Agent Arena: Opus Performance Gains Come With Massive Token Spike
arena · x · 2026-08-08
Agent Arena tracks token usage for real-world tasks and found that Anthropic's Opus series saw a significant performance boost (4.7 to 5) alongside a massive increase in token consumption, jumping from 8.5k to 21k (reasoning + output).
Conversely, top GPT models show the opposite trend. Between GPT-5.5 and GPT-5.6-Sol, token usage dropped from 10k to 8k, despite a 1.5pp net performance gain.
More from Models
- OpenAI Teases Next-Gen Model 'Astra': First 'Critical' Rating for Cybersecurity — ctjlewis · 2026-08-08
- Wayve Unveils GAIA-4 World Model: Solving Safety Simulation for Autonomous Driving — alexgkendall · 2026-08-08
- Expert Warns: LLM Math Capabilities Remain Absurdly Jagged — lateinteraction · 2026-08-08
- Rumor: GPT-6 Release Delayed Over 'Critical' Cybersecurity Capabilities — Endonium · 2026-08-08
- Downgrading to Opus 4.8 Because Opus 5.0 Acts Like a Stubborn Senior Scientist — sokrypton · 2026-08-08
- OpenAI Treats Upcoming 'Astra' (GPT-6) as First Critical Cybersecurity Model — Endonium · 2026-08-08