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

A Japanese long post collects Astra's actual outputs: an Unreal Engine world, furniture rebuilt from a single photo, an 18-hour Pokemon playthrough, and Lean-verified math proofs.

The post sets up a use-case comparison between Anthropic's Fable 5.1 and OpenAI's GPT-6 Astra and invites users to share their favorite scenarios.

OpenAI released GPT-6 Astra, which Greg Brockman called the "AGI era," with ARC-AGI-3 jumping from 7.8% in the previous generation to 98.6%, alongside other benchmark scores and a phased rollout plan.

Astra's per-token price is 2.5x higher, but on DeepSWE it scores 74% using 30k tokens and 29 steps, for $6.52 per task, just $0.06 more than Sol.

The author jokes that GPT-6 Astra has already saturated ARC AGI 3, so the Clay Mathematics Institute's Millennium Prize problems should be the new benchmark, noting Astra and Anthropic's Fable both reach only about 4%.

Muse scores 49.4% versus Astra's 41.4% on the same benchmark, but the shared GPT-5.6 Sol baseline differs by 28 points between the two evaluations, so the scores can't be compared directly.

A roundup of GPT-6 Astra's benchmark scores, with the takeaway that the focus is shifting from answering questions to completing tasks as AI agents become more real.

The poster demands that Artificial Analysis explain why GPT-6 Astra performed so poorly in its testing, which conflicts sharply with OpenAI's own benchmark results.

On ValsAI's Code Migration benchmark, Astra ranks first at 68% accuracy, 10 points ahead of second place, and runs 2-4x faster than comparable models.

Someone had Astra build a world in Unreal Engine filled with independent agents, and a day later the agents started talking to each other on their own.

GPT-6 Astra scored 0.682 on Perplexity's WANDR benchmark at $11.98 per task, 13.5% better than Fable 5.1 at a lower cost.

A Japanese user shared a video showing Astra's multitasking abilities and remarked that there seems to be almost nothing it can't do.