GPT-6 Astra in Practice: Tests, Use Cases, and Feedback

After a large coding project with GPT-6 Astra, the author finds it smart but says Opus 5.5 is the more productive model, expecting Anthropic to keep high-end workload share.

On 50 unpublished physics research problems, GPT-6 Astra solved 98% and Opus 5.5 92% at under $3 per problem; the authors note AI still lacks the taste to set research direction.

A user bets OpenAI will ship GPT-6.2 Astra within seven days of Fable 5.5 to avoid ceding the spotlight to Anthropic; no confirmation yet.

OpenAI's blog details how its team used GPT-6 models to merge 1,400+ cleanup PRs and delete 800k lines in two weeks, with the actual prompts.

A rumor says OpenAI scrapped the GPT-6.1 Astra release but kept training on the same base model, which could become its answer to Fable 5.5 if alignment is fixed.

A Japanese post argues Anthropic's revenue growth slowed exactly when enterprises moved from Claude Code to OpenAI Astra Codex over data-review terms, and says customers are returning after Opus 5.5.

An analyst says Anthropic's lower-than-expected ARR is not a sign its models fell behind, but that users moved en masse from Claude to Codex after Astra launched, with total demand unchanged.

A Japanese post says Astra answered 49 of 50 theoretical physics problems, and that several answers marked wrong turned out to be errors in the experts' own solutions, citing an arXiv paper.

A month after OpenAI launched Astra, a trader describes two days of arguing across five buy-side group chats about lab ARR, the bogey, and how portfolio managers will react.

Physicists from Harvard, TIFR, UT Austin and DeepMind wrote 50 unpublished research problems; Astra, run in Codex with internet access, answered 49 at under $3 per problem, while Opus 5.5 in Claude Code got 46.

Both models built the same Arctic scene: Claude's water reflections stand out, while Astra's penguins are easier to see but too large and unrealistic, at API costs of $26.42 versus $30.33.

On TheoPhysEval, a set of 50 unpublished research-grade theoretical physics problems, GPT-6 Astra run through Codex scored 98% at an average cost below $3 per problem.