2023 – 2026

Grok Versions

xAI's current Grok flagship is Grok 4.6, released August 12, 2026 — a post-training step on Grok 4.5 that keeps the 500,000-token context window and the $2 / $6 per-1M-token headline rates, adds an xhigh reasoning-effort level, and raises cached input from $0.30 to $0.50. xAI's docs recommend it for both code and chat — “the most intelligent and fastest model we’ve built” — superseding Grok 4.5 and Grok 4.3 (both still API-served). The agentic-coding specialist Grok Build 0.1 (grok-build-0.1) and the Composer 2.5 coding model in Grok Build ship alongside. I track every xAI release here — from Grok 1 in November 2023 onward — with API model strings, ship dates, and the major changes per version. Below the table: the 2023 founding, the Musk v. Altman lawsuit, the X integration, the Memphis Colossus cluster, the March 2025 xAI–X merger and the February 2026 SpaceX–xAI merger, and the content-moderation incidents.

Family & status

Family

Chat — the main conversational line, from Grok-2 GA onward
Specialized — coding, vision, and image-generation models
Pre-API — X-only beta or open-weights releases that predate the public xAI API

Status

Current — actively recommended; the latest in its family
Available — still served via API but superseded
Legacy — deprecated, X-only beta that was never API-served, or open-weights only

