Breaking

SpaceXAI Releases Grok 4.7, Narrowly Passing GPT-5.6 Sol on Terminal-Bench 4.0

September 21, 2026 at 21:31 EDT

Foundation Models
AI Agents
Software Dev & Coding

SpaceXAI released Grok 4.7 on September 21, positioning the model as its strongest system for coding, agentic terminal work, and professional knowledge tasks. In the company’s published comparison, Grok 4.7 rose to 38.0% on Terminal-Bench 4.0, up from Grok 4.6’s 20.3% and just ahead of GPT-5.6 Sol Max at 37.3%, while Claude Fable 5.1 Max remained far ahead at 57.9%.

Sep. 21, 2026 · SpaceXAI

Grok 4.7 moves up in agentic coding

The new model nearly matches GPT-5.6 Sol on Terminal-Bench 4.0, but Claude Fable 5.1 still leads by a wide margin.

Terminal-Bench 4.0 completion rates

57.9%

Claude Fable 5.1

38.0%

Grok 4.7

37.3%

GPT-5.6 Sol

20.3%

Grok 4.6

Grok 4.7 more than doubles Grok 4.6 on SpaceXAI's comparison and edges GPT-5.6 Sol by 0.7 percentage points.

500K

token context window

input per 1M tokens

output per 1M tokens

81K

output tokens per independent task

Why it helps

A larger base model, longer reinforcement-learning run, and better self-checking target multi-hour coding and knowledge-work tasks.

What still limits it

Independent testing found gains in coding-agent performance, but also high token use and regressions on some non-coding tasks.

The benchmark story

Grok 4.7 is a sharper competitor for agentic coding, yet the leaderboard still shows a clear separation between catching GPT-5.6 Sol and catching Claude Fable 5.1.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.