Devin tests its own code with GPT-6 Astra
Cognition is putting OpenAI's GPT-6 Astra behind Devin, so the agent can check whether the code it just wrote actually works.
In short
Cognition is using OpenAI's GPT-6 Astra so its coding agent Devin can test its own work and show that it runs, with the stated aim of letting engineers review less code and ship more.
At a glance
- Cognition runs OpenAI's GPT-6 Astra inside its coding agent Devin.
- The emphasis is testing: Devin should demonstrate that the code it wrote actually runs.
- Stated goal: less manual code review, more shipped software.
- The claims come from OpenAI's own announcement; no independent benchmarks are available.
- OpenAI's announcement page returned HTTP 403 on two fetch attempts on September 12, 2026.
Cognition has wired OpenAI's GPT-6 Astra into Devin, and the pitch is about verification rather than generation. The agent is meant to test what it built and produce evidence that the code works, so engineering teams read fewer diffs and ship more.
Generation was never the bottleneck
An agent that writes code moves the work rather than removing it, because somebody still has to check the output. That is the gap this collaboration targets. If the agent writes the tests, runs them and hands over a result a human can inspect, the reviewer's job shifts from reading line by line to judging a piece of evidence.
What the announcement actually claims
The announcement says GPT-6 Astra improves Devin's ability to test software and to show that it works, with the goal of engineers reviewing less code and shipping more. Beyond that framing, the short public summary we have carries no further hard detail.
What we could not verify
OpenAI's announcement page was unreachable at the time of writing: the server returned HTTP 403 on two fetch attempts on September 12, 2026. This report therefore rests only on the publicly circulated summary of the announcement. Benchmarks, measured results, rollout timing and availability could not be checked, and none are asserted here.
Why it matters
The direction is the story. Coding-agent vendors are shifting their marketing away from how much code gets produced and toward how trustworthy the result is. The open question is structural: an agent that writes its own tests is also grading its own homework, and only independent measurement will show whether the evidence holds up in day-to-day engineering work.
FAQ
What does GPT-6 Astra do for Devin?
According to the announcement it strengthens Devin's ability to test software and to demonstrate that the code runs. No technical specifics beyond that could be verified from the source.
Does Devin still need human code review?
The announcement does not say review disappears. It frames the benefit as engineers having less code to review, not as removing human oversight.
Where do these claims come from?
From OpenAI's own announcement about the Cognition collaboration. There is no independent benchmark or third-party evaluation behind the claims so far.