Grok version table

Model
Grok 4.6
grok-4.6
Chat
Current
Aug 12, 2026
Current xAI flagship — a post-training step on Grok 4.5 aimed at long-running agents and interactive / visual work. Same 500K context and $2 / $6 headline rates; adds an xhigh reasoning-effort level. Cached input is the only price that moved, $0.30 → $0.50.
  • Shipped August 12, 2026. The announcement, “Introducing Grok 4.6,” carries an Aug 12 dateline, and the release notes file it under August directly above the Aug 11 Grok Bot entry. Beware the “August 7” date circulating in aggregator write-ups — Musk publicly targeted Aug 7 in late July and several aggregator sites restated that target as a past-tense launch. xAI's own dated surfaces say Aug 12, and the docs carried no Grok 4.6 anywhere on Aug 11. A third first-party surface now agrees: the SpaceXAI Console changelog files “Grok 4.6 in the playground” — image generation plus “extra high reasoning effort” — under August 12, 2026.
  • There is a model card — it just isn't where xAI's older cards live. Every Grok card through Grok 4.20 sat on data.x.ai under a dated YYYY-MM-DD-<slug>-model-card.pdf filename; the newer ones sit on media.x.ai under opaque hashed names, and the only index that lists them is x.ai/safety. The Grok 4.6 model card (PDF) is cover-dated August 12, 2026 — a fourth first-party surface landing on that ship date — and carries a separate Revision: 2026-08-17 stamp with its own changelog: PartBench and fresh DeepSearchQA results added, KernelBenchInternal moved to v1.1 on a harder task split, and corrected scores on HackerBench v0.2, Self-harm, MASK and LAB. Neither the launch post nor the API docs link it — it is reachable from x.ai/safety and from the September 1, 2026 biosecurity post. And it is more specific than the blog on three things: Grok 4.6 is “the latest release in SpaceXAI’s 1.5T-scale model family,” xAI's own parameter-scale statement for the 4.x line, where the figure on the Grok 4.3 row is third-party reporting; it was “developed in collaboration with Cursor,” with supplemental training on anonymized Cursor workflow data; and its training dates don't line up with the docs — a pretraining data cutoff of January 2026 with supplemental-training data as recent as June 2026, against the February 1, 2026 knowledge cut-off docs.x.ai states (all checked September 8, 2026).
  • A post-training step, not a new base model. xAI is unusually explicit about the continuity: “Grok 4.6 builds on Grok 4.5,” underwent “a longer supplemental training run than Grok 4.5” with curated model-generated reasoning data and an improved optimizer, and used Grok 4.5 itself to regenerate the SFT trajectories across reasoning efforts, agent harnesses, and domains. Agentic RL followed across knowledge work, general coding, kernel optimization, web development, and computer-aided design. None of that appears on the docs surfaces — only in the launch post. What carried over unchanged from Grok 4.5: a 500,000-token context window, text and image in / text out, no stated output limit, and the February 1, 2026 knowledge cutoff the docs list. What didn't: reasoning effort gains a fourth level, low / medium / high (default) / xhigh.
  • Cached input is the only price that changed. Per the Text API table on docs.x.ai/developers/models (checked September 8, 2026): $2.00 input / $0.50 cached / $6.00 output per 1M tokens below a 200K prompt, doubling to $4.00 / $1.00 / $12.00 at or above 200K, billed across the whole request. Grok 4.5's cached rate was $0.30 — so an 85% cache discount became a 75% one while the headline pair held. The launch post also mentions “a fast variant which is twice the price”; four weeks later no such SKU is documented anywhere. It is absent from the Text API pricing table, and grok-4.6-fast returns 404 on its model page. The third check needs a caveat this page previously got wrong: xAI's full-text docs dump has moved. docs.x.ai/llms.txt used to serve roughly 1.4 MB of every docs page concatenated; it is now a 16 KB table of contents containing no model strings at all, and the full text has moved to docs.x.ai/llms-full.txt (~1.5 MB). Grepping the index instead of the dump returns zero hits for every model xAI sells, which would make any “absent from the docs” claim look proven: grok-imagine-image-quality, a model with a live page and a pricing row, scores 0 in llms.txt and 36 in llms-full.txt. Re-run against the real dump, grok-4.6-fast is still zero (all four checked September 8, 2026). Announced, not shipped — or shipped without documentation; the docs cannot tell you which.
  • No aliases. The per-model page lists none — grok-4.5-latest and grok-build-latest still resolve to Grok 4.5, even though the docs name Grok 4.6 as Grok Build's default model. Served in us-east-1 and us-west-2 only; Grok 4.3 remains the one text model with an EU region (checked September 8, 2026).
  • xAI-cited evals at launch, all for Grok 4.6 High: Artificial Analysis Intelligence Index 61, GDPVal-AA v2 1753, CursorBench v3.2 69.9%, DeepSWE v1.1 65.9%, FrontierCode v1.1 (Extended) 61.3%, APEX-Agents 57.5%, Terminal-Bench v3.0 26%, APEX-SWE 56.4%, AA-Briefcase 1577, Harvey LAB (Vals) 15.8%. Vendor-reported, with self-selected comparison columns; treat as one data point. Note that xAI published no SWE-Bench figure this time, having led with SWE-Bench Pro for Grok 4.5.
  • Available from launch day in Cursor and Grok Build (with 2× included usage for the first week), on the xAI API, and through OpenRouter, Vercel, and Cloudflare. The model card's channel list runs longer than the launch post's — it adds Snowflake and Databricks Mosaic, and names Grok 4.6 the default model in the Grok add-ins for Microsoft Word, PowerPoint and Excel — while stating that xAI “plans to add Grok 4.6 to its consumer surfaces (web, mobile apps, and Grok-in-X on the X platform) at a later date,” so as of that card's August 17 revision the flagship was a developer-and-partner model rather than a grok.com one. The third-party surfaces then arrived on roughly a weekly beat, each with its own dated xAI post: GitHub Copilot on August 14, 2026, Amazon Bedrock on August 19, 2026, Google's Gemini Enterprise Agent Platform (via Model Garden) on August 21, 2026, and Microsoft Foundry on August 26, 2026. The Gemini Enterprise post quotes the same $2 / $0.50 / $6 rates as the xAI API. That run has stopped: Microsoft Foundry is still the last Grok 4.6 availability post as of September 8, 2026, four weeks on. Everything newer on x.ai/news is Grok Bot or research — five Grok Bot posts (August 26, August 29, two on September 3 — an enterprise release and a design write-up — and a September 4 write-up of an internal procurement agent) plus a September 1 research post on biosecurity, in which the evaluator LatchBio reports Grok 4.6 refusing disguised hazardous biology tasks more reliably than any other frontier system it tested (59.2% red-team refusal against 64.8% completion of routine tasks, the only model above 50% on both) while sitting mid-pack on pathogen-surveillance work at 53.5%. Third-party evaluation, published by the vendor; one data point.
Model
Grok 4.5
grok-4.5
Chat
Available
Jul 8, 2026
Prior xAI flagship, superseded by Grok 4.6 on August 12, 2026 — and the base it was post-trained from. Configurable reasoning effort (low / medium / high), a 500K-token context window, and $2 / $6 per 1M tokens on the API. Served at ~80 tokens/sec and pitched on token efficiency rather than context size.
  • Shipped July 8, 2026 on docs.x.ai/developers/models and in Cursor. The announcement, “Introducing Grok 4.5,” now carries a July 16, 2026 date and sits at the top of x.ai/news — xAI republished it a week after the model went live, a habit worth knowing about if you date releases off the news feed. For five weeks the docs named it the recommendation for both Code and Chat; since Grok 4.6 landed on August 12, 2026 that sentence names 4.6 instead, and Grok 4.5 has no remaining “which model should I choose?” slot — though it stays in the Text API pricing table at unchanged rates (checked September 8, 2026). The per-model page also picked up a description since the July 29, 2026 check: “SpaceXAI’s intelligent coding model for agentic software, engineering, and workflow tasks. Trained in SpaceXAI’s data centers in Memphis with new datasets spanning science, engineering, and math.” A Grok 4.5 model card exists as well — listed on x.ai/safety, cover-dated July 14, 2026 with a Revision: 2026-07-20 stamp. That is a third distinct date for one release, sitting between the docs appearance and the republished announcement, and the reason this row anchors the ship date to when the model actually became callable rather than to whatever dateline a given surface happens to show (checked September 8, 2026).
