The Agent Burn Index · October 2026
How many tokens do AI coding agents use?
A coding agent sends a median of 99,554 input tokens with every request, and 10.7% of requests carry more than 272,000. A typical active day is 94 requests and 8.5 million tokens. 87% of the input is served from cache, which is what keeps the bill down.
Half of coding-agent requests carry more than 100K input tokens
The 90th percentile is 279,966 tokens and the 99th is 533,304. At list prices, Claude Haiku 5.5 charges more per token above 100K and GPT-6 Luna above 272K, so this is where agent bills split.
Most of what an agent sends is a cache read, and that keeps the bill down
Agents resend the same files and history on every turn, and providers bill repeated prefixes at a fraction of the price. At Claude Sonnet 5.5 list rates (see the cost method below), caching took the average request from $0.27 to $0.05.
- Cache reads 87.2%
- Fresh input 12.0%
- Cache writes 0.9%
All 10.6 billion coding-agent input tokens, as the upstream provider reported them. Instruction files and MCP tool schemas ride along on every request as part of that prefix; contexttax measures how many tokens they add per agent.
Agents read 191 tokens for every token they write
The median reply is 214 tokens, the mean 698 and the 90th percentile 1,450.
- Input
- Output 0.5%
A session is a burst of 16 requests in about 8 minutes
The median session moves 1.5 million tokens. The mean is 67.5 requests because long autonomous runs pull it up.
1,175 sessions. A session ends after 10 minutes without a request on the same key.
Coding agents keep office hours: 10:00 to 18:00 UTC carries 44% of requests
That is a third of the day. The quietest hour, 02:00 UTC, has 1.3%.
- 10:00 to 18:00 UTC
- Even spread, 4.2% an hour
What a typical agent day costs at list prices
The median day is the one in the answer above. A busy day, the 90th percentile, is 491 requests and 86.7 million tokens.
| Model | Median day | Busy day |
|---|---|---|
| Claude Opus 5.5 | $7.06 | $57.94 |
| Claude Sonnet 5.5 | $3.53 | $28.97 |
| GPT-6 Luna | $0.23 | $2.38 |
Method: every request in an account-day is priced at the model's list rates for its own fresh input, cache reads, cache writes and output (GPT-6 Luna at its higher rate above 272K), then summed. Medians over 297 active days. Prices from claude.com/pricing and developers.openai.com, checked 2026-10-13. DeepSeek, GLM and Qwen are left out because each is sold by many providers at different prices, with no single list price to apply.
How we measured
- Source
- Standard Compute gateway request logs, Sep 13, 2026 to Oct 12, 2026 (30 days, UTC). 487,662 completed, metered requests. Health checks and our own team's accounts are excluded.
- Coding agents
- Requests whose user agent or X-Title header names a coding agent, such as OpenCode, Kilo Code, Roo Code or GitHub Copilot Chat: 79,276 requests. Clients that do not identify themselves count only toward all-traffic figures.
- Tokens
- As the upstream provider reported them. Input includes cache reads and cache writes.
- Sessions and days
- A session is a run of requests on one API key with no gap over 10 minutes. An active day is one account on one UTC calendar day with at least one request.
- Weighting
- Every request counts once, so busy accounts weigh more. Weighting accounts equally instead, the typical account's median request was 79,565 input tokens.
- Privacy
- Only pooled aggregates are published. Every breakdown covers at least 50 accounts; nothing about any single customer appears here.
- Refresh
- Recomputed monthly by script. The figures on this page are in data.json.
Questions
How many tokens does an AI coding agent use per request?
In our October 2026 data the median coding-agent request sent 99,554 input tokens. The 90th percentile was 279,966 and the 99th was 533,304. Replies are short: the median request returned 214 output tokens.
How many tokens does a coding agent use in a day?
A median active day was 94 requests and 8.5 million tokens. A busy day, the 90th percentile, was 491 requests and 86.7 million tokens.
How much does prompt caching save on coding agents?
87.2% of coding-agent input tokens were cache reads. At Claude Sonnet 5.5 list prices that cut the average request from $0.27 to $0.05, 81% less than the same tokens billed uncached.
How much does it cost to run a coding agent for a day?
At list prices, a median active day cost $7.06 on Claude Opus 5.5, $3.53 on Claude Sonnet 5.5 and $0.23 on GPT-6 Luna. A busy day (90th percentile) cost $57.94, $28.97 and $2.38.
Do coding agents use more tokens than other AI apps?
Yes, per request. Across all 487,662 requests on our gateway the median was 68,471 input tokens and 32% passed 100K. For identified coding agents the median was 99,554 and 49.7% passed 100K.
Can I cite or reuse these numbers?
Yes. Cite "Standard Compute, The Agent Burn Index, October 2026" with a link to https://standardcompute.com/agent-burn-index. The figures are pooled aggregates from Sep 13, 2026 to Oct 12, 2026; the full data file is at https://standardcompute.com/agent-burn-index/data.json.
Cite this report
Standard Compute. “The Agent Burn Index, October 2026.” Gateway data, Sep 13, 2026 to Oct 12, 2026. https://standardcompute.com/agent-burn-index
Run your agents on a fixed monthly budget, with the routing behind these numbers.
Plans from $19/mo