SpaceXAI releases Grok 4.7

Background

Grok 4.6 had been released on 12 August 2026 as a model for long-running agents. Grok 4.7 followed on 21 September 2026 as the next model in the line.

What happened

On 21 September 2026 the company released Grok 4.7. It has a larger base model than Grok 4.6 and a longer reinforcement-learning run, which is a training method in which the model improves from feedback on its own attempts^[1].

Key details

The model kept Grok 4.6's 500,000-token context window and its $2 and $6 per million token pricing^[2][1]. A token is a small unit of text that a model reads and writes.

A Fast variant was also released at twice the price. It was available only in Cursor and Grok Build^[2].

The company reported 71.0% on DeepSWE v1.1 and 37.6% on Terminal-Bench 4.0^[1]. The release also included safety results: 62.4% on LatchBio's biosafety benchmark, and 3.3% of risky dual-use prompts allowed on HackerBench v0.3^[1].

It launched in Cursor, Grok Build and the API. Access for red-team testers, who look for failures in the model, was limited to selected cybersecurity partners who were invited^[1].

Impact

The release extended the Grok 4 line's coding and knowledge-work focus, with the same price per token as before. Its Fast variant and red-team program were limited to a small set of tools and partners^[1][2].

For readers, a token is the unit in which both the context window and the prices are measured. At twice the standard rates, the Fast variant works out to $4 per million input tokens and $12 per million output tokens, against the standard $2 and $6. The safety figures are the company's own results and cover two specific tests: a biosafety benchmark and a set of dual-use prompts^[1]. The main points to remember are the release date, the larger base model and longer training run, the unchanged prices, and the limited access to the Fast variant and to red-team testing^[1][2].