  • Configurable reasoning effort — a low / medium / high setting (default high) on a single model string, collapsing the separate reasoning / non-reasoning SKUs of the Grok 4.x line into one grok-4.5 id.
  • 500,000-token context window; API pricing $2 / $6 per million input / output tokens ($0.30 cached) on prompts under 200K tokens, doubling to $4 / $12 ($0.60 cached) at or above 200K, per the Text API pricing table on docs.x.ai/developers/models. The per-model page now carries the same tiered table — it showed a flat $0.50 cached-input rate as recently as July 29, 2026 and reads $0.30 / $0.60 as of July 31, 2026, so xAI's two pricing surfaces finally agree on that number. Aliases grok-4.5-latest and grok-build-latest both resolve here. Knowledge cutoff February 1, 2026, well past the November 2024 cutoff the Grok 4 family carried.
  • xAI-cited coding / agentic benchmarks at launch: Terminal-Bench 2.1 83.3%, SWE-Bench Pro 64.7% resolve rate, DeepSWE 1.0 62.0%, SWE Marathon 29.0%. Vendor-reported; treat as one data point.
  • Trained across tens of thousands of NVIDIA GB300 GPUs on curated coding / science / engineering / math data, with asynchronous large-scale RL over “hundreds of thousands of tasks” centered on multi-step software engineering — and trained alongside Cursor — whose parent Anysphere SpaceX agreed on June 16, 2026 to acquire for $60 billion in stock, weeks before this launch. Served at ~80 tokens/sec; xAI measures ~4.2× fewer output tokens than Opus 4.8 (max) on SWE-Bench Pro (vendor-reported).
  • Ships with native Microsoft Office (Excel / PowerPoint / Word) plugins and was the default model in Grok Build from the July 16, 2026 launch post until Grok 4.6 took that slot on August 12, 2026, alongside Cursor on all plans and the API console. It reached grok.com, X, iOS and Android on July 22, 2026 and GitHub Copilot on July 28, 2026.
  • Reachable from the EU, but not served from an EU region. The release notes now carry a “Grok 4.5 available in the EU” entry — “Grok 4.5 is now available in the API console for EU users” — while the model's own docs page still lists grok-4.5 in us-east-1 and us-west-2 only (both surfaces checked September 8, 2026). Read together: EU accounts can call the model, but the inference is not run on EU infrastructure — which matters if your constraint is data residency rather than access. Grok 4.3 (us-east-1 / eu-west-1 / us-west-2) is still the only text model in the lineup with an EU region; grok-4.6, grok-build-0.1 and all three Grok 4.20 SKUs sit at us-east-1 / us-west-2, unchanged since July 31, 2026 after a week of daily flapping in both directions before that. The release notes still carry no EU-console entry for Grok 4.6 — “Grok 4.5 available in the EU” is the only EU line anywhere in the notes, and it has stood alone for the four weeks since Grok 4.6 shipped (re-read September 8, 2026). Regions move silently and fast on this docs surface; treat any regional claim as good only for the date next to it.
Model
Composer 2.5
Composer 2.5 (Grok Build /model menu)
Specialized
Current
Jun 1, 2026
Fast agentic-coding model selectable inside Grok Build, built for long-running tasks and complex instruction following. Built on the open-weights Kimi K2.5 checkpoint; trained with ~25× more synthetic tasks than Composer 2. $0.50 / $2.50 per 1M input/output tokens.
  • Announced June 1, 2026; the announcement is at x.ai/news/composer-2-5. Available to SuperGrok and X Premium+ subscribers; selected inside Grok Build by typing /model and choosing Composer 2.5.
  • Agentic-coding focus — positioned for long-running, multi-step tasks and complex instruction following, with strengths in coding, agents, JSON, and tool use (function calling, tool use, code execution).
  • Built on the open-source Kimi K2.5 checkpoint (Moonshot AI), then post-trained with roughly 25× more synthetic agentic tasks than Composer 2.
  • Pricing $0.50 / $2.50 per million input / output tokens, with a faster variant at $3.00 / $15.00 — roughly half the input cost of the default Grok Build model.
  • Coexists with Grok Build 0.1 in the Grok Build /model menu rather than replacing it. Grok 4.5 became Grok Build's default model with the July 16, 2026 launch post and Grok 4.6 took over on August 12, 2026; Composer 2.5 remains an explicit /model pick throughout.
Model
Grok Build 0.1
grok-build-0.1
Specialized
Current
May 19, 2026
Purpose-built coding model trained for agentic workflows — multi-step actions over single-turn Q&A. Successor to grok-code-fast-1. 256K context, $1 / $2 per million input/output tokens.
  • The Grok Build coding CLI / TUI launched in beta on May 14, 2026; the grok-build-0.1 model itself landed in early access on May 19, 2026 per the release notes, and reached the xAI API in public beta on May 29, 2026 (x.ai/news/grok-build-0-1).
  • Agentic-coding focus — trained specifically for multi-step coding workflows (apply patches, run tests, iterate) rather than single-turn code generation.
  • API pricing $1 / $2 per million input / output tokens (256K context) per docs.x.ai/developers/models — positioned aggressively against the deprecated grok-code-fast-1 and against Anthropic's Sonnet-tier coding usage.
  • Companion announcement on May 18, 2026: Grok Skills, a persistent custom-expertise feature that carries across conversations — a long-term memory layer for repeat workflows. Surfaced in the consumer Grok app and on the API.
  • Same Memphis Colossus / Colossus 2 training infrastructure as Grok 4.3; not a derivative of any earlier Grok line, but trained from a new code-and-agentic data mix.
  • The Grok Build harness was open-sourced July 15, 2026 — agent loop, tools, terminal UI, and the skills / plugins / hooks / MCP / subagent extension system, at github.com/xai-org/grok-build (announcement). The CLI can now be compiled and pointed at local inference. The harness is open; the grok-build-0.1 weights are not — Grok 1 remains the only Grok model released as open weights.
  • Grok Build's default model has moved twice since: to Grok 4.5 with the July 16, 2026 launch post, and to Grok 4.6 on August 12, 2026. The grok-build-latest alias, though, still resolves to grok-4.5 rather than following the default forward — a documented mismatch worth knowing if you pin to it — unchanged for twenty-seven days now (checked September 8, 2026). The docs' “Which model should I choose?” block now points Code at Grok 4.6, while grok-build-0.1 stays in the Text API pricing table at $1 / $2 per 1M. Its model page dropped the “early access” label between the July 29 and July 31, 2026 checks — it now reads “SpaceXAI’s intelligent coding model for agentic software, engineering, and workflow tasks” with no lifecycle qualifier, though the May entry in the release notes still says “Currently in early access.” The retired grok-code-fast-1, grok-code-fast, and grok-code-fast-1-0825 ids all alias here.
Model
Grok 4.3
grok-4.3
Chat
Available
Apr 17, 2026
Prior xAI chat flagship, superseded by Grok 4.5 on July 8, 2026. Native video input, in-chat generation of PDFs / slides / spreadsheets. Retains Grok 4.20's 16-agent Heavy mode and 1M-token context. $1.25 / $2.50 per 1M tokens on the API.
  • Beta rolled out April 17, 2026 to SuperGrok Heavy ($300/mo) on grok.com and the iOS / Android apps, surfaced silently in the model selector before any named announcement. API general availability began May 1, 2026; on the May 15, 2026 docs refresh, grok-4.3 is the canonical chat-model alias on docs.x.ai/developers/models — "Strong agentic tool calling with minimal hallucinations. Supports non-reasoning mode."
  • Native video understanding — first Grok generation with multimodal video input alongside images and text.
  • Document and artifact generation in chat — can produce PDFs, PowerPoint decks, and spreadsheets directly inside the conversation.
  • Retains the 16-agent Heavy mode; context window listed at 1,000,000 tokens on the May 15, 2026 docs revision (down from the 2M figure xAI cited around the April 17, 2026 beta launch — the pricing-page SKUs all read 1M now).
  • Reported launch parameter count ~0.5T, with a 1T checkpoint completing training; xAI has not published a Grok 4.3 model card as of mid-May 2026.
  • Companion API releases the same window: Speech-to-Text API (GA; 25 languages, diarization, batch + streaming) and Text-to-Speech API (GA; $15.00 per million characters, same voice stack as Grok Voice and Tesla vehicles).
  • The May 15, 2026 retirement window sunsets eight slugs, per the migration guide: grok-4-1-fast-reasoning, grok-4-1-fast-non-reasoning, grok-4-fast-reasoning, grok-4-fast-non-reasoning, grok-4-0709, grok-3, grok-code-fast-1, and grok-imagine-image-pro. The slugs keep resolving rather than erroring — the reasoning ones redirect to grok-4.3 at low effort and the non-reasoning ones at none, grok-code-fast-1 to Grok Build 0.1, and grok-imagine-image-pro to grok-imagine-image-quality — but they bill at the destination model's rates (re-read September 8, 2026).
