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

A long post pushes back on the backlash against OpenAI, saying Astra's experience and usage limits have stabilized and that it remains the best model even with Opus 5.5 out.

The author says Astra still has the edge in polish, but Sonnet 5.5 is making it a real competition, and calls on OpenAI to respond at DevDay.

The poster predicts DevDay will bring a $500 tier, GPT-6.1 Astra, and another Hermes Agent knockoff, and complains that the $20 plan gets nothing useful.

The author asked GPT-6 Astra to write a new piece in LilyPond combining the sensibilities of Bach with early popular song composers, and shared the prompt.

The poster predicts Astra 6.1 and no new image models at DevDay, complains that Astra is lazy and Sol is slow, and senses a bubble forming around OpenAI.

The author asked Opus 5.5 to make a video about its Unciv match against GPT-6 Astra; it went through 3.7 GB of logs, notes and screenshots and produced a 15-minute documentary that left the author speechless.

Synthetic Sciences' open-source research agent scored 75.7% on Terminal-Bench-Science, above Codex with GPT-6 Astra at 68.1%.

The author notes Astra was quickly called AGI at launch, but now that Opus 5.5 beats it on several benchmarks and costs 60% less per token, almost no one uses the label anymore.

Retests put Opus 5.5 at 99.2% and GPT-6 Astra at 102.8% of their launch scores, a spread within normal variance, so NerfBench found no nerf detected.

The poster says Sonnet 5.5's first test output beats Opus 5.5 and GPT-6 Astra, and that it releases tomorrow.

A user says one prompt on Astra Medium exhausted their limits, and even a reset did not restore them.

The author built the ARPG Amberbrook in two weeks with GPT-6 Astra, Opus 5.5 and three.js, and has received a publishing offer.