Same agent. Different model. Different behavior.
The certification layer for model swaps — built on the Inference Index, the live price-per-quality index for LLM inference.
NarasingaMoorthy V · Pranav Guhathakurta · amperesai.github.io/inference-index
Every agent platform re-routes the model beneath its agents — for cost, availability, routing. The prompt is identical; the behavior is not: different tools called, malformed arguments, broken JSON, silent refusals. There is no standard, portable pre-swap test — every team hand-rolls its own, or ships and prays.
Our research (under review at TMLR, 2026): downshifting a workload 70B→8B cuts cost 98.35% — but when 10 independent judges scored the identical output pairs, preservation ranged 61–97%, median ~90%. A single judge is an anecdote. A panel with a disclosed spread is a measurement. Nobody was pricing this — so we built the instrument.
Run your agent's task suite across a baseline and N swap candidates on any OpenAI-compatible endpoint. Five behavioral dimensions. One verdict: CERTIFIED / CAUTION / FAIL.
The live market index that prices LLM inference by $/MQT — dollars per million quality-adjusted tokens. 40 models, 11 providers, every price linked to its official source. Verify decides if you can swap; the Index decides what you swap to.
The collection layer: website-content-crawler over 11 official pricing pages (JS-heavy, bot-gated — the honest reason a headless crawler exists) + status pages. Ready-to-run actor inputs in the repo.
The nervous system: schedule → Apify → LLM price extraction → diff vs live index → Slack market-move alert → GitHub dispatch. Importable workflow JSON in the repo.
The Pilot Console: workload sliders → per-workload savings simulation, fetching the live index data. One-paste build prompt in the repo.
Recompute → commit → redeploy. A market move becomes a live index change in minutes. Zero servers anywhere.
Live on stage: three swap candidates race through the task suite — every tool call, argument, and schema check on screen — then the verdict stamps land. Simulation clearly labeled; the identical console renders your fleet's real run from one CLI command.
Third-party $/MQT scores for Foundation Model API serving — price-performance proof no vendor deck can match. Verify runs against serving endpoints natively.
Frontier coding spend routed by quality-adjusted price; code-task panel runs next on the roadmap.
The flagship verified pair runs on exactly the open models Nebius serves — third-party economics for their catalog.
Free neutral index → inbound → Verify runs on real fleets → paid pilots on Amperes, the inference control plane underneath.
Early conversations underway — no signed pilots or endorsements yet. That candor is the pitch.
It's live right now. One command runs Agent Verify against your endpoints — Databricks serving, Nebius, OpenAI, Groq, vLLM. We're signing three fleet design partners to turn sample reports into certified swaps on real traffic.
Amperes — the AI inference control plane · amperes.pro