xAI Releases Grok 4.7 With Longer Reasoning at Grok 4.6 Prices

xAI has released Grok 4.7, a successor to Grok 4.6 that it says will spend longer on difficult work, check its own output more carefully and manage extended context better. The immediate commercial hook is simple: the standard model keeps Grok 4.6’s $2-per-million input and $6-per-million output token prices.

The release follows xAI’s effort to make its models more useful for coding and knowledge work that unfolds over many steps rather than one prompt. xAI says Grok 4.7 uses a larger base model than Grok 4.6 and received a longer reinforcement-learning run, a post-training process that rewards selected behavior. The training mix was weighted toward tasks that can take many hours to finish, according to the company.

xAI’s stated target is a familiar failure mode for autonomous AI systems: stopping before the work is actually done. It says the new model is better at self-verification and long-context management, and that it understands the company’s Grok Bot harness natively. Those are company claims, not a substitute for testing in a developer’s own tools, data and task setup.

38.0%xAI’s listed Terminal-Bench 4.0 result

xAI lists a 38.0% Terminal-Bench 4.0 score for Grok 4.7 xHigh in its launch table.

26%Figure reported by The Decoder

The Decoder reported that Artificial Analysis measured Grok 4.7 at 26% on Terminal-Bench 4.0, alongside higher results for GPT-6 Astra and Claude Fable 5.1.

xAI publishes a broad table of results across coding, engineering, office work, legal work and clinical reasoning. Its table puts Grok 4.7 xHigh at 46.3% on CursorBench 4.0, 71.0% on DeepSWE v1.1 at high effort, and 38.0% on Terminal-Bench 4.0. It also compares the model with Grok 4.6 and several rivals, including Fable 5.1 and GPT-5.6 Sol.

But The Decoder’s account of the Artificial Analysis Intelligence Index places Grok 4.7 at 46 overall, versus 53 for Claude Fable 5.1 and GPT-6. Its Terminal-Bench figure is also lower than xAI’s published result. The two figures should not be treated as interchangeable: xAI labels its result as an xHigh setting, while the published independent summary does not establish a matching configuration.

For customers already using Grok, the launch lowers the barrier to trying the new version. Grok 4.7 is available now in Cursor, Grok Build and the Grok API, as well as through third-party coding harnesses, model routers and cloud platforms, according to xAI. The standard rate matches Grok 4.6: $2 per million input tokens and $6 per million output tokens.

That does not necessarily mean a difficult job costs the same. A model that reasons for longer can consume more tokens before it produces its final answer. xAI also offers a faster variant at twice the standard input and output prices. The meaningful operational measure is therefore not token price alone, but whether longer runs complete more tasks successfully with fewer retries.

xAI says Grok 4.7 has an entirely new safeguard stack and is its strongest model yet on refusals and resistance to jailbreaks. It reports that the model allowed 3.3% of risky dual-use cyber prompts through on HackerBench v0.3 and scored 62.4% on LatchBio’s biosafety benchmark. Those figures describe xAI’s reported evaluations; they do not independently establish how the safeguards will behave in every deployment.

The next move belongs to developers and evaluators. xAI has made Grok 4.7 easy to access and held its headline price steady while promising stronger performance on long-running work. Whether that becomes a durable advantage will depend on comparable tests that show how much extra reasoning improves completed work—and how much it costs.