LIVE
AGENT·VERIFY → PILOT·CONSOLE →
as of —

Tokens are commodities. Quality isn't. The first index that prices LLM inference by $/MQT — dollars per million quality-adjusted tokens, scored by a 10-judge evaluation panel.

98.35%
verified cost reduction · flagship pair (judge-independent)
10
independent judges per verified preservation score
90%
panel-median quality preservation · flagship pair
61–97%
full judge spread — why single-judge scores are fiction

Market Index — scraped from official pricing pages

#ModelProvider In $/1MOut $/1MBlended $/1M Panel preservation$/MQTStatus
loading index…

Blended = 3:1 input:output. Every price links to the official page it was scraped from. Panel preservation appears only where the 10-judge panel has scored a verified route pair — no modeled guesses, ever.

Verified route pair — flagship

Llama-3.1-70B → Llama-3.1-8B

The identical response pairs were re-scored by 10 independent judge models. Cost reduction is arithmetic and judge-independent; preservation is reported as the panel median with the full spread disclosed.

QIX = savings × preservationmedian
    = 98.35% × 0.90 = 88.5
judge spread: Mistral-7B 97% (lenient) · Qwen-32B / DeepSeek-67B 82% · Gemma-27B 74% (strict)
⚠ Real-traffic caveat, stated up front: on live chat workloads the cheap model is usable ~92–95% of the time, but rated equivalent-or-better only ~21–32%. Downshifting is a managed trade, not free money — which is exactly why routing needs escalation and an honest index.
QUALITY-PRESERVED SAVINGS
QIX 88.5

88.5% of this workload's spend is eliminable at panel-median quality — on the open models served by Nebius, Together & co.

Pilot tracks — seeking design partners

Provider track · target: Databricks

A neutral, third-party $/MQT score for Foundation Model API serving — price-performance proof no vendor deck can match.

Buyer track · target: Cursor

Frontier coding spend routed by $/MQT instead of raw price — code-task panel runs are next on the roadmap; the general-benchmark pair is verified today.

Open-model track · target: Nebius

The flagship verified pair runs on exactly the open models Nebius serves — third-party economics for their catalog.

Proposed tracks in early conversations — no signed pilots, partnerships, or endorsements yet.

How it stays live

Apify crawls 11 official pricing pages (they're JS-heavy — that's why a headless crawler, not curl) → n8n extracts & diffs prices every 6h, fires market-move alerts, and dispatches to GitHub → GitHub Actions recomputes $/MQT and commits data.json → this page redeploys. Zero servers. A market move becomes a live index change in minutes. Bolt powers the Pilot Console — paste a workload, get a per-workload savings simulation. Full pipeline in the repo.