Grok 4.7

Grok 4.7 is a large language model released by xAI on 21 September 2026, positioned by the company as its most capable model for coding and knowledge work^[1].

What changed from Grok 4.6

Grok 4.7 is built on a larger base model than Grok 4.6 and was trained with a longer reinforcement-learning run^[1]. The company also points to better self-verification — the model checking its own intermediate work — and to improved long-context management^[1].

Many of the surrounding settings stayed the same. Grok 4.7 keeps the 500,000-token context window of Grok 4.6, the same prices of $2 per million input tokens and $6 per million output tokens, and the same reasoning-effort options, which let a user trade speed for more deliberation^[2][1]. The model was additionally trained to natively understand the Grok Bot harness, which the company says makes it better at conversational tasks and at general knowledge work^[1].

The Fast variant

Alongside the standard model, xAI offered Grok 4.7 Fast, which delivers twice the output speed at twice the price; long-context requests are billed at 1.5 times the standard rate^[1][2]. Fast was not a general release: it was offered only inside Cursor and Grok Build^[2].

Benchmark results

The company published a set of benchmark numbers, measured at the xhigh reasoning setting unless noted otherwise. On CursorBench 4.0 the model scored 46.3%; on DeepSWE v1.1, at high effort, 71.0%; on Terminal-Bench 4.0, 37.6%; on HealthBench Professional, 56.7%; and it reached an Elo of 1695 on GDPval^[1]. Evaluations of this kind test a model carrying out multi-step tasks rather than answering single questions, so they say more about sustained tool use than about one-shot accuracy. Two further comparisons were given against the predecessor: 64.0% on the EEBench electrical engineering benchmark versus 53.0% for Grok 4.6, and 19.6% on the Harvey Legal Agent Benchmark versus 15.8%^[1].

Safety

xAI said the model was built with an entirely new safeguard stack and described it as the strongest model it had tested on refusals and on resistance to jailbreaks^[1]. Two safety figures were published with the release: 62.4% on LatchBio's biosafety benchmark, and 3.3% of risky dual-use prompts allowed on HackerBench v0.3^[1].

Where it is available

Grok 4.7 could be used in Cursor and Grok Build, through the Grok API, in third-party coding harnesses, through model routers and on cloud platforms^[1]. Red-team access for a set of cybersecurity partners was invite-only^[1].

Reception

Coverage of the release noted a practical caveat: the per-token price did not change, but token consumption on the same tasks went up, so a given job could cost more in practice than it did on earlier models^[3]. The same coverage noted that Cursor data was used in training the model^[3].