GPT-6 'Astra' expected tomorrow: codename spotted in API, rumored 4x GPT-5.6 Sol
daniel_mac8 · x · 2026-09-03
danielmac8 compiles leaks pointing to an imminent GPT-6 "Astra" launch: the codename returns a 404 in the OpenAI API while other non-existent model names return 400, suggesting it's staged.
Key rumored details:
- Sam Altman claims human parity at computer use at superhuman speeds
- Per @alexeheath, trained for ultra long-horizon tasks with a new 'persistent' reasoning effort
- The Information reports a new 'recurrent depth' reasoning algo that processes data recurrently in Transformer layers, acting like a much larger model
- 4x stronger than GPT-5.6 Sol on an internal ExploitGym benchmark
- Speculation: end-to-end multi-agent orchestration, natively delegating to Luna, Terra and Sol
The author expects the AI model release of the year, a leap comparable to GPT-3 to GPT-4.
Related event: GPT-6-Astra Codename Surfaces in OpenAI API Ahead of Rumored Launch(7 posts)→
More from Models
- GLM-5.3 hits 310 tok/s on Databricks inference, claimed strongest OSS coding model — jeremyphoward · 2026-09-03
- Fable 5.1 benchmarked against 8 rivals across coding, 3D, SVG and data-viz — arena · 2026-09-03
- Databricks: GLM-5.3 hits 310 tok/s, claims strongest open-source coding model — DavidSKrueger · 2026-09-03
- Qwen3.8-Flash-Next may be the best open model on 128GB unified memory — victormustar · 2026-09-03
- GPT-5.6 and Claude feel slow — where are the diffusion coding models? — viksit · 2026-09-03
- Meta's Muse Spark 1.3 reportedly tops Gemini 3.8 Flash 62-59 hours after launch — AIatMeta · 2026-09-03