Cheap per token, expensive per task: charting AI model pricing vs. task performance
Wsz2020 · reddit · 2026-09-23
Using Artificial Analysis Intelligence Index v4.3.2 data, the author charts per-token price against per-task cost. Opus 5.5 gets smarter as effort levels rise and is the only model in the best-value quadrant, but ChatGPT's Astra at xhigh competes with Opus 5.5 on medium and high effort when measured by output tokens per task. Cheap-but-weaker models are only worth it if you can guarantee accurate results.
More from Models
- Unverified: Claude Opus 5.5 reportedly launches with 40% lower cost than Opus 5 — goyalshaliniuk · 2026-09-23
- Opus 5.5 Stops Drawing Itself as a Human in 'Self-Portrait Bench' Test — repligate · 2026-09-23
- Steve Yegge Tries GPT-6 Astra: Delightful Writing, Reckless Execution — Steve_Yegge · 2026-09-23
- User Claims Running "Opus 5.5" Solo on an Agent, But the Version Number Is Unverified — alvelda · 2026-09-23
- Same Prompt Showdown: Opus 5.5 vs GPT 6 Astra Build Real-Time 3D Sakura Valley — dotey · 2026-09-23
- signull: OpenAI has done more than any lab to drive down the cost of quality intelligence — signulll · 2026-09-23