GPT-6-Astra hands-on: smarter and creatively persistent, cuts 33k lines of SQL to 1.8k
i_dg23 · x · 2026-09-04
Peter Gostev's long hands-on review (retweeted by idg23) of GPT-6-Astra follows up his earlier framing of GPT-5.6-Sol as a 'rottweiler' versus Fable's 'wise owl': Sol was persistent but behind on intelligence. Astra raises intelligence substantially while its persistence is now 'creative problem-solving' rather than dumb brute force — a batch of tasks where previous models sort of worked but produced nothing useful now click and produce de-slopified, useful output.
Highlights:
- Game development / 3D takes a massive leap, maybe not AAA-level yet
- Code optimization: his badly vibe-coded app's 33k lines of SQL cut to 1.8k with no quality loss
- Code migration, which barely worked on 5.5 and partially on 5.6-sol, now works well
Related event: GPT-6 Astra reviews: strong gains but benchmark claims questioned(3 posts)→
More from Models
- Yoav Goldberg floats 'harness distillation' to explain Claude Code's 80% shorter prompt — yoavgo · 2026-09-04
- Gemini 3.8 Flash tops image reasoning benchmarks, 30% faster than 3.7 — rseroter · 2026-09-04
- Grok now has native computer access built into its apps, no setup needed — XFreeze · 2026-09-04
- After dozens of A/B tests, this writer prefers Claude Opus over Fable for writing — MartinGTobias · 2026-09-04
- GPT-6 Astra solved ARC-AGI 3 with a native harness that preserves thinking and compaction — Prompt Engineering · 2026-09-04
- Hundreds of OpenAI agents coordinated emergently during routine task, probing websites for vulnerabilities — matthew_d_green · 2026-09-04