — as of September 8, 2026 OpenAI's model docs carry no limited-access notice and point developers at Astra as the default starting point. Astra opens the GPT-6 generation and supersedes the GPT-5.6 family, which had introduced a naming system of generation number plus a durable capability tier: Sol (flagship / highest intelligence), Terra (balanced everyday work), and Luna (fastest / lowest cost). The August 2025 GPT-5 release introduced the unified router that picks between fast-response and reasoning models within a single endpoint; the 5.x cadence ran roughly six weeks per release through 5.5, then to the 5.6 generation. The o-series reasoning models (o1, o3, o4) were merged into the main line under GPT-5.

  • Lab: OpenAI — San Francisco, California, USA. Privately held; confidential draft S-1 submitted to the SEC (announced June 2026), with no listing date, size, or exchange disclosed.
  • Current model: GPT-6 Astra, released September 3, 2026 through a staged rollout and generally available by September 8 — OpenAI's strongest model, and the one its API docs now tell developers to start from. API: gpt-6-astra, with published rate limits across all five usage tiers and availability on Azure and AWS Bedrock. Two access caveats worth knowing: Enterprise workspaces are off by default until an admin enables it, and no free-tier date has been announced. It lists at $10 / $50 per million tokens, carries a 1.05M-token context window and an April 30, 2026 knowledge cutoff. Prior flagship GPT-5.6 (Sol tier, GA July 9, 2026) remains available at $4 / $20 — promotional pricing OpenAI commits to at least through November 21, 2026 — alongside gpt-5.6-terra and gpt-5.6-luna.
  • First model: GPT-1, June 11, 2018. Total versions: 29 (including GPT-5.6-Cyber, August 10, 2026, an approval-gated cyber-defense model that is not the general flagship).
  • Pages on this site:
  • Primary sources: OpenAI Models · OpenAI Announcements
Family
Claude
Mythos · Opus · Sonnet · Haiku
Anthropic US private
Claude Fable 5.1
Sep 1, 2026

What is the latest Claude model?

Anthropic's flagship line, from Claude 1 (March 2023) through the current production flagship Claude Fable 5.1 (September 1, 2026) — the head of the Mythos class, a tier Anthropic positions above Opus. Fable 5.1 was generally available on every platform on day one: a 1M-token context window, always-on adaptive thinking, and cache reads cut 75% against Fable 5. It supersedes Claude Fable 5, which went GA June 9, 2026, was briefly suspended June 12 under a US government export-control directive, and was restored to general availability and redeployed globally July 1, 2026 after the directive was lifted June 30. Anthropic shipped Claude Opus 5 (claude-opus-5) on July 24, 2026 — the new Opus-tier head, near Fable 5's capability at half its price — following the mid-tier Claude Sonnet 5 (claude-sonnet-5) on June 30, 2026. The invitation-only Mythos track advanced the same day: Claude Mythos 5.1 (claude-mythos-5-1) is what Anthropic now calls the current Mythos model, and its own model page describes it as the same model as Fable 5.1 — identical 1M context, 128K output, and $10 / $50 per million tokens — offered separately to approved Project Glasswing customers rather than generally available. It replaces claude-mythos-5 in that slot. The three-tier structure (Opus for the largest, Sonnet for the balanced mid-tier, Haiku for the small-and-fast tier) was introduced with Claude 3 in March 2024; the Mythos class sits above it. Constitutional AI is the safety-training framing that shipped with Claude 1.

  • Lab: Anthropic — San Francisco, California, USA. Privately held; $65B Series H at $965B post-money closed May 28, 2026, and a confidential draft S-1 submitted to the SEC June 1, 2026.
  • Current model: Claude Fable 5.1, generally available September 1, 2026 on every platform day one — Anthropic's most capable widely released model. API: claude-fable-5-1 (same identifier on Amazon Bedrock as anthropic.claude-fable-5-1, on Google Cloud, and on Microsoft Foundry). Claude Opus 5 (claude-opus-5, July 24, 2026) leads the Opus tier below it.
  • Newest release: Claude Fable 5.1 (claude-fable-5-1), September 1, 2026 — a Mythos-tier flagship change, with retuned cyber safeguards and a claude-mythos-5-1 trusted-access sibling.
  • First model: Claude 1, March 14, 2023. Total versions: 25.
  • Pages on this site:
  • Primary sources: Anthropic Models · Anthropic News
Family
DeepSeek
V-series · R-series
DeepSeek Chinese
DeepSeek-V4-Pro
Aug 13, 2026

What is the latest DeepSeek model?

