Arena reports that the top 10 agents on its Agent Arena platform process roughly 93 input tokens for every output token they generate, a ratio that helps explain why AI labs are competing to reduce the cost of cached input.
Agent Arena platform · Token usage analysis
Top AI agents read about 93 tokens for every token they write
Input-heavy workloads explain why AI labs are racing to cut the cost of cached input, the context that agents resend at every step of a long task.
93:1
Input tokens to output tokens across the top 10 agents
26,000+
Agent sessions included in the analysis
Input
Instructions, tool definitions and conversation history dominate each call
Input vs. output tokens, drawn to scale
Each column height is proportional to token count (93 vs. 1)
93
1
Input tokens
Output tokens
At this ratio, the output token is a sliver: most of the work an agent does is reading its own accumulated context.
Why input cost becomes the constraint
A repeating loop in multi-step agent workflows
1. Agent steps
Each step resends instructions and history to the model
→
2. Cached input reused
Previously processed context is reused instead of recomputed
→
3. Lower cost per run
Long, multi-step workflows become cheaper to run
Arena's framing: as agents take more steps and carry more context, input cost decides whether they can scale economically.
Reading the number
A platform-level observation, not a universal benchmark
Selection
How the top 10 agents were chosen is not detailed
Verification
No independent check of the session data is included
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...