BREAKING
Agent Arena Maps Token Efficiency
More Tokens Don't Mean More Quality
Net Improvement by Model
Claude Fable 514
Opus 4.89.2
GPT-5.58.04
GLM-5.25.1
Efficient vs Wasteful Agents
GPT-5.5Frontier
High gains, fewer tokens
On efficiency frontier
Grok Build 0.1Negative
Over 20K tokens used
Negative improvement
0
sessions
0
tasks
0
models
Efficiency Shapes Real-World Use
AI NEWS BLITZ
arena.ai just published Agent Arena data comparing how efficiently AI agents spend tokens.