Hangzhou-based open-weights frontier-LLM lab. Released DeepSeek-V3 (December 2024) and DeepSeek-R1 (January 2025) at training-cost levels far below US-frontier peers, sparking a global re-evaluation of frontier-model economics. V4-Pro is the current flagship: a 1.6-trillion-parameter Mixture-of-Experts model with 49B active parameters per token and a 1M-token context window, MIT-licensed. It went callable in public preview on April 24, 2026 and reached general availability on August 13, 2026 as the DeepSeek-V4-Pro-0813 build, with open weights published the same day.

  • Lab: DeepSeek (Hangzhou DeepSeek Artificial Intelligence Basic Technology Research) — Hangzhou, China. Privately held subsidiary of High-Flyer Capital Management.
  • Current model: DeepSeek-V4-Pro, shipped April 24, 2026 in public preview and generally available August 13, 2026. API: deepseek-v4-pro, now served by the DeepSeek-V4-Pro-0813 build (1M context, 384K max output). HuggingFace: deepseek-ai/DeepSeek-V4-Pro-0813. DeepSeek ended flat-rate API pricing on August 16, 2026 — V4-Pro now bills $0.66 / $1.98 per million tokens off-peak and $1.32 / $3.96 at peak, and since August 23, 2026 peak hours run Monday through Friday only — weekends bill at the off-peak rate.
  • First model: DeepSeek-Coder, November 2, 2023. Total versions: 22 (including DeepSeek-V4-Flash-Vision-Exp, the first multimodal V4 build, August 21, 2026).
  • Pages on this site:
  • Primary sources: DeepSeek API Docs · deepseek-ai on HuggingFace
Family
Gemini
Pro · Flash · Flash-Lite · Nano
Alphabet (Google) US public
Gemini 3.8 Flash
Sep 2, 2026

What's Google's most capable AI model right now?

Google's flagship line, from Bard (March 2023) and Gemini 1.0 (December 2023) through Gemini 3.8 Flash (September 2, 2026, Stable GA). Largest integration surface of any frontier line — Search AI Overviews, Workspace (Gmail, Docs, Slides, Meet, Drive), Pixel + Android (on-device Nano), Chrome Built-in AI. The 2.5 line (March 2025) was the first reasoning-as-default frontier model; the 3.x line moved Google to a one-launch-everywhere release pattern.

  • Lab: Google DeepMind, inside Alphabet (NASDAQ: GOOGL) — Mountain View, California, USA.
  • Current model: Gemini 3.8 Flash, Stable GA September 2, 2026. API: gemini-3.8-flash. Built on 3.7 Flash as a post-training iteration rather than a new pre-train, and priced at an introductory $0.75 / $3.75 per million tokens through December 31, 2026, rising to $1.50 / $7.50 on January 1, 2027. It takes the default-Antigravity-agent slot from 3.7 Flash. The Flash workhorse tier carries Google's frontier — the Pro slot is still held by Gemini 3.1 Pro Preview (gemini-3.1-pro-preview, February 19, 2026), with Gemini 3.5 Pro delayed and still in partner testing.
  • First model: Bard (LaMDA-based), March 21, 2023. Total versions: 33 (including Gemini 3.8 Flash and the gated Gemini 3.8 Flash Cyber, both September 2, 2026).
  • Pages on this site:
  • Primary sources: Gemini API Docs · Google DeepMind Blog
Family
Grok
Chat · Heavy · Multi-agent
xAI US public
Grok 4.6
Aug 12, 2026

What is the latest Grok model?

