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

The post compares Haiku, Sonnet and Opus 5.5 against each GPT-6 tier and concludes OpenAI is losing on all fronts, drawing a heated comment section.

Given only a virtual gamepad and 512px split-screen screenshots with no game state, Astra learned to double jump by trial and error and watched its teammate's moves, taking about 40 minutes for one run.

Epoch AI notes that newer models have already seen the original innovation during training, making the task easier, yet GPT-6 Astra and Claude Fable 5.1 still struggled to reimplement it.

GPT-6 rolls out globally to Plus, Pro, Business and Enterprise today and to Free and Go tomorrow, with Sol powering paid tiers and Luna powering free tiers in the Chat tab, while Work and Codex models stay unchanged.

The author shows handing repeatable operations to the Dreamina command line, driven by Codex and GPT-6 Astra.

charliermarsh says a project's README was written by a human, but all code changes, PR summaries and additional documentation were generated by GPT-6 Astra in Codex.

Opus 5.5, GPT-6 Astra, Sonnet 5.5 and GPT-6.1 Sol each got the same one-line prompt to recreate Sonic in Three.js with no reference image, to see which does best.

A team used meta-harness to train a 15M-parameter agent in three weeks that beats Craftax's final boss 44% of the time, versus GPT-6 Astra's 1% at 24 hours and $83.

A post says an anonymous user used GPT-6 Astra to improve a result from OpenAI's math announcement, with a comparison image attached.

The author shares the full prompt used to build a spaceport project with GPT-6 Astra and Opus 5.5, along with the finished video.

A bare comparison post asserts that Claude's Opus 5.5 is better than two GPT-6 versions, with no evidence offered, and predictably stirs up the comments.

The same prompt for a plush whale: Opus 5.5 spent $11.40 on a cute one, while Astra spent $8.66 on a goofy open-mouthed one.