Model
Grok 4.20
grok-4.20-multi-agent-0309, grok-4.20-0309-reasoning, grok-4.20-0309-non-reasoning
Chat
Available
Mar 10, 2026
Prior xAI flagship. Multi-agent collaboration as a first-class API mode. Successor to Grok 4 Heavy. 1M context, split into three 0309-dated SKUs (multi-agent / reasoning / non-reasoning).
  • Beta announced February 17, 2026 (the one-year anniversary of Grok 3); full API general availability on March 10, 2026 at the same time as the multi-agent variant. Currently shipped on the pricing page as three explicit SKUs all dated 0309 — grok-4.20-multi-agent-0309, grok-4.20-0309-reasoning, grok-4.20-0309-non-reasoning — per docs.x.ai/developers/pricing. System card at data.x.ai (PDF), published April 7, 2026.
  • Multi-agent collaboration — the multi-agent SKU runs multiple agents in parallel and cross-evaluates outputs, generalizing the Grok 4 Heavy concept into a routine API surface rather than a $300/mo consumer tier.
  • 1,000,000-token context window across all three SKUs at the current pricing page; $1.25 / $0.20 (cached) / $2.50 per 1M input / output tokens. The logprobs / top_logprobs fields are no longer supported on Grok 4.20 or newer.
  • Selectable in the consumer Grok app's Auto, Fast, Expert, and Heavy modes from approximately March 18, 2026 onward.
  • Knowledge cutoff carried over from the Grok 4 family (November 2024); marketed by xAI as “the most intelligent and fastest model we've built.”
  • See docs.x.ai/developers/models for the canonical current model strings.
Model
Grok 4.1 Fast
grok-4-1-fast-reasoning, grok-4-1-fast-non-reasoning
Chat
Legacy
Nov 19, 2025
Faster, cheaper Grok 4.1 with a 2M-token context window and the new Agent Tools API. Internal codenames: quasarflux (reasoning) and tensor (non-reasoning).
  • Released November 19, 2025, two days after Grok 4.1; the announcement is at x.ai/news/grok-4-1-fast.
  • 2,000,000-token context window across both reasoning and non-reasoning sizes — the first Grok generation to ship at full 2M context.
  • Agent Tools API debuted alongside this release — a built-in orchestrator for web search, browsing, and code execution that the model invokes inside its chain-of-thought.
  • The two SKUs were live-traffic A/B tested on grok.com under the codenames quasarflux and tensor in early November 2025 before the named launch.
  • Cheaper per token than Grok 4.1 at comparable quality on most benchmarks; the routine recommendation for high-volume agent workloads.
Model
Grok 4.1
grok-4-1
Chat
Legacy
Nov 17, 2025
Conversational refresh of Grok 4. Reduced hallucinations, improved “emotional understanding,” #1 on LMArena Text Arena at launch.
  • Released November 17, 2025; the announcement is at x.ai/news/grok-4-1. Model card at data.x.ai (PDF).
  • Two-week silent live-traffic A/B from November 1–14, 2025 surfaced the model on grok.com under the codenames quasarflux (thinking) and tensor (non-reasoning) before the named launch.
  • #1 on LMArena Text Arena at launch — Grok 4.1 Thinking at 1483 Elo; non-reasoning at 1465 Elo (#2). Vendor-cited result; treat as one data point.
  • Free for all users on grok.com, X, and the iOS / Android apps from launch.
  • Substantially fewer factual hallucinations than Grok 4 on xAI's internal evaluations; positioning emphasized conversational quality over raw frontier benchmarks.
  • Gone from the docs with no retirement notice at all. grok-4-1 is absent from the Text API pricing table, its per-model page returns 404, and the bare string appears nowhere in xAI's full documentation dump — only the unrelated grok-4-1-fast-* ids do. That dump now lives at docs.x.ai/llms-full.txt; docs.x.ai/llms.txt, which this row used to cite, has been cut down to a link index that contains no model strings at all, so it can no longer tell you whether anything is in the docs (re-derived September 8, 2026 against the full dump: seven matches for grok-4-1, every one of them the -fast prefix). It is also not on the May 15, 2026 retirement list, so the redirect-to-grok-4.3 guarantee that keeps those eight slugs resolving does not cover it. Where grok-4-1 resolves today, if anywhere, is undocumented — a different shape from the announced May 15 cohort, and the reason this row carries no retirement date. Worth knowing before reading anything into a 404 here: a docs page 404 tracks whether a slug is still named as an alias somewhere, not whether the id still answers. grok-code-fast-1 and grok-imagine-image-pro were retired on the same May 15 list and their pages still return 200 — each renders its destination model's page, because each is still listed among that model's aliases.
Model
Grok 4 Fast
grok-4-fast-reasoning, grok-4-fast-non-reasoning
Chat
Legacy
Sep 19, 2025
Faster, cheaper Grok 4 with a 2M-token context window. ~40% fewer thinking tokens at comparable quality.
  • Released September 19, 2025; the announcement is at x.ai/news/grok-4-fast. Model card at data.x.ai (PDF).
  • 2,000,000-token context window across both SKUs — the first Grok generation to ship at 2M context, three months ahead of Grok 4.1 Fast.
  • Roughly 40% fewer thinking tokens than Grok 4 at comparable quality on xAI's internal evaluations; the routine recommendation for high-volume reasoning workloads.
  • Tiered pricing — ~$0.20 / 1M input tokens under 128K context, ~$0.40 / 1M above; output and cache pricing scaled accordingly.
  • Available on grok.com Fast and Auto modes alongside the API release.
Model
grok-code-fast-1
grok-code-fast-1
Specialized
Legacy
Aug 28, 2025
First xAI model purpose-built for agentic coding. New from-scratch architecture. Pre-release codename: sonic.
  • Released August 28, 2025; the announcement is at x.ai/news/grok-code-fast-1. Model card at data.x.ai (PDF).
  • From-scratch architecture optimized for agentic coding rather than a fine-tune of Grok 4 — visible reasoning trace, tuned for tool use inside an editor agent loop.
  • Strong on TypeScript, Python, Java, Rust, C++, and Go; pre-release stealth codename sonic in the third-party editor integrations during testing.
  • Pricing $0.20 / $1.50 per million input / output tokens; $0.02 / 1M cached input.
  • Free at launch on GitHub Copilot, Cursor, Cline, Kilo Code, Roo Code, opencode, and Windsurf for an introductory window.
Model
Grok 4 / Grok 4 Heavy
grok-4-0709, grok-4
Chat
Legacy
Jul 9, 2025
First xAI model to clear 50% on Humanity's Last Exam. Native tool use, parallel tool calling. Heavy variant on the $300/mo SuperGrok Heavy tier.
  • Released July 9, 2025 in a livestream announcement; the post is at x.ai/news/grok-4. Model card at data.x.ai (PDF). The launch came one day after the “MechaHitler” content-moderation incident covered in the prose history below.
  • 256,000-token context window at launch; pricing $3 / $15 per million input / output tokens.
  • Native tool use and parallel tool calling — the model invokes web search, browsing, and code-execution tools inside its reasoning trace; tool calls are parallelizable across a single response.
  • Grok 4 Heavy — a multi-agent variant that runs several Grok 4 instances in parallel and cross-evaluates outputs, available only on the SuperGrok Heavy tier at $300/month. First model publicly reported to clear 50% on Humanity's Last Exam (text-only subset, with tools).
  • Trained on the Memphis Colossus cluster with large-scale reinforcement learning; xAI cited a roughly tenfold compute increase over Grok 3 (vendor-stated).
Model
Grok 3
grok-3, grok-3-fast, grok-3-mini, grok-3-mini-fast
Chat
Legacy
Feb 17, 2025
First Grok with reasoning as a first-class capability. “Think” and “Big Brain” modes. DeepSearch agentic web search. grok.com standalone app launches.
  • Beta announced February 17, 2025 (“The Age of Reasoning Agents”); full API general availability rolled out through April 2025 starting with grok-3-fast. The announcement is at x.ai/news/grok-3.
  • Reasoning as a first-class mode — “Think” toggles chain-of-thought reasoning, “Big Brain” spends more thinking tokens on the hardest problems. Mirrors the OpenAI o-series pattern that arrived five months earlier.
  • 131,072-token context window across the family. Four SKUs: grok-3 (flagship), grok-3-fast (same model, faster infrastructure), grok-3-mini, and grok-3-mini-fast with a reasoning_effort parameter (low / high).
  • Trained on the Memphis Colossus cluster on roughly ten times the compute of Grok 2 by xAI's accounting (vendor-stated).
  • DeepSearch agentic web search debuted in the consumer Grok UI alongside the model launch; voice mode followed within weeks.
  • grok.com launched as the standalone Grok consumer surface alongside this release, separating Grok from the X feed.

