Cognition Uses GPT-6 Astra to Let Devin Test Its Own Work

Information published by OpenAI says Cognition is using GPT-6 Astra to strengthen the self-testing ability of Devin, its AI software engineer. The central claim is that Astra improves Devin's ability to test software and show that it works.

The stated goal is straightforward: engineers review less code and ship more. In practice that means AI-written changes are not simply queued for line-by-line human inspection; they are expected to run, pass, and arrive with evidence first.

The underlying engineering problem is volume. The more code a model generates, the heavier the review burden becomes, and the bottleneck moves from writing to reviewing. Evidence produced by the model itself is one of the few ways to relieve that pressure.

Devin is positioned as an agent that takes on development tasks, and testing is one of the hardest steps for any such agent. Changing code is not the same as confirming the change is correct, so adding self-verification is a key step toward closing the loop.

The information again comes from OpenAI's own channels and describes Astra's effect inside a specific product, without quantitative figures on test coverage or defect rates. It is best read as a public signal about a direction of capability rather than a measured benchmark.

Two things to watch: whether self-testing holds up in large, long-lived codebases, and how review policies and quality gates need to change once a model can both write code and vouch for it.