GITHUB → OPENROUTER
Same workload.
Compare model costs.
Compare current OpenRouter prices against a public repository’s sampled workload. Inspect the evidence and change the assumptions.
Public repositories only. No code is executed or sent to a model.
Compare the cost
Estimate · current rates
Solid: lower estimateCost using the lower modeled token count. The same meaning applies to the grey baseline and coloured alternatives.
Shaded tail: upper estimateThe lighter tail extends from the lower to the upper modeled cost. It is a scenario range, not a statistical confidence interval.
Fixed baseline scaleThe baseline upper estimate is 100%. Empty track is unused scale, not another cost. A › marker means the estimate exceeds the scale; its full dollar value remains shown.
Increase retries to test whether the cost advantage holds.
Context, output limits and tools are checked. Similar coding quality remains an assumption; increase alternative retries to stress-test the price advantage.
Observed work
Work types come from commit titles, not a code-quality assessment.
What this report establishes
- Commit dates, titles and attribution are GitHub metadata. A Claude trailer or model name is a self-reported claim, not proof of model usage.
- Only the latest up to 30 default-branch commits are inspected. Merge commits are excluded from task costs to avoid counting merged work twice.
- Dates are shown as evidence. We do not infer the historical model from release dates or price the past with today's rates.
- Features, fixes, tests, docs and other are deterministic title classifications. One commit can include several kinds of work.
- Three category-diverse non-merge commits are sampled. Actual added text and parent-file character counts anchor the workload range. Context reads, rewrites, reasoning and attempts remain explicit assumptions. Current catalog prices and a separate platform fee determine the comparison. No historical bill or measured savings is inferred.
- Candidates are recent tool-capable models from GLM, DeepSeek and Qwen that fit the upper workload range. Coding-oriented or full-size names are preferred to the cheapest small model. Every family has an editable model selector. This is a compatibility shortlist, not a quality score.
- Opening a commit fetches its patch on demand. GitHub may omit or truncate patches. Files and patches shown are capped and marked; this is not a full code audit.
- This is a cost estimator. It does not prove that two models solve the same task equally well. Increase the alternative retry allowance to test how much of the price advantage survives.