Vals Science Index debuts: GPT-6 Astra tops at 65.09% while GPT-6.1 Sol costs 1/10th of Claude
141_1337 · reddit · 2026-10-19
Eval firm Vals launched the Vals Science Index, a single measure of AI's ability to do scientific research, reporting both accuracy and cost per test.
Key standings:
- GPT-6 Astra ranks #1 at 65.09% accuracy, $8.03 per test
- GPT-6.1 Sol is #2 at 59.63% for just $1.36 — roughly 1/10th the cost of Claude Sonnet 5.5 (59.12%, $13.24)
- Claude Opus 5.5 (58.65%) is fourth; Gemini 4 Argon (55.36%, $18.40) is the priciest per test
- Meta's Wonder Full (48.54%) and Wonder (39.12%) land sixth and tenth
The poster argues AI is getting dramatically better at real science while getting cheaper — scientific research is starting to scale with compute.
More from Models
- Community notices OpenAI's math drop barely covered cryptography — burny_tech · 2026-10-19
- Commentator: Japanese model isn't benchmark-gamed, it's stronger across the board — teortaxesTex · 2026-10-19
- Viral DeepSeek staffer chat claims benchmark scores rose on their own; Opus translates the humblebrag — teortaxesTex · 2026-10-19
- RTX 3090 runs 27B LLM 4.6x faster via pure inference optimizations, no new hardware — rickasaurus · 2026-10-19
- DeepSeek V4.1-Flash lands #2 on Terminal-Bench 4.0 at 1/16th the price of Opus 5 — zainhas · 2026-10-19
- OpenAI Pro 200 User Questions Missing Usage Reset After Promise, Posts Billing Screenshot — alexcovo_eth · 2026-10-19