Breaking

SpaceXAI Says Grok 4.7 Reaches 38.0% on Terminal-Bench 4.0, Just Ahead of GPT-5.6 Sol

September 21, 2026 at 23:46 EDT

Software Dev & Coding
AI Agents
Foundation Models

SpaceXAI introduced Grok 4.7 on Sept. 21 as its strongest model for coding and knowledge work, reporting a 38.0% Terminal-Bench 4.0 result for Grok 4.7 xHigh with Grok Build. The figure narrowly clears GPT-5.6 Sol max at 37.3%, but it does not make Grok the benchmark leader.

Sep 21, 2026 · SpaceXAI

Grok 4.7 narrows the terminal-agent gap

Grok 4.7 posts 38.0% on Terminal-Bench 4.0, up from the Grok 4.5 public row at 12.4% and just above GPT-5.6 Sol at 37.3%.

38.0%

Grok 4.7 on Terminal-Bench 4.0

3.1x

Versus Grok 4.5 at 12.4%

+0.7 pp

Ahead of GPT-5.6 Sol at 37.3%

500k

Listed context window

Terminal-Bench 4.0 score stack

12.4%

Grok 4.5

38.0%

Grok 4.7

37.3%

GPT-5.6 Sol

57.9%

Fable 5.1

Each column is scaled to a 60% ceiling; taller stacks mean higher task resolution rate.

Read the score as an agent-system result

Model
Grok 4.7 at xHigh effort

Harness
Grok Build runs the terminal loop

Caveat
Effort and harness settings vary by row

Positive read

Grok Build moved from lagging public rows to near parity with GPT-5.6 Sol on hard terminal work.

Critical read

A narrow margin and differing benchmark methods make the 38.0% figure a signal to verify, not a final verdict.

Continue reading

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