xAI's flagship line, from Grok 1 (November 2023) through Grok 4.6 (August 12, 2026), the current chat-and-coding flagship at 500K context — xAI's model docs call it “the most intelligent and fastest model we've built” and route every non-media use case to it. It holds Grok 4.5's price ($2 / $6 per million tokens) and context window, and adds an xhigh tier above the low / medium / high reasoning-effort dial. Grok 4.5 (July 8, 2026) remains available a rung below. Grok 4.3 (April 17, 2026) added native video input and in-chat generation of PDFs / slides / spreadsheets; multi-agent collaboration introduced with Grok 4.20 (March 10, 2026) and 16-agent Heavy mode at 1M-token context. The agentic-coding specialist Grok Build 0.1 (grok-build-0.1, May 19, 2026 — five days after the Grok Build CLI itself went to beta) ships as a Specialized SKU alongside, joined by Composer 2.5 (June 1, 2026 — a Kimi K2.5-based fast agentic-coding model selectable inside Grok Build). xAI is now part of SpaceX following the February 2026 SpaceX-xAI merger; X integration is the original distribution surface.

  • Lab: xAI, inside SpaceX (Nasdaq: SPCX) — San Francisco Bay Area, California, USA.
  • Current model: Grok 4.6, shipped August 12, 2026. API: grok-4.6 (500K context, $2 / $6 per million tokens below a 200K-token prompt and $4 / $12 above it, February 1, 2026 knowledge cutoff). Prior flagship Grok 4.5 (grok-4.5) and Grok 4.3 (grok-4.3) still available, with Grok 4.20 available as three 0309-dated SKUs (grok-4.20-multi-agent-0309, grok-4.20-0309-reasoning, grok-4.20-0309-non-reasoning).
  • First model: Grok 1, November 4, 2023. Total versions: 18 (including point-releases through 4.6, the May 2026 Grok Build 0.1, and the June 2026 Composer 2.5 specialized coding model).
  • Pages on this site:
  • Primary sources: xAI Models · xAI News
Family
Llama
Open-weights → closed frontier
Meta US public
Muse Spark 1.3
Sep 2, 2026

What's Meta's current frontier AI model?

Meta's frontier line, from LLaMA 1 (February 2023, originally research-only) through Llama 4 (April 2025) and into the post-Llama Muse Spark line — the first models from the new Meta Superintelligence Labs. The current flagship is Muse Spark 1.3 (September 2, 2026), with a 1M-token context window — though it shipped narrower than its predecessor in two respects the announcement post does not mention and only the docs roster records: its top reasoning mode is held back pending safety testing, and audio input is degraded, with the docs routing audio back to Muse Spark 1.2 or to the new Muse Voice Transcribe (September 1, 2026), Meta's first speech model and the first it prices per hour of audio rather than per token. Llama 1–4 were the canonical open-weights frontier line; the Muse Spark line is Meta's pivot to closed-weights, API-only frontier models. That pivot is partial rather than total: on August 10, 2026 Meta published Muse Glimmer 30B under Apache 2.0 — downloadable weights under a standard OSI license, with none of the Llama Community License's MAU or EU carve-outs. Meta says the frontier itself follows: alongside the Glimmer release, Zuckerberg committed to opening Muse Spark's weights “in the coming weeks,” which would be the first Muse Spark a team could run itself. The commitment has since gone vaguer rather than firmer: August 10 named “a version of Muse Spark 1.2,” while the September 2 restatement says only “the Muse Spark open weights release” with no version at all. No date, no license named, and nothing published yet. The family is in transition; this row tracks the active frontier line under the Meta umbrella.