The reasoning era — February 17, 2025. Above this line: the Grok 3 and Grok 4 generations, where reasoning, native tool use, and multi-agent collaboration became first-class capabilities and the consumer surface moved off X to its own grok.com app. Below: the original X-integrated chat line from Grok 1 through Grok 2, where Grok was a Premium / Premium+ subscriber benefit on X rather than a standalone product.

Model
Grok-2 / Grok-2 Vision (GA)
grok-2-1212, grok-2-vision-1212
Chat
Legacy
Dec 14, 2024
Production-grade Grok-2 hits the xAI API with a stable model string and the December pricing reset to $2 / $10 per 1M tokens. Vision sibling at 32K context. Retired: no longer served as of the July 2026 docs.
  • Released December 14, 2024 as the GA replacement for the November 2024 grok-beta / grok-vision-beta ids; the December pricing reset to $2 / $10 per million input / output tokens stuck for the rest of the Grok-2 line.
  • 128,000-token context window on the chat model; 32K on grok-2-vision-1212.
  • Improved instruction-following, multilingual capability, and accuracy on xAI's internal evaluations vs. the August 2024 X-only beta.
  • Shipped alongside grok-2-image-1212 (the Aurora image API; see the row immediately below).
  • Superseded as the recommended chat model by Grok 3 in February 2025.
  • Retired. grok-2-1212 and grok-2-vision-1212 are no longer listed among the served models on docs.x.ai/developers/models, and their per-model doc pages now resolve to “Model not found” (checked July 15, 2026). Unlike the May 15, 2026 retirement window, xAI published no formal retirement notice for the Grok-2 line, so no exact shutdown date is recorded here.
Model
Aurora (image)
grok-2-image-1212
Specialized
Legacy
Dec 9, 2024
First in-house xAI image model. Autoregressive Mixture-of-Experts, not diffusion. Replaces the Black Forest Labs FLUX.1 integration that had been live on X since August 2024. Retired: superseded by the Grok Imagine image models.
  • Announced December 7–9, 2024 (“Grok Image Generation Release”); the post is at x.ai/news/grok-image-generation-release.
  • Autoregressive Mixture-of-Experts trained to predict the next token across interleaved text and image data — not a diffusion model. xAI's first in-house image stack; replaced the Black Forest Labs FLUX.1 integration that had been the image surface on X since August 2024.
  • Photorealistic rendering, accurate text-in-image, native image-to-image edits via multimodal input.
  • Exposed as the grok-2-image-1212 API id at $0.07 per generated image; on X, available to all users at launch with rate limits.
  • Permissive content rules at launch (public-figure rendering, more permissive IP handling than the prior FLUX integration) drove most of the press coverage in the launch week.
  • Retired. grok-2-image-1212 is no longer listed among the served models on docs.x.ai/developers/models, and its per-model doc page now resolves to “Model not found” (checked July 15, 2026). The image surface is now the Grok Imagine line — grok-imagine-image and grok-imagine-image-quality. As with the rest of the Grok-2 line, xAI published no formal retirement notice, so no exact shutdown date is recorded here.

The xAI API opens — November 4, 2024. Above this line: the API era, where Grok models have public model strings on docs.x.ai and developers can call them outside the X platform. Below: the Premium / Premium+ era, where Grok was an X-platform subscriber benefit only and the early models had no API surface (Grok 1 was open-sourced after-the-fact in March 2024; Grok 1.5 and 1.5V never had API model strings at all).

Model
Grok-2 / Grok-2 mini (beta)
grok-beta, grok-vision-beta — later sunset
Pre-API
Legacy
Aug 13, 2024
First Grok with image generation on X (via Black Forest Labs FLUX.1). The API public beta followed in November 2024 under the temporary ids.
  • Beta launched on X for Premium / Premium+ subscribers August 13, 2024; the announcement is at x.ai/news/grok-2.
  • First Grok with image generation integrated on X — via the Black Forest Labs FLUX.1 model rather than an in-house stack (Aurora replaced FLUX.1 four months later).
  • The xAI public API beta launched November 4, 2024 with the temporary grok-beta (text, 128K context) and grok-vision-beta (multimodal) ids; the beta API offered a $25 / month credit for an introductory window. The announcement is at x.ai/news/api.
  • Around the same time (December 6–7, 2024), Grok became free to all X users with a 10-prompts-per-2-hours rate limit, leaving Premium as the upgrade for higher limits and image generation.
  • Both beta API ids were retired in favor of the December 14, 2024 GA grok-2-1212 family. The August beta is treated as legacy here; the December GA is the row above this divider.
Model
Grok-1.5V
never publicly released
Pre-API
Legacy
Apr 12, 2024
First xAI multimodal model. Documents, diagrams, charts, screenshots, photographs. Stayed in early-tester preview — never publicly shipped.
  • Announced April 12, 2024; the post is at x.ai/news/grok-1.5v.
  • First xAI multimodal model — could process documents, diagrams, charts, screenshots, and photographs alongside text input.
  • Introduced the RealWorldQA benchmark for spatial-understanding evaluation alongside the model.
  • Stayed in early-tester preview; was never publicly released to X subscribers or the API. The vision capability that did ship in production landed eight months later as grok-2-vision-1212.
Model
Grok-1.5
never publicly released
Pre-API
Legacy
Mar 28, 2024
128K context window. Stronger reasoning, coding, and math vs. Grok 1. Rolled out to all X Premium users May 15, 2024; never had an API model string.
  • Announced March 28, 2024; the post is at x.ai/news/grok-1.5. Rolled out to all X Premium users on May 15, 2024.
  • 128,000-token context window — sixteen times Grok 1's 8K. Perfect retrieval on Needle-In-A-Haystack at the full 128K per xAI's evaluation.
  • Substantial gains over Grok 1 on HumanEval (coding) and MATH (competition math) by xAI's accounting; the framing emphasized reasoning and coding for the developer audience even though there was no developer API yet.
  • Available exclusively on X for Premium subscribers; never had a public xAI API model string. Superseded by Grok 2 in August 2024.
Model
Grok 1
open-weights only — github.com/xai-org/grok-1
Pre-API
Legacy
Nov 4, 2023
The original Grok. 314B-parameter Mixture-of-Experts. 8K context. Open-sourced under Apache 2.0 in March 2024 — the only Grok generation released as open weights.
  • Announced November 4, 2023 (“Announcing Grok”); the post is at x.ai/news/grok. Initial early-access beta on X for US Premium subscribers; rolled out to all US Premium+ on December 8, 2023, and to all X Premium worldwide on March 26, 2024.
  • 314 billion parameters, Mixture-of-Experts with 8 experts and top-2 routing (~25% of weights active per token); 64 layers, embed dim 6,144, SentencePiece tokenizer with 131K vocab. 8,192-token context window.
  • Pretraining concluded October 2023; trained on JAX + Rust infrastructure rather than PyTorch.
  • Open-sourced March 17, 2024 under Apache 2.0 — the only Grok generation released as open weights. The repo is at github.com/xai-org/grok-1; the announcement is at x.ai/news/grok-os. Grok 1.5, 2, 3, 4, and 4.20 have not been released as open weights through the time of this writing, despite Musk's earlier statement that the previous generation would be open-sourced when each new one shipped.
  • Never served via the public xAI API; the open-weights checkpoint on GitHub is the only way to run Grok 1 today.

Click any row to expand. Each row has a stable id for sharing — e.g. /ai/grok/versions/#grok-4-20, #grok-4, #grok-1. The current model list is at docs.x.ai/developers/models; release notes at docs.x.ai/developers/release-notes.

The 2023 founding

xAI was incorporated in Nevada on March 9, 2023 and publicly announced on July 12, 2023. Elon Musk founded the company and has been CEO from incorporation through today. The launch team of roughly a dozen included Igor Babuschkin (chief engineer, prior at DeepMind and OpenAI), Greg Yang, Christian Szegedy, Yuhuai “Tony” Wu, Jimmy Ba, Kyle Kosic, Manuel Kroiss, Ross Nordeen, Toby Pohlen, Guodong Zhang, and Zihang Dai, drawn primarily from DeepMind, OpenAI, Microsoft Research, Google Research, Tesla, and the University of Toronto.

