Grok 4.7

xAI discloses that Grok 4.7 uses "a new, larger base model" than Grok 4.6 and a longer reinforcement-learning run on harder, multi-hour tasks. Everything else is undisclosed: architecture type, total and active parameters, expert count, layers, attention mechanism, training tokens and compute. The field is marked unknown rather than assuming a mixture-of-experts design. For context, xAI said Grok 4.5 was trained across tens of thousands of NVIDIA GB300 GPUs in its Memphis data centers, with reinforcement learning over hundreds of thousands of tasks; it has not said how much of that carries into Grok 4.7. Documented behavior: a 500K-token input context, always-on reasoning with four effort levels, summarized reasoning output plus encrypted reasoning content on the Responses API, text and image input, and a May 2026 knowledge cutoff. Like Anthropic, OpenAI and Google, xAI keeps its frontier internals closed.

Scores below lean on xAI's own launch figures, which no independent evaluator had reproduced at verification time. They are set conservatively for that reason.

• Coding (8.8): xAI reports CursorBench 4.0 46.3% at xhigh (Claude Opus 5 at max 46.6%, GPT-5.6 Sol at max 41.7%, Claude Fable 5.1 at max 51.8%), DeepSWE v1.1 71.0% at high (Sol 72.7%, Fable 5.1 70.0%) and Terminal-Bench 4.0 38.0% (Sol 37.3%, Fable 5.1 57.9%). That places it with GPT-5.6 Sol and below Fable 5.1, a large step up from Grok 4.3's 6.5. It is the default model in Grok Build.
• Reasoning (8.5): No GPQA, HLE or AIME figure has been published. HealthBench Professional 56.7% trails Sol (60.5%) and Fable 5.1 (62.1%); EEBench 64.0% beats both. Scored on the strength of the professional-task results, held below the coding score for lack of standard reasoning data.
• Math (8.0): No math benchmark published for Grok 4.7. Held at Grok 4.3's level until one appears.
• Agentic / tool use (9.0): The main pitch. Training weighted toward multi-hour tasks, xAI's AA Briefcase figure of 1,657 Elo for multi-hour office work (second to Fable 5.1's 1,678 on xAI's chart), GDPval 1,695 Elo, and Terminal-Bench 4.0 level with Sol. Context compaction and WebSocket mode support long agent loops.
• Long context (7.5): 500K tokens is half Grok 4.3's 1M and half the 1M-class windows of GPT-5.6 Sol, Claude Sonnet 5 and Gemini. xAI says long-context management improved, but no long-context benchmark was published, and prompts of 200K tokens or more double the price of the whole request.
• Multilingual (6.5): No multilingual data published.
• Vision / document OCR (6.5 / 6.5): Image input (jpg and png, up to 20 MiB, no image-count limit) with no published vision benchmark. xAI reports better document and presentation creation, which is output quality, not OCR.
• Creative writing (7.0): No data. xAI's emphasis is on documents, slides and professional artifacts rather than fiction.
• Instruction following (8.0): No IFEval or IFBench figure. Scored on Grok 4.3's baseline.
• Function calling (8.5): Function calling and structured outputs on both the Responses and Chat Completions APIs, plus server-side web search, X search and code execution tools.
• Safety calibration (7.0): xAI reports a new safeguard stack: 3.3% of risky dual-use cyber prompts allowed through on its own HackerBench v0.3, and a leading 62.4% on LatchBio's biosafety benchmark. That is a clear improvement in xAI's own framing over Grok 4.3's 5.5, but none of it has been independently reproduced.
• Real-time data (10): Native X search (posts, profiles, threads) and web search as server-side tools. No rival has first-party access to the live X corpus. Real-time access requires enabling those tools; without them the model knows only its training data.

All scores in this table are vendor-reported by xAI in its 2026-09-21 launch post. Grok 4.7 figures are at xhigh effort unless noted; competitor figures are as charted by xAI.