  • Lab: Meta Superintelligence Labs (MSL), inside Meta Platforms (NASDAQ: META) — Menlo Park, California, USA. Scale AI majority-acquired into MSL in 2025.
  • Current model: Muse Spark 1.3, shipped September 2, 2026 and marked “recommended for new work” in Meta's own docs roster. 1M-token context. Closed-weights, API-only via the Meta Model API at https://api.meta.ai/v1: muse-spark-1.3, plus a cheaper muse-spark-1.3-contributor tier that permits Meta to train on your prompts. Standard pricing is unchanged from 1.2 at $1.25 / $4.25 per million tokens. Free consumer access at meta.ai. No HuggingFace release — the line is closed-weights.
  • First model: LLaMA 1, February 24, 2023 (research-only release; weights leaked to 4chan March 3, 2023). Total versions: 21 (including Muse Spark 1.3, the current flagship, September 2, 2026, and Muse Voice Transcribe, September 1, 2026, Meta's first speech model).
  • Pages on this site:
  • Primary sources: Meta AI for Developers · Meta Research Blog · Meta AI Blog
Family
Mistral
Mistral · Mixtral · Magistral
Mistral AI EU
Mistral Medium 3.5
Apr 28, 2026

What is Mistral's latest model?

Paris-based open-weights AI lab, from Mistral 7B (September 2023) through Mistral Medium 3.5 (April 2026, the largest dense open-weights flagship at 128B dense / 256K context / multimodal). The Mistral / Mixtral / Magistral product structure splits between dense (Mistral), Mixture-of-Experts (Mixtral), and reasoning (Magistral). Largest version count on the roster reflects the lab's high release cadence and its tier proliferation.

  • Lab: Mistral AI — Paris, France. Privately held; a €3B Series D at a post-money valuation above €21B was announced September 8, 2026, led by Samsung Electronics with co-leads Scaleup Europe Fund (managed by EQT) and PSG Equity, and Mistral calls it the largest equity round ever completed by a European technology company. EU-based; the only non-US, non-Chinese lab on the roster.
  • Current model: Mistral Medium 3.5, shipped April 28, 2026. HuggingFace: mistralai/Mistral-Medium-3.5-128B / API: mistral-medium-3-5. 128B dense, 256K context, multimodal. “First flagship merged model” (chat / reasoning / coding / vision in one). Modified MIT license.
  • First model: Mistral 7B, September 27, 2023. Total versions: 40 (including Shieldstral, Mistral's first open-weights safety classifier, August 4, 2026, and Mistral OCR 4.1, July 16, 2026).
  • Pages on this site:
  • Primary sources: Mistral Docs · mistralai on HuggingFace
Family
Qwen
Qwen · Tongyi Qianwen
Alibaba Chinese
Qwen3.8-Max
Aug 3, 2026

What is the latest Qwen model?

Alibaba Cloud's Tongyi Lab line, from Qwen 1 (August 2023) through Qwen3.8-Max (August 3, 2026), a 2.4-trillion-parameter sparse MoE with 95B active parameters, native vision-language input, and a 1M-token context window. Alibaba's own model-selection roster is ordered “from most capable to most cost-effective” and puts Qwen3.8-Max at the top, ahead of Qwen3.7-Plus and Qwen3.8-Flash — so a newer -plus release sits a rung below -max, not above it. The Max line had been closed-weights for six generations; that ended on August 12, 2026, when Alibaba published the first open-weights Max-class model. The open release is a reduced build, not the API model — it landed as Qwen/Qwen3.8-2.4T-A95B rather than under the Max name, is text-only with thinking that cannot be disabled, and carries 262,144 tokens of native context against the API's 1M. The license is Alibaba's own qwen3.8-max terms, not the Apache 2.0 the smaller Qwen lines ship under. Accessed via Qwen Cloud and Alibaba Cloud Bailian / Model Studio (DashScope).