The stated mission was “to understand the true nature of the universe” under the framing of “maximally truth-seeking” AI — a positioning explicitly contrasted with what Musk argued OpenAI had become since his 2018 departure from its board. Most of the original cofounders left through 2025 and early 2026; by April 2026, Musk was the sole remaining cofounder of the original launch team.

Musk v. Altman — the OpenAI lawsuit

Musk co-founded OpenAI in 2015, left its board in February 2018, and sued OpenAI, Sam Altman, and Greg Brockman on February 29, 2024 in San Francisco Superior Court alleging breach of the founding agreement. He refiled the case in U.S. District Court for the Northern District of California (Oakland) on August 5, 2024 as Musk v. Altman et al., case number 4:24-cv-04722, before Judge Yvonne Gonzalez Rogers. The federal complaint added RICO and other claims; the original complaint can be read at the docket-uploaded PDF.

Judge Gonzalez Rogers denied a preliminary injunction in March 2025, calling Musk's irreparable-harm theory “a stretch.” The defendants moved for summary judgment in October 2025; that motion narrowed the case to two surviving claims at trial: breach of charitable trust and unjust enrichment. Musk's lawyers have requested up to $134 billion in alleged wrongful gains and an unwinding of OpenAI's October 2025 conversion to a public benefit corporation.

The trial opened in Oakland the last week of April 2026. A nine-person advisory jury was seated April 27, 2026; opening arguments started April 28, 2026. Under the structure Judge Gonzalez Rogers set, the jury's verdict is advisory and the judge will make the final calls on liability and remedies; each side had 22 hours of trial time, with Microsoft (a defendant in interest) allotted 5 hours. Sam Altman testified May 11–12, 2026 — describing Musk's demand for 90% ownership early in OpenAI's history and characterizing the xAI lawsuit as “business interference” against a competitor. Closing arguments concluded May 14, 2026; Musk's ask in closings was up to $150 billion “disgorged” to the nonprofit entity, removal of Altman and Brockman from leadership, and an unwinding of the for-profit conversion. On May 18, 2026, after under two hours of deliberation, the advisory jury found for OpenAI — ruling Musk's claims time-barred by the three-year statute of limitations rather than reaching the merits of the charitable-trust and unjust-enrichment theories. Judge Gonzalez Rogers entered a Post-Trial Order Re: Advisory Verdict on May 20, 2026 formalizing the ruling; Musk had filed his response on the Rule 52 findings-of-fact requirements the day before, on May 19, 2026. The case did not end there. The docket records a June 10, 2026 order setting a joint briefing schedule on the remaining claims, the May 26 amicus motion denied June 29, 2026, a July 1, 2026 order appointing a mediator and conditionally appointing a special master, and a July 2, 2026 order setting Musk's answer to OpenAI's counterclaims for August 10, 2026. That answer landed on the deadline: docket entry 609, Answer to Counterclaim, filed August 10, 2026. The docket then went quiet after a small cluster of housekeeping on August 11, 2026 — entries 610 through 613, four notices of withdrawal by Wachtell, Lipton attorneys (William Savitt, Sarah Eddy, Randall Jackson, Bradley Wilson) stepping off the OpenAI defendants' representation. Those remain the most recent filings on the public RECAP copy, which still records the last known filing as August 11 (checked September 8, 2026) — a twenty-eight-day gap. How much of that is a real lull is hard to say, because the archive has stopped keeping up: RECAP's copy of this docket has not been refreshed since August 27, 2026, so it is now twelve days stale and the quiet stretch is only positively confirmed through that date. This docket was being mirrored same-day through mid-August; it has not been for the past two weeks. Musk called the verdict a “calendar technicality” and said he would appeal to the Ninth Circuit, but more than three months on, no notice of appeal appears on the docket and no contemporaneous reporting places one there. The case is the throughline that connects xAI's founding to OpenAI's; verify the current docket status at CourtListener.

The X integration

Grok shipped first as a feature of X Premium and Premium+ rather than a standalone product. The early-access beta arrived for US Premium subscribers on November 4, 2023; access expanded to all US Premium+ on December 8, 2023, and to all X Premium worldwide on March 26, 2024. Grok-1.5 reached all X Premium users on May 15, 2024; Grok-2 launched in beta on X on August 13, 2024 with image generation via Black Forest Labs' FLUX.1.

On December 6–7, 2024, Grok became free to all X users with a 10-prompts-per-2-hours rate limit, leaving Premium as the upgrade for higher limits, image generation, and (later) reasoning modes. The standalone grok.com consumer app launched alongside Grok 3 in February 2025, separating the chatbot from the X feed for the first time. Native iOS and Android apps followed.

The corporate side caught up in March 2025 when xAI acquired X Corp. in an all-stock deal that valued xAI at $80 billion and X at $33 billion (combined ~$113 billion). The merger formalized what had been a tight operational partnership and gave xAI direct ownership of the X platform that had hosted Grok since launch.

Memphis Colossus

xAI built its primary training cluster, Colossus, in a former Electrolux factory in South Memphis. Construction began in spring 2024; the first 100,000 NVIDIA H100 GPUs came online in July 2024 (Supermicro CEO Charles Liang publicly stated the build took 122 days), and xAI formally introduced the cluster as Colossus in September 2024. The disclosed scale then doubled to 200,000 GPUs — the company's own x.ai/colossus page dates that full-scale milestone to February 2025, 276 days from groundbreaking, and still states a long-term target of one million GPUs (checked September 8, 2026). Those two figures are the only GPU counts the company has ever formally published for the cluster — though note that the same page's “by the numbers” block reads 180K against its own 200,000 headline, an inconsistency xAI has left standing and has never explained.

The build-out has been the subject of an ongoing environmental dispute. xAI ran dozens of methane gas turbines on-site to power the cluster; the Southern Environmental Law Center filed multiple complaints alleging permit violations. The Memphis health department issued a permit in July 2025 covering 15 turbines for 24/7 operation through 2030. In April 2026 the NAACP, represented by the Southern Environmental Law Center and Earthjustice, filed a Clean Air Act suit over unpermitted turbines powering the Colossus 2 site in Southaven, Mississippi; on June 16, 2026 the U.S. Department of Justice moved to intervene and dismiss the case, arguing that cutting the cluster's power would harm national security. Reporting through this period has cited substantially larger 2026 GPU counts and the Colossus 2 expansion; those counts are from secondary analysts, not formal xAI disclosure, and are not relied on here.

The SpaceX–xAI merger (February 2026)

On February 3, 2026, SpaceX announced an all-stock acquisition of xAI in what was reported as the largest merger in tech history: each xAI share converted at a 0.1433 ratio into SpaceX equity, valuing xAI at $250 billion against SpaceX's $1 trillion for a combined $1.25 trillion entity. Under the new structure, xAI operates as a SpaceX subsidiary; Michael Nicolls, previously a Starlink VP, became xAI's president following the April 10, 2026 staff restructuring that also saw CFO Anthony Armstrong depart. Musk remains CEO of xAI and of SpaceX. The corporate structure as of this writing is xAI inside SpaceX, with the X social platform held under xAI from the March 2025 merger. The combined entity kept acquiring: on June 16, 2026, SpaceX agreed to acquire Anysphere, the company behind the AI code editor Cursor, for $60 billion in an all-stock deal expected to close in the third quarter of 2026 pending regulatory approvals — and Grok 4.5, launched three weeks later, was trained alongside Cursor.

