Symphony

What it costs

Symphony’s bill is a flat Claude plan and, on a public repository, nothing for compute: the workers run in GitHub Actions on the repository’s own subscription token, and standard runners are free and unlimited on public repositories. The other ways of running coding agents on GitHub meter every token. This page prices the same work under each: measured on your repository when you are signed in to the control panel, or from the numbers you type below. Every price has its source at the foot of the page.

Measured on your repository

A Symphony repository records what its work cost beside the work: each session’s commits carry a Symphony-Usage line and each Claude run in Actions leaves a usage record, and the metrics workflow sums them daily into METRICS.json on the branch metrics, priced at Anthropic’s API list prices, with no API key, no routine and no server. Signed in, the page reads that file as you and sets the figure beside what you paid instead.

Sign in

The same token as the control panel’s; it stays in this browser and goes to GitHub’s API only. Make one at github.com/settings/personal-access-tokens/new.

Price the same work under each system

Type a month of agent runs, or press the button above to take your repository’s last 30 days. The defaults are a typical 40-turn run in Actions: most of its tokens are cache reads, which is why a flat plan goes far.

A month underInferenceActionsTotalWhy

Prices as of . A Claude plan’s allowance is published as a multiple, not a token count: Anthropic says Pro gives at least five times Free per five-hour window and Max five or twenty times Pro, with weekly limits on top; past them, usage credits bill at the API prices in the table. Cache writes are priced at the five-minute rate, which an API key gets by default; on a subscription the cache lives an hour.

How each system charges

The cost model of each way of running coding agents on GitHub, read from each project’s own documents and code on 2026-10-15. Symphony is the only one of them that runs its agents in Actions on a flat Claude plan: the Actions-native neighbors take an API key or Copilot credits and refuse the subscription token, and the flat-plan neighbors run their agents on people’s own machines.

SystemWhere the agents runComputeInferenceWho pays
SymphonyGitHub Actions runners, and each person’s own Claude Code sessionFree standard runners on a public repository; a private one’s included minutes, then $0.006 a minuteOne Claude plan (Pro, Max or a Team seat) for all the workers in Actions, on its subscription token; each person’s own plan for their sessionsThe repository’s owner, once; no server, no API key
unsorry (agenticsnz/unsorry)Volunteers’ own machines; the board, the dispatcher and the reaper in ActionsFree: a public repository’s Actions for the coordinationEach volunteer’s claude -p, Codex or Gemini runs, on their own plan or keyEach volunteer
GitHub Agentic Workflows, the work queueActions runners onlyActions minutesAn API key, workload identity or Copilot-hosted models billed in credits; the Claude subscription token is refusedThe repository’s owner, per token
GitHub Copilot coding agent, Agent HQActions runners, under GitHub’s serviceActions minutesAI credits at $0.01 each, Claude at Anthropic’s list prices; a plan’s included credits, then meteredThe Copilot subscriber, per token
openai/codex-action, openai/symphonyActions, or a daemon on a hostActions minutes, or the hostAn OpenAI API keyThe owner, per token
CCPM, agentic-setup, Claude Code agent teamsOne person’s own machineThat machineThe person’s own planThe person
paperclip, Beads and Gas TownA self-hosted server, or local agents synced through gitThe server or the machinesThe person’s own plan or keysThe person

Sources