What an AI coding agent costs per merged pull request

COST PER MERGED PULL REQUEST

169 merged deliveries, measured on Orbi Cloud

The clearest unit is a merged delivery, not an abstract token quota. Orbi Cloud's usage records start on 2026-09-16; at 2026-10-05 15:36 UTC+8 they held 169 merged deliveries with token counts, every one of them on deepseek-flash in a Pi session. Per delivery the median cost was $0.082 off-peak, the 90th percentile $0.22, and the mean $0.11.

Amounts use DeepSeek's official deepseek-flash pricing at off-peak rates: cache hit $0.003 per million tokens, cache miss $0.15 per million, and output $0.60 per million. Peak weekday hours (01:00–04:00 and 06:00–10:00 UTC) double every figure, so the numbers here are the floor. The records keep one total per delivery with no per-request timestamps, so peak and off-peak cannot be separated: the true median is somewhere between $0.082 and twice that.

What the sample covers: Cloud's usage records, one row per issue and branch, written by Pi sessions. A resumed run adds to its row; a new attempt on the same branch overwrites it. 141 of the 169 were in Orbi's own three repositories, 11 in our organization's forks and test repositories, and 17 in 7 repositories that belong to Cloud users — those 17 had a median cost of $0.090. Most of the input is cache reads: per delivery, the median share of input tokens read from cache was 97.8%. See also Devin comparison, Cloud, and auto-merge AI PRs guide.

THE MEASUREMENT · PER DELIVERY (INCLUDING UNMERGED)

46 real deliveries, measured on our own repository

Every Orbi delivery runs two agent sessions — the implementation, then an independent review that can repair the branch and rerun the suite. As of 2026-09-12 we aggregated the session logs of n=46 deliveries of the orbi repository itself (Python, full CI, issues written to spec). The sample is whatever worktrees are on disk that day — it moves as worktrees get cleaned up — so n is a snapshot as of the stated date, not a permanent fact. Here is the complete per-delivery distribution of totalTokens — the count the provider actually bills, cache reads included:

READ THE LIMITS FIRST

Three things this page does not claim

• OUR REPO ONLY
  This is our repository's measured range — not a promise to everyone

  Python, a full CI suite, and issues written to spec keep our context short. A repository with longer context, slower tests, or vaguer issues lands elsewhere in the distribution. Treat the tables above as what they are: our repo's measured range.

• CACHING IS LOAD-BEARING
  95.9% cache reads assumes the provider supports prompt caching

  The whole cost structure rests on cache hits. On an endpoint without prompt caching — where every token bills as a miss — the same work costs roughly an order of magnitude more. Check your provider before extrapolating.

• TOTAL TOKENS, AS BILLED
  Numbers are totalTokens, cache reads included

  That is the line the provider actually bills. Counting only input + output would hide about 96% of the real consumption and flatter the number — the opposite of what this page is for.

WHY IT'S SO CHEAP

A cache hit costs 1/50 of a miss — and 95.9% of the work is cache hits

Each delivery re-reads the same repository context: the tree, the conventions, the issue, the failing tests. Prompt caching turns that repeated reading into billed hits at $0.003/1M instead of misses at $0.15/1M. That single ratio is why a full delivery — implementation plus independent review that can repair and rerun — costs cents. It is also the honest question to ask any BYOK agent vendor: what share of your tokens are cache hits, and what does your provider charge for each?

EARLIER SAMPLE · n=20 · KEPT FOR THE RECORD

Earlier sample: 20 merged pull requests, 22–24 September 2026

Before Cloud's usage records existed, we measured a smaller sample by hand: from 2026-09-22 09:44 UTC to 2026-09-24 12:03 UTC, 20 merged PRs in the orbi-build/orbi repository, every one of them on deepseek-flash (DeepSeek-V4.1-Flash). Their median was $0.125 off-peak and $0.249 at peak, over a median 10,612,802 tokens per merged PR; the median runtime was 60.4 minutes across a median 156 model requests. We keep this sample published, but the 169 merged deliveries above are the current figure.

These estimates use DeepSeek's official deepseek-flash pricing, checked 2026-09-24: cache hit $0.003 per million tokens, cache miss $0.15 per million, and output $0.60 per million. Peak weekday hours (01:00–04:00 and 06:00–10:00 UTC) are double. The sample's token mix: cache reads account for 97.1%; input 1.9%; output 1.0%.

Why this sample differs from the 46-delivery snapshot above: that one counts worktrees, including unmerged work and small changes; this one counts only merged PRs. The 20 PRs are the public Issues #1289, #1290, #1231, #1262, #1304, #1306, #1307, #933, #898, #856, #610, #1259, #1294, #1322, #1324, #1088, #1328, #1332, #1337, and #1336.