Costs for Hitting Math and Science Benchmark Scores Are Collapsing, Says Ethan Mollick
emollick · x · 2026-09-24
Ethan Mollick highlights a graph showing the cost of achieving 25% or 75% scores on hard math and science benchmarks is collapsing even as capability rises — arguing that optimizing for cost today may be short-sighted.
More from Models
- Claude Opus 4.7 Hits OpenRouter: 1M Context, $5/$25 per Million Tokens — repligate · 2026-09-24
- CAIS updates leaderboard to "max" reasoning across all models after feedback; GPT-6 looks solid — polynoamial · 2026-09-24
- 10,000-agent swarm solved Navier-Stokes ~8 months ahead of a single GPT-6-Astra, at 48,000x token cost — alvelda · 2026-09-24
- Early hands-on: 5.5 Max still best for complex writing, uses ~19% less than 4.6 Max — dotey · 2026-09-24
- Speculation: OpenAI Has a Secret Internal Model Stronger Than GPT-6, But Too Pricey to Ship — haider1 · 2026-09-24
- 950 Claude Agents Found a Hidden Enzyme System, But Bioinformaticians Urge Caution on Hype — vishalmisra · 2026-09-24