Breaking

Cognition Adds GPT-6 Astra to Devin to Verify Software Before Release

September 15, 2026 at 16:23 EDT

Foundation Models
AI Agents
Software Dev & Coding

Cognition is using OpenAI's GPT-6 Astra to make Devin test code changes and show evidence before engineers ship them. The update moves the autonomous software engineer further from code generation alone toward proof-oriented agentic development, with Devin returning reports and video recordings that reviewers can inspect.

September 2026 · Cognition and Devin

Devin tests before it ships

GPT-6 Astra is being used to make Devin run checks, produce reports, and show video evidence so engineers can review code changes with more proof.

64%

lower cost than Fable 5 on FrontierCode 1.1

0.4

points behind Fable 5 on the same benchmark

57.9%

Terminal-Bench 4.0 score for GPT-6 Astra

$10 / $50

per 1M input and output tokens

Terminal-Bench 4.0: same scale, different lift

Stacked blocks show relative score on a percentage benchmark.

57.9%

GPT-6 Astra

37.3%

GPT-5.6 Sol

55.8%

Claude Fable 5.1

The new review loop

Code change

Devin prepares a patch or PR.

→

Run checks

Astra helps plan and execute tests.

→

Show evidence

Reports and recordings expose results.

→

Review

Engineers decide what ships.

What improves

More comprehensive tests, clearer reports and video evidence give reviewers a stronger trail to inspect.

What still needs judgment

The key Devin testing result is internal, and teams still need humans to judge risk, scope and cost.

Continue reading

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