GPT-6.2 Astra claimed to top pro & science evals at under one-third the per-task cost of GPT-6
Dr_Singularity · x · 2026-11-18
Account DrSingularity claims GPT-6.2 Astra "sets new highs on evals across professional work and scientific research at less than a third of the cost per task of GPT-6 Astra," posting three benchmark charts:
- AutomationBench 1.0.6 (realistic business workflows: sales, marketing, ops, support, finance, HR): GPT-6.2 Astra leads the low-cost segment, scoring 50%+ under $1;
- Vals Science Index v1 (biological data analysis, system reasoning, research workflows): 68% at around $4, clearly above Claude Fable 5.2 (64% at $8.4) and Claude Opus 5.5 (59%);
- Agents' Last Exam v1.1 (long-horizon professional workflows): 78% at roughly $1, far cheaper than other models at similar scores.
Note: from a third-party account, not officially confirmed by OpenAI.
More from Models
- GPT-6.2 Sol paints Mount Fuji in 15 minutes for ~$3 on Ultrafast API rates — gabrielchua · 2026-11-18
- GitHub Copilot Fusion Tops Terminal Bench 4.1 at Half the Cost — jennifermarsman · 2026-11-18
- Ethan Mollick on GPT-6.2 Astra: impressive but hard to steer, 'uneven and weird' — emollick · 2026-11-18
- User Claims New GPT 6.2 Astra Got Worse at Making Slides — zck · 2026-11-18
- Ethan Mollick on GPT-6.2 Astra: hard to steer, but delivers on ambitious projects — emollick · 2026-11-18
- Polymarket's GPT Astra release market hits 100% as OpenAI ships GPT-6.2 — Polymarket · 2026-11-18