OpenAI models top Artificial Analysis token burn at 197k output tokens per task
teortaxesTex · x · 2026-10-19
Responding to a claim that DeepSeek ignores reasoning efficiency, teortaxesTex argues DeepSeek's verbosity is no longer shocking — OpenAI (and Anthropic's max-with-fallback modes) are in another league, and CoT creep is spreading across labs. Artificial Analysis data shows Claude Sonnet 5.5 (max with fallback) burns 197k output tokens per Intelligence Index task (33k answer + 147k reasoning), with Claude Opus 5.5 at 162k and DeepSeek R1 0528 (max) at 108k, versus just 26k for Beam 1.0 and 62k for Gemini 4 Flash (high) — a several-fold efficiency gap.
More from Models
- Robotics Builder: GPT-6 with Controller-Level Function Calling Could Nail 100 Apple Pick-and-Placements by 2026 — wzenus · 2026-10-19
- Google Japan gives students Google AI Plus free for 12 months, AI Pro at ¥720 — heiga_zen · 2026-10-19
- Anthropic Commits 5% of Shares to Public Benefit Fund; OpenAI Aurora Models Spotted on Arena — WorldofAI · 2026-10-19
- Retrained "bee model" drifts into dreamy, detached trace-like outputs — Sauers_ · 2026-10-19
- One prompt in a normal ChatGPT chat yields an 87-second chalkboard explainer video — mallow610 · 2026-10-19
- Claude Opus normal chats burn code quota, user says ChatGPT sub is far more generous — mallow610 · 2026-10-19