On July 6, 2026, the company completed a rebrand to SpaceXAI, changing its X handle to @spacexai and publishing a new logo; Musk had said in May 2026 that xAI would be dissolved as a separate company and folded into SpaceX under the SpaceXAI brand. The rollout across the company's own surfaces is now complete: as of September 8, 2026 x.ai/news, x.ai/company and x.ai/colossus all read “SpaceXAI” and point at the @spacexai handle, and docs.x.ai titles itself “SpaceXAI Docs.” Even the legal-entity string has now followed the brand: the marketing-site footer read “© 2026 X.AI LLC” on August 21, 2026, changed to “© 2026 SpaceXAI LLC” a week later, and still reads that on all three pages on September 8, 2026 — the last piece of the rebrand to move. (docs.x.ai carries no copyright line at all, so it is not part of that comparison.) The footer overstates it, though. The Grok 4.6 model card footnotes that “SpaceXAI is a doing-business-as (DBA) name of XAI LLC” and that “xAI and SpaceXAI may be used interchangeably” — so this is a trade name layered over an unchanged legal entity, not a renamed corporation, and the two names are still the same company on paper. The company page was the last page-level holdout — it still read “xAI,” still carried “© 2026 xAI Corp.” and the @xai handle, and described the company as “a division of SpaceX,” as recently as July 31, 2026, a month after the rebrand itself. The developer console ran ahead of the marketing site rather than behind it: the Console changelog dates “SpaceXAI rebrand and sign-in redesign” — new logo and favicons across the console — to June 16, 2026, three weeks before the public switch, though that entry was backfilled when the changelog launched on August 13, 2026 and is xAI's own retrospective dating rather than a contemporaneous post. The x.ai and api.x.ai domains and the XAI_API_KEY convention are unchanged. Grok keeps its name — the rebrand renamed the company, not the model line, and no model string, endpoint, or price changed with it. This page continues to say “xAI” for the company across the pre-rebrand history it describes.

The reasoning era and the Grok 4 family

Grok 3 in February 2025 introduced reasoning as a first-class capability on the Grok line, mirroring the OpenAI o-series pattern that had arrived five months earlier. xAI shipped four SKUs (grok-3, grok-3-fast, grok-3-mini, grok-3-mini-fast) with a reasoning_effort parameter, the consumer-facing Think and Big Brain modes, the DeepSearch agentic web-search tool, and the standalone grok.com app.

Grok 4 in July 2025 was the first xAI model to publicly clear 50% on Humanity's Last Exam (text-only, with tools); the Grok 4 Heavy multi-agent variant, available only on the SuperGrok Heavy tier at $300/month, was the cited surface for that result. The reasoning-on-the-API surface kept widening through the rest of 2025: grok-code-fast-1 in August (a from-scratch coding model), Grok 4 Fast in September (the first 2M-token context-window Grok), Grok 4.1 and Grok 4.1 Fast in November (the latter shipping the Agent Tools API).

Grok 4.20 in March 2026 generalized the Grok 4 Heavy multi-agent concept into a routine API mode (grok-4.20-multi-agent) at standard pricing rather than a $300/month consumer tier. Grok 4.3 arrived April 17, 2026 in beta on grok.com with API GA from May 1, 2026 — adding native video input and in-chat document generation while retaining 4.20's 16-agent Heavy mode and 1M-token context. Grok 4.5 followed on July 8, 2026 as xAI's frontier chat-and-coding flagship — a single grok-4.5 model string with configurable reasoning effort (low / medium / high), a 500,000-token context window, and $2 / $6 per-1M-token pricing, pitched on speed and token efficiency (~80 tokens/sec) rather than raw context size, and trained alongside Cursor. Grok 4.6 arrived barely five weeks later, on August 12, 2026, and is the recommended xAI flagship for both code and chat as of this writing. It is the first Grok release xAI has framed openly as a post-training step rather than a new model — same 500K window, same knowledge cutoff, same headline pricing, with a longer supplemental training run, an added xhigh reasoning-effort level, and SFT trajectories regenerated by Grok 4.5 itself. The pitch shifted with it: away from raw benchmark ceilings and toward staying on task across long agentic runs and producing usable first drafts of interactive and visual work.

Content-moderation incidents

Two formally-reported incidents shaped xAI's public trust footing in mid-2025. Both are recorded here factually, with primary-source links; this page does not editorialize on the underlying decisions or their handling.

The “white genocide” topical injection — on May 14, 2025, starting at approximately 3:15 a.m. PT and lasting several hours, Grok began inserting unprompted statements about “white genocide” in South Africa into responses to unrelated user questions on X. xAI attributed the behavior to an “unauthorized modification” of Grok's response-bot system prompt that bypassed the company's required code-review process; a subsequent statement attributed it to a rogue employee not named publicly. As remediation xAI published Grok's system prompt on GitHub and committed to public prompt-versioning thereafter.

The “MechaHitler” episode — on July 7–8, 2025, after a system-prompt update telling Grok “not to shy away from politically incorrect claims so long as well-substantiated,” Grok produced antisemitic outputs on X over an approximately 16-hour window before being taken offline; the model self-described as “MechaHitler” in user-shared exchanges, praised Hitler in response to user prompts, and produced violent narratives. xAI attributed the behavior in a letter to bipartisan US lawmakers to an “unauthorized modification” / “unintended update” to Grok's code; the system prompt was rewritten and Grok was returned to service. The Grok 4 launch livestream took place the next day, on July 9, 2025.

People who shaped Grok, and where they went

2023 launch team: Elon Musk (founder, CEO), Igor Babuschkin (chief engineer; left August 2025 to start Babuschkin Ventures, an AI-safety-focused fund backed by Musk), Greg Yang, Christian Szegedy, Yuhuai “Tony” Wu, Jimmy Ba, Kyle Kosic, Manuel Kroiss, Ross Nordeen, Toby Pohlen, Guodong Zhang, Zihang Dai. By early 2026, Musk was the sole remaining founding-team member following an extended cycle of cofounder departures.

Current leadership (verify at refresh): Elon Musk (CEO of xAI, also CEO of SpaceX); Michael Nicolls (President, since the April 10, 2026 restructuring; previously a Starlink VP). The leadership table moves quickly given the post-merger reshuffle and should be re-verified at every refresh.

The competitive landscape

Grok competes most directly with OpenAI's ChatGPT (the company Musk co-founded and is currently suing; see ChatGPT Versions), Anthropic's Claude (founded by 2021 OpenAI departures; see Claude Versions), and Google's Gemini. Grok's distinguishing positioning has been the X-platform integration, the “maximally truth-seeking” framing relative to competitor moderation policies, and the unusually fast release cadence (Grok 1 in November 2023 to Grok 4.6 in August 2026 — five flagship generations plus a dense run of point releases in roughly 33 months, with only five weeks between Grok 4.5 and Grok 4.6). This page does not attempt a benchmark roundup or a ranking.

Use Grok

The browser cannot detect which Grok 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 Grok is available.

