Insider leak: OpenAI's Astra tested as 'ultima-alpha' and 'vega-alpha' checkpoints
Ok_Display_3159 · reddit · 2026-09-03
Tester Lentils80 reports two GPT Astra checkpoints in testing: 'ultima-alpha' (likely the public release candidate) and 'vega-alpha' (a cybersecurity variant for select enterprises). Based on extensive hands-on testing, Astra runs fully autonomously for very long periods without a '/goal', is strong at orchestrating subagents, applies academic literature well on hard graphics-optimization problems, and writes the best-maintainability code of any model tested — though it's weaker than Fable on code, and weakest on frontend and 3D. Creative writing is noticeably better than 5.6 Sol with less slop. Typical usage will likely be Astra as the main agent with cheaper models as subagents.
Related event: OpenAI's Next-Gen Model Astra Reportedly Trained on Over 100K GPUs(10 posts)→
More from coding & agent
- LukeW built an agent-orchestration app for a foldable: managing agents is now an all-day job — LukeW · 2026-10-21
- Agentic SDLC: engineers shift from implementers to agent orchestrators as cycles shrink from weeks to hours — Pavan_Belagatti · 2026-10-21
- OpenAI's dots can now use company tools via Merge for Workforce with employee-bound permissions — shensi · 2026-10-21
- The topology of an LLM agent: where the model, loop, and tools run shapes cost and security — davidcrawshaw · 2026-10-21
- Agents with full-disk and accessibility access are an enterprise security nightmare, argues Runlayer — leveragedupside · 2026-10-21
- Acellera runs an unsupervised virtual lab of AI scientists on a real IL-23 molecular discovery program — gdefabritiis · 2026-10-21