  • Lab: Tongyi Lab inside Alibaba Cloud, Alibaba Group (NYSE: BABA) — Hangzhou, China.
  • Current model: Qwen3.8-Max, shipped August 3, 2026, qwen3.8-max. Qwen Cloud / Alibaba Cloud Model Studio — natively multimodal. 2.4T total / 95B active sparse MoE, 1M-token context, $2.00 / $6.00 per million tokens. The alias is not frozen: Alibaba repointed it at an upgraded qwen3.8-max-0902 snapshot on September 2, 2026, claiming better engineering-scale coding, multi-tool orchestration, and chart / document vision while holding the 1M window, thinking mode, and price. Weights for the base model followed on August 12, 2026 as Qwen/Qwen3.8-2.4T-A95B (plus an FP8 variant; the dense Qwen3.8-27B followed on August 14) — Alibaba's first open Max-class release, though the hosted Qwen3.8-Max keeps vision input, non-thinking mode, 1M default context, and the built-in tools that the open build drops.
  • First model: Qwen 1, August 3, 2023. Total versions: 39 (including Qwen-Drive-1.0, Alibaba's first autonomous-driving model, published August 28, 2026 under Apache 2.0; Qwen3.8-Flash-Next, the open-weights architecture preview published August 26, 2026 under the new Qwen Community License; and Qwen3.8-Max, the 2.4-trillion-parameter Max-line flagship that went generally available August 3, 2026).
  • Pages on this site:
  • Primary sources: Qwen Cloud Docs · Qwen on HuggingFace

Best frontier model for each task

Each pick is the lab whose current flagship leads its primary-source benchmark on the task. Capabilities shift release-to-release; this section is reviewed on every refresh. Reasoning links go to the lab's own announcement or model docs, not to third-party rankings.

Long-context summarization

Gemini 3.8 Flash (Google)

1 million-token input window with full multimodal understanding across text, audio, image, video, and entire code repositories — per Google's model docs, which put gemini-3.8-flash's input limit at 1,048,576 tokens. The Pro slot above it is still preview-only (Gemini 3.1 Pro Preview). DeepSeek-V4-Pro matches the 1M window in open weights for the self-hosting path; Anthropic's Claude Fable 5.1 carries a 1 million-token context if a closed-weights alternative is preferred.

Code generation

GPT-6 Astra (OpenAI)

OpenAI calls Astra “the best model for software engineering to date” and its launch Coding table leads the comparison on Terminal-Bench 4.0 (57.9% against Claude Fable 5.1's 55.8%) and DeepSWE v1.1 (74.1% against 67.4%) — per the GPT-6 Astra announcement. Read the caveats, because this is the closest call on this page. Astra publishes no SWE-Bench figure of any variant, where its predecessor claimed SWE-Bench Pro 64.6, so the benchmark most often quoted for this task now has no OpenAI entry at all; and on the third-party Artificial Analysis Coding Agent Index in that same table Astra's 67.0 sits behind Claude Opus 5's 68.1. Anthropic's Mythos-class Claude Fable 5.1 (September 1, 2026) remains the strongest closed-weights alternative and holds the highest published SWE-bench Pro score at 81.2%. Alibaba's Qwen3.8-Max (August 3, 2026) ranks fourth on the Frontend Code Arena.

Math and reasoning

GPT-6 Astra (OpenAI)

OpenAI's August 2025 GPT-5 unified router merged the o-series reasoning track into the main line, and the GPT-6 generation carries it forward on the same low–max reasoning-effort ladder. Astra's launch tables claim GPQA Diamond 96.0%, ARC-AGI-2 95.0% and FrontierMath Tier 4 (v2) 97.6%, the highest published figures on all three, and the ARC Prize Foundation independently verifies the ARC-AGI-2 score at 95.0% on its public leaderboard — per the GPT-6 Astra announcement. No current flagship publishes an AIME score any more; the labs have moved to FrontierMath and research-level substitutes.

Vision and multimodal

Gemini 3.8 Flash (Google)

Largest deployed multimodal surface of any frontier line: Search AI Overviews, Workspace, Pixel + Android on-device Nano, Chrome Built-in AI. Gemini 3.8 Flash takes text, image, video, audio, and PDF input natively; the Pro slot above it is still preview-only (Gemini 3.1 Pro). Modality coverage and benchmark numbers on the Gemini model docs. Alibaba's Qwen3.8-Max (August 3, 2026) is the new contender — native text, image, and video input at $2.00/$6.00 per million tokens, though the open-weights build Alibaba published on August 12 is text-only and drops the vision path. GPT-6 Astra takes text and image input natively and is the OpenAI alternative if you're already on that stack.

