The funnel
1
Estimate (free)
Anyone gets the cost table — no signup.
→
2
Sign in
GitHub → 5 free credits.
→
3
Run
Reserve credits, benchmark, settle & refund the unused hold.
→
4
Decide
Cheapest model that passes. Buy more credits to keep going.
What a run does (cost per correct answer)
For each case × each target model: call the model, check the answer, measure tokens & cost.
Rank by cost ÷ correct answers — so a cheap model that passes beats an expensive one.
You're charged only the actual spend; the rest of the reservation is refunded.
QualityMax ↔ TokenHunger over MCP
QualityMax ⇄
TokenHunger →
Model providers
QM → TH
① server: QM (or any agent) calls TokenHunger's MCP tools —
estimate, run, list_models — with a th_… token. Runs debit credits.TH → QM
② client: a run with
casesFrom pulls cases from QM's MCP (e.g. list_test_cases), maps rows → {input, expect}, and benchmarks on real QM data.TH → models
runs each case across the chosen models, scores, and returns the cost-per-success ranking.