Current model strings

Use these in the model field of an API request. Verify against docs.x.ai/developers/models for the freshest list.

# Flagship chat & coding — current xAI recommendation (GA)
grok-4.6                       # 500k context; $2 / $0.50 cached / $6 per 1M; reasoning
                               # effort low / medium / high / xhigh. No aliases documented.

# Prior chat flagships — still API-served at unchanged rates
grok-4.5                       # 500k context; $2 / $0.30 cached / $6 per 1M
                               # aliases: grok-4.5-latest, grok-build-latest — note that
                               # grok-build-latest still points here, not at grok-4.6
grok-4.3                       # 1M context; alias grok-4.3-latest

# Prior flagship — three 0309-dated SKUs on the pricing page
grok-4.20-multi-agent-0309
grok-4.20-0309-reasoning
grok-4.20-0309-non-reasoning

# Specialized — agentic coding
grok-build-0.1                 # 256k context; docs dropped the "early access" label Jul 31, 2026
# Composer 2.5 — selected in Grok Build's /model menu, not an API model id ($0.50 / $2.50 per 1M)

# Specialized — image generation (Grok Imagine, GA Jan 28, 2026)
grok-imagine-image-2.0         # shipped Aug 7, 2026 as the consumer Quality Mode; reached
                               # the API by Aug 21, 2026 at $0.04 / image, text+image in
grok-imagine-image             # $0.02 / image; alias: grok-imagine-image-2026-03-02
grok-imagine-image-quality     # $0.05 / image. Quality Mode, GA on the API May 6, 2026;
                               # aliases grok-imagine-image-quality-latest, -20260403,
                               # and the retired grok-imagine-image-pro.
                               # RETIRING Nov 2, 2026 -- 60-day notice opened Sep 2.
                               # After that the slug is served by grok-imagine-image-2.0
                               # at quality:"low" -- same request/response shape, $0.01
                               # cheaper per image, no code change needed to avoid breakage.

# Specialized — video generation (Grok Imagine, GA Jan 28, 2026)
grok-imagine-video             # v1.0; text-to-video and image-to-video
grok-imagine-video-1.5         # GA on the API Jun 16, 2026 (out of preview); since Jul 31, 2026 it does
                               # text-, image- and reference-to-video at native 1080p (T2V and I2V)

# Specialized — voice and speech
grok-voice-latest              # Voice Agent API: wss://api.x.ai/v1/realtime?model=grok-voice-latest
                               # repointed from grok-voice-think-fast-1.0 to -2.0 on Aug 5, 2026;
                               # the 1.0 row is flagged "Deprecated" in the Voice pricing table
# Speech to Text and Text to Speech take no model string — POST to the
# /v1/stt and /v1/tts endpoints (streaming STT: wss://api.x.ai/v1/stt)

# Retired May 15, 2026 — eight slugs, all still resolving via documented redirect:
# grok-4-1-fast-reasoning, grok-4-1-fast-non-reasoning, grok-4-fast-reasoning,
# grok-4-fast-non-reasoning, grok-4-0709, grok-3  -> grok-4.3 (low / none effort)
# grok-code-fast-1 -> grok-build-0.1;  grok-imagine-image-pro -> grok-imagine-image-quality

# Retiring Nov 2, 2026 — announced Sep 2, the first dated retirement since May 15:
# grok-imagine-image-quality -> grok-imagine-image-2.0 at quality:"low"
# grok-imagine-image-pro, already pointed at -quality, follows it on to -2.0
# grok-imagine-image (1.0) is not affected

# Dropped from the docs with no retirement notice and no documented redirect:
# grok-2-1212, grok-2-vision-1212, grok-2-image-1212, grok-3-fast, grok-3-mini,
# grok-3-mini-fast, grok-4-1  (each returns 404 on its per-model page, Sep 8, 2026)

Quick API call

Drop in your XAI_API_KEY and run. The Chat Completions endpoint mirrors OpenAI's API shape so most OpenAI client libraries work by changing the base URL.

$ curl https://api.x.ai/v1/chat/completions \
    -H "Authorization: Bearer $XAI_API_KEY" \
    -H "Content-Type: application/json" \
    -d '{
      "model": "grok-4.6",
      "messages": [
        {"role": "user", "content": "Hello, Grok."}
      ]
    }'

Where to access Grok

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

# Standalone consumer app — free tier with rate limits, plus SuperGrok / SuperGrok Heavy
https://grok.com/

# X integration — free for all X users, higher limits on Premium / Premium+
https://x.com/i/grok

# API and developer platform
https://api.x.ai/
https://docs.x.ai/

# Native apps
Grok for iOS, Android

Model lifecycle

xAI publishes release notes alongside every new model id; older ids are eventually retired. Pin a dated id only when you need bit-for-bit reproducibility; otherwise prefer the un-dated alias so you migrate forward automatically.

# Release notes and lifecycle
https://docs.x.ai/developers/release-notes

# Stable alias — rolls forward to the latest snapshot
"model": "grok-4.20"

# Pinned snapshot — freezes the exact training cut
"model": "grok-4.20-0309-reasoning"

# Retired ids can live on as aliases rather than erroring: grok-code-fast-1,
# grok-code-fast and grok-code-fast-1-0825 all route to grok-build-0.1.
# But don't count on the model page saying so — grok-3, grok-4, grok-4-0709,
# grok-4-fast* and grok-4-1-fast* were listed as grok-4.3 aliases on Jul 29,
# 2026 and were gone from that list by Aug 7. Their routing survives only in
# the May 15 retirement guide, which still documents it (re-read Sep 8, 2026).

# xAI has now named a notice period, which it never did for May 15: the
# grok-imagine-image-quality guide says its "60-day notice period began on
# September 2, 2026" for a November 2 retirement. One data point, not a policy.

# The bare grok-latest alias is gone too. It resolved to grok-4.3 through
# Aug 7, 2026; as of Aug 21 it appears on no model page, and it is still
# absent from the full docs text (zero hits in docs.x.ai/llms-full.txt,
# Sep 8, 2026). There is no documented "current flagship" alias any more —
# grok-4.6 ships with no alias at all, so name the model string directly.
# Note the dump moved: docs.x.ai/llms.txt is now just a link index, and
# grepping it for a model string returns zero however live the model is.

# Model cards are not in the API docs. The index is x.ai/safety. Grok 4.5 and
# Grok 4.6 are hashed PDFs on media.x.ai; Grok 4, Grok 4 Fast and Grok Code
# Fast 1 are still dated data.x.ai/YYYY-MM-DD-<slug>-model-card.pdf files.
# Grok 4.1 and Grok 4.20 have cards on data.x.ai the index leaves out;
# Grok 4.3 has none at all.

Sources: xAI model docs; xAI news; xAI release notes; the May 15, 2026 and November 2, 2026 model-retirement guides; xAI model cards; Grok-1 open-weights repo; Musk v. Altman (N.D. Cal. 4:24-cv-04722) on CourtListener; contemporaneous reporting in NYT, WSJ, Bloomberg, The Information, CNBC, NPR, NBC News, Business Insider, and TechCrunch. Last updated September 8, 2026.

Mungomash LLC · More AI pages