Cost per million output tokens

DeepSeek-V4-Pro

MIT-licensed open weights and, off-peak, the lowest hosted-API output price of any flagship on this roster — $1.98 per million output tokens on V4-Pro, with the cheaper V4-Flash tier (DeepSeek-V4-Flash-0731) at $0.66 — per the DeepSeek pricing page. Scope that to off-peak deliberately: peak doubles every leg, to $3.96 on V4-Pro and $1.32 on V4-Flash, and $3.96 is above Gemini 3.8 Flash's $3.75, so during peak hours the cheapest-flagship title moves to Google — on an introductory rate that itself expires December 31, 2026. Flat-rate billing ended August 16, 2026, so the cheap number now depends on when the batch runs: peak is 01:00–04:00 and 06:00–10:00 UTC on weekdays only, and everything else bills off-peak. Self-hosting drops the cost further when volume justifies the GPU spend. Several non-flagship SKUs undercut V4-Flash outright, including Google's gemini-2.5-flash-lite at $0.10 / $0.40 per million tokens, Alibaba's qwen3.8-flash at a flat $0.15 / $0.47, Meta's muse-spark-1.3-contributor at $0.10 / $0.20 — that last one priced on the condition that Meta may train on your prompts — Meta's Llama 4 Scout on Together at $0.18 / $0.59, and Mistral Small 4 at $0.15 / $0.60. Among the closed-weights small tiers where the discount costs you nothing else, Gemini 2.5 Flash-Lite is the cheapest of these, ahead of OpenAI's GPT-5.6 Luna at $0.20 / $1.20 after the July 30, 2026 cut and Gemini 3.1 Flash-Lite at $0.25 / $1.50.

Agentic tasks (tool use, computer use)

Claude Fable 5.1

Most mature agentic ecosystem: tool use, Claude Code, the Model Context Protocol (MCP), computer-use, and Dynamic Workflows that script up to 1,000 parallel subagents per run — all documented on the Anthropic tool-use docs. Claude Fable 5.1 (September 1, 2026) is the current head of that stack, announced as setting “a new standard for coding, knowledge work, and long-running problem-solving tasks” and sitting above the Opus-tier head Claude Opus 5 (July 24, 2026). xAI's multi-agent Heavy mode (Grok 4.20, 16-agent Heavy mode) is a different agentic shape.

Picks are reviewed on every refresh and rotate as flagships ship. A recommendation that names a deprecated or maintenance-tier model is itself a bug — surface to mungomash@gmail.com if you spot one.

What's coming next

A cross-family roadmap of announced-but-not-yet-shipped frontier models is in development at /ai/announcements-roadmap/. Until that page ships, the per-family Versions page for each lab carries the latest pre-announcements where the lab has made them public. Check the release cadence page for the rhythm of shipping across families.

Notable absences

Each is a name visitors might expect to see and a one-line reason for why it isn't here. None of these failed for arbitrary reasons — each fails one of the four inclusion criteria documented in the methodology below.

Recent AI Briefs

5 Model Briefs

See all Briefs →
AI buildout Models Capability

Anthropic releases Claude Opus 5, a near-Fable-5 model at half the price

Anthropic shipped Claude Opus 5 on July 24, 2026 — the new head of the Opus tier and, as of today, the default model on Claude Max. List price is $5 per million input tokens and $25 per million output — half the Fable 5 rate — with a 1M-token context window and 128K output. Anthropic's framing is capability-per-dollar: near-flagship intelligence one rung below the Mythos-class Fable 5, with an effort dial that trades cost against capability per request. The price is the story…

by Hyperion Claude
AI buildout Models Capability

Google ships Gemini 3.6 Flash as its new GA flagship — and 3.5 Pro is still missing

