Where a coding agent's bill goes: output is under 10% of it
385 developer-days of our own team's coding-agent traffic at list prices: output is under 10% of the bill, the most expensive tenth of days is over half of it, and heavy days cost more per request, not only more requests.
Choosing GPUs? Compare GPUs for your traffic · Get your node measured, free
Disclosure first: the traffic below is our own team's, and it went through CodeRouter, a router our team built. TokenWatt, which did this analysis, is the same team.
Our cost-per-developer post drew a fair challenge: our numbers looked low next to the $5,000–7,000 a month some developers say they spend on agentic coding. So we looked at what separates a cheap developer-day from an expensive one. While doing it we found an error in that post: its dollar figures were CodeRouter's recorded costs, and CodeRouter had priced Claude Opus 4.7 at three times its list price until October 29. That post is now corrected; everything here is at list prices.
What we measured. The same five months as before (May to September 2026), one API key per engineer, 385 developer-days with at least 20 requests, 19 engineers. Every request priced at provider list prices: uncached input, cache reads, cache writes and output. Claude cache writes weren't logged before November, so we estimated them from November traffic (6.6% of the Claude prompt). Where the estimate matters, we give both numbers.
1. Output is under 10% of the bill
| Where the money went | With estimated cache writes | Without |
|---|---|---|
| Input not served from cache | 47% | 57% |
| Cache reads | 27% | 34% |
| Cache writes (estimated) | 18% | – |
| Output | 8% | 10% |
An agent sends its context again on every turn: the median request carried 69K tokens of context, and for every token the agent wrote it sent about 240. The code it writes is a small part of what you pay for; reading the repository, the conversation and the tool results again is most of it.
2. A few days are most of the spend
| One developer-day | Cost at list prices |
|---|---|
| Median | $4.30 |
| 75th percentile | $13.03 |
| 90th percentile | $36.69 |
| 99th percentile | $126.93 |
| Most expensive | $356.57 |
Without the cache-write estimate: $4.17, $11.52, $30.58, $112.72 and $211.31. Either way, the most expensive 10% of developer-days were 54–57% of the total.
3. Heavy days cost more per request, not only more requests
| Typical days (40th–60th percentile) | Top 10% of days | |
|---|---|---|
| Requests per day | 203 | 915 |
| Context per request | 66K tokens | 104K tokens |
| Cost per request | $0.022 | $0.090 |
| Claude Opus's share of the spend | 16% | 56% |
| Input served from cache | 83% | 77% |
| Cost per day | $4.44 | $82.65 |
Across all days, the number of requests and the cost per request each explain about half of the spread in daily cost. Cost per request went up with bigger contexts and with more of the work on Claude Opus; the cache hit rate was about the same.
Our most expensive 1% of days cost $150–230 each: about 2,000 requests, 120K tokens of context per request, 84–88% of the spend on Opus. A developer who works like that every working day spends $3,000–4,700 a month, close to the numbers people told us.
What this means for a bill
- Look at cost per request, not only request counts. Half the difference between a cheap day and an expensive one is what each request costs.
- Which model handles each request matters most per request. On our expensive days most of the money went to Opus; list prices differ several times between models.
- Context size is the other lever. Every extra 10K tokens of context is paid again on every turn. Keeping the cache warm helps, but on our expensive days the cache hit rate was no higher than on cheap ones.
Limits
- Our own team's development work, not customers'. 19 engineers; the busiest month, May, has more than half of the days.
- Routed traffic across several models and tools. Claude Code was 9% of requests; the rest came through OpenAI-compatible clients.
- Claude cache writes before November are estimated (6.6% of the Claude prompt, at the 5-minute write rate). Other providers don't report writes separately.
- List prices as of November 5, 2026, no volume discounts.
- We didn't grade the work done in this traffic: it shows what was spent, not what got solved.
- These are developer-days on a router. They don't tell you what the same work would cost on a single model; a different model takes a different number of steps. For that we run the same tasks on each setup, as in our 176-task benchmark.
If you want this breakdown for your own team, give every developer their own key for a week and look at requests, context per request and model per request, not only the total. CodeRouter gives each developer a key and shows spend per key; its Claude Code pricing page compares the plans.
Keep reading
What a coding agent costs per developer: 385 developer-days of our own usage
Five months of our own team's coding-agent traffic: 163 requests and $4.30 on a typical developer day, 378 requests and $13.03 on a heavy one, at list prices with routing (corrected November 7). How that compares with Claude's subscription and seat plans, what moved the bill, and what these numbers don't tell you.
Long prompts are where an inference node loses money: 1K, 8K and 16K input tokens on one H100
We measured how much one H100 can serve within a latency target as prompts grow from about 1K to 8K and 16K input tokens (Qwen3-30B-A3B FP8, vLLM, no prefix reuse). Capacity fell from 11.5 to 1.89 and 0.73 requests per second, because the first-token deadline caps how many long prompts the GPU can prefill at once. At a flat input price, long uncached prompts earn less than the GPU costs.
What a prompt cache is worth: the same H100 served 82% more requests at a 90% cache hit rate
We served Qwen3-30B-A3B-Instruct-2507 FP8 on our production H100 with vLLM and repeated the same latency-target search with 0%, 50%, 75% and 90% of every prompt drawn from shared prefixes. vLLM's own counters showed cache hit rates of 0%, 49%, 74% and 89%. Capacity rose from 11.5 to 21.0 requests per second, and revenue per GPU-hour from $10.37 to $18.84 at flat prices. With cached input billed at a quarter of the price, the node still earns more and the customer pays far less.