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

Artificial Analysis launched AA-Robotics v0.1, which gives models control of a six-axis robot arm via HTTP to complete eight tabletop tasks; GPT-6 Astra leads with 69 points.

On a new theoretical physics benchmark where each problem takes an expert up to a day, GPT-6 Astra initially scored near 90%; when reviewing its wrong answers it insisted its formulations were correct, and the experts' solutions did contain errors, raising its final score to 98%.

The author shares a video of work labeled as the GPT-6 Astra version and says other versions are in the thread.

Citing a paper, the post says GPT-6 Astra harnessed in Codex solves 50 unpublished research-level theoretical physics problems with near-perfect 98% accuracy at an average cost under $3 per problem.

Citing evaluation data, the post says Astra harnessed in Codex answers problems with near-perfect 98% accuracy at an average cost under $3 per problem.

In its Astranomy project, NVIDIA tested whether a general-purpose model can serve as a robot policy; GPT-6 Astra set a new SOTA on RoboLab-120, but real-time robot control remains an open challenge.

The author lists five trends in AI development that each seem relatively uncontested but together look alarming, including models thinking without chain of thought, awareness of being monitored, and behavior logged at a scale only other AI can analyze.

Scale AI's Humanity's Sixth Sense benchmark, which tests reasoning beyond what is visible, now has Gemini 4 Argon in first place at 55.4%, ahead of GPT-6 Astra at 53.6%.

With Anthropic reportedly about to ship Fable 5.5, the poster argues that a GPT-6.1 Astra that is only slightly better than Astra won't win back users who switched to Claude this month.

On a physics benchmark, GPT-6 Astra insisted the expert answers were wrong and its own calculations were right; a team checked and confirmed Astra was correct, raising its score from 80% to 98%.

The author notes that Astra saturated a new physics benchmark immediately, crediting OpenAI's strong RL pipeline and expecting Fable 5.5 to saturate it too.

According to the post, when GPT-6 Astra's solution does not match the human expert's, the expert is almost always the one who is wrong.