62% of Your Agent Bill Is Context You Already Sent
A 2026 cost breakdown of real agentic coding sessions found the single biggest line item wasn't the model's output — it was re-sending the same context every turn. One team cut $87k/month to $24k without slowing down.
NeuroX AI · July 22, 2026
A 2026 breakdown of agentic coding costs sorted a real session's token bill by where the money actually went. The largest slice wasn't the code the agent wrote. Re-sent context was 62% of the bill — tool definitions took 14%, reasoning output 11%, system prompts 8%, and wasted retries the last 5%.
The mechanism is the stateless API. Every turn re-sends the entire conversation so far, so context compounds while output stays flat. A separate analysis clocked the ratio at roughly 25:1 input to output — by turn 50, a session was pushing 35,000 input tokens per turn against about 800 tokens of new work. You pay to re-read the transcript far more than you pay to extend it.
This is why premium model bills feel irrational. The 90th-percentile engineer runs about $1,650/month in tokens, and most of it is re-reading, not writing.
It's also fixable without downgrading the model. One SaaS team attacked context discipline — pruning transcripts, scoping tool surface, killing retry loops — and cut spend from $87,000 to $24,000 a month with no loss in output. That's $756k a year recovered by treating context as a budget, not a default.