Grok 4.7 Guide: xAI's Flagship, and What Changes on Upgrade

Grok 4.7 is the model xAI now recommends for "everything else, including code", and its models page calls it the most capable model the company has built. It was announced on 21 September 2026, under six weeks after Grok 4.6, at exactly the same price.

A note on names: xAI, the company behind Grok, was renamed SpaceXAI in July 2026 after SpaceX acquired it. This guide uses the xAI name throughout.

Same price and better scores sounds like a free upgrade. Mostly it is. But Grok 4.7 also changes how reasoning comes back on the Responses API, it leans hard on prompt caching to hit its quoted costs, and its benchmark profile has one clear gap. This guide covers all three.

What is new in Grok 4.7

According to xAI's announcement, Grok 4.7 is built on a larger base model than Grok 4.6, followed by a longer reinforcement learning run on a harder task mix weighted towards problems that take many hours to complete. xAI says it is better at checking its own work and managing long context, was trained to work natively in its Grok Bot harness, and produces better documents and presentations. It also ships on a new safeguard stack.

The spec sheet, from xAI's documentation:

• Context: 500,000 tokens, unchanged from Grok 4.5 and 4.6.
• Knowledge cut-off: May 2026.
• Input: text and images; output is text.
• Reasoning effort: low, medium, high (default) and xhigh.
• Tools: function calling plus built-in web search, X search and code execution.
• Batch: not supported.

There is still no published parameter count and no open weights. "Larger base model" is as specific as xAI gets.

Benchmarks: strong on knowledge work, behind on long terminal tasks

xAI's launch table compares Grok 4.7 at xhigh effort with Grok 4.6 at high, GPT-5.6 Sol at max and Fable 5.1 at max. All figures are vendor-reported:

• Harvey Legal Agent Benchmark: 19.6%, first in the table.
• EEBench (electrical engineering): 64.0%, first.
• DeepSWE v1.1: 71.0% at high effort, behind GPT-5.6 Sol at 72.7% and ahead of Fable 5.1 at 70.0%.
• AA Briefcase v1.1 (multi-hour office work): 1,657, just behind Fable 5.1 at 1,678.
• CursorBench 4.0: 46.3%, behind Fable 5.1 at 51.8%.
• HealthBench Professional: 56.7%, behind both competitors.
• Terminal-Bench 4.0 (multi-hour terminal work): 37.6%, against 57.9% for Fable 5.1.

Grok 4.7 improves on Grok 4.6 on every row, often substantially; Terminal-Bench 4.0 nearly doubles from 20.3%. But that last row is the gap worth noting. On long, unattended terminal work, the kind an autonomous coding agent does, Fable 5.1 is twenty points ahead in xAI's own table. If that is your workload, test both before deciding. Vendor-claimed vs independent benchmarks covers how far to trust a table like this.

Integration details that bite

Three documented behaviours matter when you move an existing integration to Grok 4.7.

Encrypted reasoning is always returned

On the Responses API, Grok 4.7 always includes reasoning.encrypted_content in its output, even if you did not ask for it in include. xAI's guidance is to pass reasoning items back unchanged in the next request's input to keep multi-turn reasoning intact. Code that filters the response down to text before storing the turn will silently drop that context. Chat Completions behaviour is unchanged.

Set a cache key

xAI strongly recommends setting prompt_cache_key on the Responses API, or the x-grok-conv-id header on Chat Completions. Both route a conversation to the same server so cache hits are reliable. Without one, you may pay full input price on turns that should have been cached. On a long agent loop, it is one of the cost settings most worth getting right.

Unsupported parameters

As on other Grok reasoning models, presencePenalty, frequencyPenalty and stop return an error, and logprobs is silently ignored. Clear them out of shared request builders before you switch.

Cost: one list price, two multipliers

Grok 4.7 lists at $2.00 per million input tokens, $0.50 cached and $6.00 output for prompts under 200K tokens. OpenRouter lists the same base rates at the time of writing. At or above 200K, every token in the request is billed at double: $4, $1 and $12.

Two multipliers sit on top of that:

• US regional endpoint. us.api.x.ai keeps inference in the United States, with token usage priced 10% higher.
• Grok 4.7 Fast. The same model on faster infrastructure at twice the token rate, or 1.5 times for long-context requests. It is only available inside Cursor and Grok Build, billed through those plans, and not on the public API.

xAI's announcement describes Grok 4.7 as twice as fast at half the price of comparable models. That depends on which models you compare it with, and on output tokens per task. Measure cost per completed task on your own work rather than relying on the claim.

Upgrading from Grok 4.6

For most teams on Grok 4.6 this is a straightforward move: same window, same rates, same effort levels, better scores. The work is in the details above. Check how you persist Responses API turns, add a cache key if you have not already, and re-run your evaluation set before switching production traffic.

If you are coming from Grok 4.3, the change is bigger. You lose half the context and batch support, the default effort jumps from low to high, and output costs 2.4 times as much at list price. Route only the work that needs a flagship to Grok 4.7 and keep the cheaper model for the rest. Model routing and fallbacks covers that pattern.

Decision rule

• Use Grok 4.7 for coding agents, document-heavy knowledge work and long multi-step tasks that fit in 500K tokens.
• Test an alternative if your core workload is long, unattended terminal work, where xAI's own table shows a large gap.
• Stay on Grok 4.3 for high-volume, batchable or over-500K workloads where the flagship's extra quality does not change the outcome.