Grok 4.7: what the flagship can do, what it costs and how it really scores

Release and specs

SpaceXSI (formerly xAI) released Grok 4.7 on September 21, 2026. It replaced Grok 4.6, which came out on August 12, as the company’s flagship. SpaceXSI says Grok 4.7 is built on a larger new base model, with longer reinforcement learning focused on tasks that run for several hours.

• Context window: 500k tokens on the API.
• Input: text and images.
• Knowledge cutoff: May 2026.
• Output cap: none for text, according to the docs.
• Size: Musk has described it as about 2.1T parameters. SpaceXSI has not published an official figure.

Effort levels and API settings

Grok 4.7 has four reasoning effort levels, set with the reasoning_effort parameter: low, medium, high (the default) and xhigh. Higher effort spends more tokens thinking before it answers. Most published scores, including SpaceXSI’s own, use xhigh.

Two API details matter for cost. Calls through the Responses API always return encrypted reasoning. SpaceXSI also strongly recommends setting prompt_cache_key or the x-grok-conv-id header. Without it, a request can land on a server with a cold cache and pay full input price instead of the cached rate.

Price and the Fast variant

On the API, Grok 4.7 costs the same as Grok 4.6. Long-context rates apply once a prompt passes 200k tokens and then cover the whole request.

Prices are USD per million tokens. The US endpoint keeps inference in the US at 1.1x global rates.

Grok 4.7 Fast is the same model on faster infrastructure. SpaceXSI describes it as about twice the output speed for twice the price, or 1.5x for long-context requests. It is only available in Cursor and Grok Build. It is not on the public API and not in the Grok Build free tier.

What SpaceXSI claims

SpaceXSI published these scores for Grok 4.7 at xhigh effort. They are vendor claims, not independent tests.

SpaceXSI calls the LatchBio result the top score. Its own charts are candid about the rest: Claude Fable 5.1 Max leads on CursorBench, Terminal-Bench and GDPval, and GPT-5.6 Sol Max leads on DeepSWE.

What independent tests found

Outside testers paint a more modest picture.

• Artificial Analysis Intelligence Index: 46, two points above Grok 4.6. Artificial Analysis says this puts SpaceXSI among the top four labs. The model itself ranks around 25th to 30th of 217 models, depending on how effort variants are counted (as of October 18, 2026).
• Coding Agent Index: 56 when run inside Grok Build, fourth behind Claude Fable 5.1, GPT-6 Astra and Claude Opus 5, according to Artificial Analysis.
• Hallucination rate: 29%, down from 34% for Grok 4.6, per Artificial Analysis. A hallucination is a confident answer that is false.
• LMArena text leaderboard: #89 with a score of 1444, as of October 15, 2026. That is well outside the top 30.
• Token use: Artificial Analysis found it token-hungry, at about 81k output tokens per task versus 38k for Grok 4.6. Its model page also notes a very slow time to first token.

For how Grok stacks up against OpenAI’s models in everyday use, see Grok vs ChatGPT.

Where you can use Grok 4.7

• API: as grok-4.7, on the global endpoint and the US regional endpoint.
• Grok Build: Grok 4.7 now powers the terminal coding agent.
• Cursor: standard and Fast versions, drawing on the shared weekly pool.
• Grok apps: the X Subscription Premium plan ($30) lists “latest models” among its features. SpaceXSI does not publish a model-by-plan table for the apps.

Grok 4.8 has not shipped as of October 18, 2026. See all Grok models for the full history.

Questions people ask