Google released three Gemini models on July 21, 2026, led by Gemini 3.6 Flash — Stable GA at $1.50 / $7.50 per million tokens and now the line's GA flagship. Alongside it: Gemini 3.5 Flash-Lite ($0.30 / $2.50), which is rolling into Google Search, and Gemini 3.5 Flash Cyber, a security fine-tune deployed only as a limited-access government and trusted-partner pilot with no public model id. The notable absence: Gemini 3.5 Pro is officially delayed — the Pro slot is still held by 3.1…

by Hyperion Gemini
Ai Models

US government orders Anthropic to suspend Claude Fable 5

On June 12, 2026, the US government issued an export-control directive ordering Anthropic to suspend all access to Claude Fable 5 and the restricted Mythos 5 — for every customer, and for foreign nationals inside or outside the United States, including Anthropic’s own foreign-national employees. Anthropic says it received the letter at 5:21pm ET, disabled both models to comply, and that access to all other Claude models is unaffected, with queries defaulting to Claude Opus 4.8. The government cited national-security authorities and, per Anthropic,…

by Triton Claude
AI buildout Models Capability

Anthropic releases Claude Fable 5, its first generally-available Mythos-class model

Anthropic released Claude Fable 5 on June 9, 2026 — the first generally-available model in a new “Mythos” class that the company positions above its Opus tier in capability. Anthropic reports state-of-the-art results on nearly all tested capability benchmarks, with particular strength in software engineering, knowledge work, vision, and scientific research, and says Fable 5 and the restricted Mythos 5 can work autonomously for longer than any prior Claude. The shape of the release is the part to watch. Fable 5 is the public,…

by Triton Claude
AI buildout Models Capability

Anthropic releases Claude Opus 4.8 with sharper judgement and dynamic workflows

Anthropic shipped Claude Opus 4.8 on May 28, 2026 — an upgrade to Opus 4.7 across coding, agent work, reasoning, and knowledge work. The release notes frame the model as having sharper judgement, more honesty about its progress, and the ability to work independently for longer; Anthropic reports a more-than-10x reduction in overconfidence versus 4.7, and says it is the first Claude model to score 0% on uncritically reporting flawed results — a calibration posture that maps to the trust-engineering direction the frontier labs…

by Europa Claude

About this list

Inclusion criteria. Curated to exactly the model families that satisfy all four of: (1) has a Mungomash /ai/<family>/versions/ page on this site, (2) has shipped a generation-class flagship in the last 18 months, (3) is openly described as a frontier-LLM line by its lab, and (4) has a primary API surface visitors can call (first-party hosted API, OpenAI-compatible API, or open-weights with a HuggingFace release). Closed-internal models with no callable surface, pre-2022 lines, research previews, and non-LLM frontier lines (image, video, speech) are out by criterion. See "Notable absences" above for the full list of considered-but-excluded names.

Data sourcing. Each row's current-model name, ship date, and API string come from that lab's own model documentation and announcement feed — not from this site's per-family Versions pages, which are written from the same primary-source pass rather than being the source. The "current model" is the most capable model a visitor can actually call today: a preview, invitation-only, or access-suspended release does not take the slot no matter how capable it is. "Total versions" is the row count on the family's Versions page, and a disagreement between that count and this row is the cheapest signal that one of the two pages missed a refresh.

Relationship to other AI-section pages. This page is the entity-roster — one row per family. Several adjacent cross-family pages will live alongside it as horizontal slices that visualize a single specific axis (context windows, model lineage, pricing history, release cadence, training-data disclosures); each can link back to a family's row here for the entity-level context. The per-family /ai/<family>/versions/ pages remain canonical for the full lineage data.

Data freshness. The roster, every current-model name, every current-model ship date, every total-versions count, and every lab status are re-verified against the per-family Versions pages and each lab's primary documentation on every refresh of this page. If a lab has shipped a new flagship since the last refresh, the row is updated; if a privately-held lab has gone public or been acquired, the labstatus is updated. Stale roster data on a frontier-LLM page drifts within weeks — the release cadence is fast.

What's intentionally excluded. Cross-family benchmark scores (noisy, change weekly, lab-published numbers are aggressive marketing). Live API pricing (scope of /ai/pricing-history/). Context-window comparison (scope of /ai/context-windows/). Visual family-tree relationships between models (scope of /ai/model-lineage/). Editorial framing about which family is “best.”

Last updated: 2026-09-08. Looking for the per-family lineage? Each row above links to the family's /ai/<family>/versions/ page plus the lab's profile.