See what each AI-written pull request costs

Record every model call against the task that made it, then roll the task up into the pull request it produced: a PR's cost is its implementation run plus every review and follow-up on its branch. ProPR keeps that record for all five agents and shows the model, tokens and cost on the task and in the pull request summary, with no token markup on top.

Cost estimation for pull requests starts before the run

The biggest cost driver is how much code the model reads. Planning shows which files the planner will read, ranked by relevance, with the selection's estimated cost, and you widen or trim it before a token is spent.

The token bar after a context refresh: 94k of the 588k-token budget used (16.1%), with the estimate for generating the plan — The context budget of a plan: the files selected and what reading them will cost.

Cap the run

Set a spend limit per task, per repository in .propr/workflow.yml, or for the whole instance. At the cap, ProPR stops the agent, publishes the partial work as a pull request and tells you in the Inbox and on GitHub. Cost and usage.

Task timeline: the run reaches its spend cap, an estimated $5.03 of a $5.00 cap from .propr/workflow.yml, and ProPR stops the agent and creates the pull request — A run stopped at its spend cap, with the partial work published as a pull request.

Per-task and per-PR usage

The usage log ties every model call to its model, tokens, cost, duration, task and repository. The pull request carries a run summary with status, duration, token cost and model, and every review and follow-up on that PR lands on the same task history, so the running total of a branch is one click away until it merges. Every task's record.

The LLM log: each model call with its model, tokens and cost, from a $1.31 Claude Opus 5.5 implementation to a three-cent repository summary — The usage log: each call with its model, tokens, cost, duration and task.

Cost per merge when using multiple models

Put several model labels on one issue and ProPR opens one pull request per model, each with its own cost, so the comparison rests on real diffs. Over time, Analytics reports agent efficacy by model: where each model's review score starts, where fixes take it, how many runs its pull requests need to merge, and its merge rate. Historical stats count tasks, success rate and spend over 7 or 30 days.

Route by phase once the numbers are in: the expensive model for planning and review, a faster one for routine implementation, with one branch and one history across all of them.

Agent efficacy by model: evaluated PRs, initial and final review score, score delta, runs to merge and merge rate for Claude Opus 5.5, GPT-6 Astra and Claude Fable 5.1 — Agent efficacy by model: initial and final review scores, runs to merge, merge rate.

Subscriptions versus API keys

Claude Code, Codex and Antigravity CLI run on the provider subscriptions you already hold; OpenCode and Mistral Vibe run on their own plans or keys. A subscription has a flat price and usage limits, so the question becomes which tasks spend the capacity. Agent Tank, ProPR's open-source tracker, measures provider usage before and after each task and charges the difference to that task, for those three agents.

API keys bill per token at the provider, which buys exact per-task accounting and provider-level controls. Mix the two: subscription-backed agents where they cost less, API billing where metering matters, and the usage log records every call either way.

One run on Claude Opus 5.5: 19 minutes, 2.4M tokens in, and 2.4% of the Claude session and 0.6% of the weekly window — Agent Tank on a task: the subscription capacity this run consumed.

What the per-PR number leaves out

A pull request's token cost sits beside time nobody prints: cleaning the branch, writing the PR description, turning review comments into another agent session. ProPR does those on the same branch, and the cost comparison weighs them against seats, credits and cloud tasks.