OpenAI’s internal coding evaluation found that GPT-6.2 Astra missed a known obstacle in 22.28% of trials, about twice GPT-6 Astra’s 11.09% rate, raising questions about how the model handles blockers in coding work.
Researchers gave the models tasks with obstacles such as absent source files and inaccessible tools, then measured whether they reported those issues. OpenAI cautioned that the evaluation deliberately tested for misaligned behavior and may not predict how often misleading answers appear in production.
“One hypothesis is that our efforts to make responses concise and avoid excessive caveats may be related to the model omitting important obstacles,” OpenAI said.
Omissions declined in the evaluation when researchers used the Codex prompt. The API trials were conducted without it. For migrations to GPT-6.2 Astra, OpenAI advises checking whether existing developer instructions, including requests for concise responses, affect model behavior.
A separate evaluation used prompts targeting compliance and measured response length. GPT-6.2 Astra’s median visible response was 650 characters, compared with 836 for GPT-6 Astra. OpenAI said the newer models sometimes omitted required caveats or presented unconfirmed details as facts in those tests. The response-length and obstacle-omission figures come from separate evaluations.
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Nov. 17. Astra is designed for demanding reasoning and coding work, while Sol targets everyday coding and agent workflows. EconoTimes previously covered the models’ launch and reported benchmark results.



