Brain · Output

Output

See which model is worth its bill - per thousand output tokens. A per-model rollup of unit economics from your Claude Code telemetry: cost per 1k output tokens, share of output, share of cost, sessions. The answer to “is the more expensive model worth it for this kind of work?”

See it in motion

Where to find it

  • Localhost: /quality-output.html?repo=<id>
  • API: GET /api/quality/output?repo=<id>&sinceDays=<days>
  • Keyboard: ⌘ K then output
  • Sidebar: Brain Level up Output

What it does for you

Compare models on the only number that matters: $ per 1k output.One row per model, sorted by output tokens - sessions, output tokens, cost, $ / 1k out, output share, cost share. The amber bars (output share) vs rose bars (cost share) tell you instantly whether a model is over- or under-earning its keep.
Period picker keeps the comparison fair.7, 30, and 90 day windows (90 days is the widest window the read supports). A model that was great two months ago but is now a rounding error stops dominating the longer-window view.
Attribution by language, on top of the per-model rollup.The page also breaks output down by language (“Sonnet is great for TypeScript, Haiku for prose”) and by language and model, from cc-telemetry's per-turn file_path capture. It shows an empty state only when the window holds no file-editing tool calls.

Configure

Period picker on the page. No other knobs - the data is pulled from cc.aggregate and analyzed by the pure lib/output-analysis.mjs module (tested, deterministic).

Use it well

Open Output before the next routing-tier decision. If a high-tier model has >50% of your spend but <50% of your output, you're probably over-routing. The LLM routing control tab is where you fix it - Output is where you decide what to fix.

Read more

Last updated