- Published on
Gains, privacy, and cost
- Authors
- Name
- DeepHorus
OpenAI's latest AI model, GPT-6 Astra, has demonstrated significant improvements in complex code review tasks, particularly in identifying bugs that span multiple files. In an early evaluation, Astra caught approximately 4% more actionable bugs than its predecessor, GPT-5.6 Sol, and a substantial 22% more than Anthropic's Opus 5 model. This advancement is most pronounced in challenging cross-file reviews, where Astra's performance gains reached 20% over Sol and 33% over Opus 5.
The evaluation focused on 'actionable bug coverage,' measuring how many labeled bugs a model can detect through findings a developer can directly act upon. While the overall gain over GPT-5.6 Sol appears modest, the results become more encouraging when isolating the harder subset of reviews. This suggests Astra's strength lies in understanding the intent behind a code change and connecting it to its distributed consequences across a codebase.
These findings are described as an early, directional result. They do not establish an overall ranking of review quality or guarantee identical gains for every software team. The interpretation is that Astra's advance stems from its ability to identify and connect relevant pieces of information from a large context window, a key requirement for complex reasoning.
Pricing for Astra's API is set at $10 per million input tokens and $50 per million output tokens. An illustrative comparison using a fixed token count shows Astra costs 2.5 times more than Sol and significantly more than other models like Terra and Luna. OpenAI notes that higher token prices do not always translate to higher total task costs, as more efficient models may require fewer tokens or attempts.
The model's enhanced reasoning over distributed information has potential applications beyond code review. Researchers could evaluate its use for tasks like research synthesis, which involves reconciling conflicting reports and connecting claims to evidence. OpenAI positions Astra for multistep work across code, browsers, and professional software.
For development teams, the key consideration is identifying tasks where Astra's stronger reasoning justifies the premium cost. Difficult tasks with scattered evidence are prime candidates for investigation. Teams are advised to measure total cost per successful outcome on their specific work rather than relying solely on token prices or capability scores.