Grok 4.3, the previous model in this catalog, has no scores on these benchmark versions, so the predecessor column compares against Grok 4.6 from the same xAI table. None of these results fills the catalog's standard compare fields. xAI published no MMLU-Pro, GPQA Diamond, AIME, SWE-bench Verified, LiveCodeBench, MMMU, HLE or LMArena figure for Grok 4.7. Its Terminal-Bench result is on version 4.0, a harder revision than the Terminal-Bench 2.x scores stored for other models, so that compare field stays empty rather than mixing versions. The GDPval and AA Briefcase Elo values also differ from the GDPval-AA v2 and AA-Briefcase values in xAI's Grok 4.6 post, so compare them only within the same chart. No independent evaluator had published a verifiable Grok 4.7 result at verification time.

No independent throughput or time-to-first-token measurement of Grok 4.7 existed at verification time, so both fields are null. xAI says Grok 4.7 is served at the same speed as Grok 4.6, and its launch tagline claims it is "twice as fast" as comparable models. In July xAI stated Grok 4.5 ran at about 80 tokens per second. Three things shape real latency. Reasoning is always on and defaults to high, so even simple requests spend time thinking; low is the fastest setting, and xhigh adds latency for the hardest work. xAI's code samples set one-hour client timeouts for reasoning requests. Grok 4.7 Fast doubles output speed at double the price, but only inside Cursor and Grok Build. On the API, Priority Processing buys higher scheduling priority at 2x token rates. Long agent loops can use context compaction and WebSocket mode to cut time-to-first-token and round-trip overhead.

Requests blocked before generation for usage-guideline violations on the Responses API carry a $0.05 fee; requests flagged after generation are billed normally.

Proprietary and API-only: no open weights, no self-hosting. The xAI API serves Grok 4.7 through the Responses and Chat Completions endpoints at api.x.ai/v1, with the native xai_sdk, the OpenAI SDK and the Vercel AI SDK all documented. The global endpoint may route between regions. The US regional endpoint (us.api.x.ai/v1, 10% premium) guarantees US handling, inference, moderation and retained data, but Files, Collections and server-side tools (web search, X search, code execution) fall outside that guarantee. xAI lists Grok 4.7's regions as us-east-1, us-west-2 and us-central-1, with no EU region. Beyond the API it runs as the default model in Grok Build (xAI's coding agent CLI), on all Cursor plans, and through the OpenRouter, Vercel and Cloudflare gateways. xAI's pages do not confirm Amazon Bedrock, Google Vertex AI or Microsoft Foundry availability for Grok 4.7.

• Safeguards (vendor-reported): Grok 4.7 ships an "entirely new safeguard stack". xAI says it is the strongest model it has tested on refusals and jailbreak resistance, tops LatchBio's biosafety benchmark at 62.4%, and lets through only 3.3% of risky dual-use prompts on HackerBench v0.3, xAI's own cyber-risk benchmark, while rarely blocking legitimate security work. Select cybersecurity partners get invite-only access to its red-team capabilities. No named safety framework or capability tier is published alongside the model.
• Training on inputs: xAI states it never trains on API inputs or outputs without explicit permission. By default, API requests and responses are kept encrypted for 30 days for abuse auditing, then deleted.
• Zero Data Retention: available as a team-wide setting. It disables the stateful Responses API, Files, Collections, the Batch API and deferred completions. On the Responses API, Grok 4.7 always returns encrypted reasoning, and xAI notes this default does not stop it storing the reasoning trace server-side for rehydration.
• Compliance: xAI states it is SOC 2 Type 2 compliant, with details in its Trust Center under NDA. A HIPAA Business Associate Agreement is available through a questionnaire. No GDPR-specific certification is stated, and no EU region is listed for Grok 4.7.
• Consumer surfaces: Grok Build asks users whether xAI may retain data to improve the product and model, and code-data retention can be turned off in its privacy settings. Build-mode sessions on grok.com may be used for training if the "improve the model" toggle is on.

Official SDKs cover Python (xai_sdk) and TypeScript through the Vercel AI SDK's @ai-sdk/xai provider, and the OpenAI SDK works against api.x.ai/v1 for both Responses and Chat Completions. Grok 4.7 is the default model in Grok Build, xAI's coding agent (TUI, headless CLI and Agent Client Protocol), and is available on all Cursor plans. Model gateways OpenRouter, Vercel and Cloudflare serve it. xAI trained it to understand the harness of Grok Bot, its persistent cloud agent product. Popularity is growing: Cursor distribution gives it a large developer audience, while enterprise penetration still trails OpenAI, Anthropic and Google.