Index · pro

API Price Index Pro

The prices are free — the whole dataset is one GET awayunder CC BY. What's not free is the decision: given a model you've picked, which host should you actually buy it from right now, on priceand first-party measured speed — with the grain that makes routing safe: tail latency, error rate, how many samples back the number.

One MCP tool, recommend_endpoint. You give it a model (or a tier); it returns the best endpoint on the Pareto front of price × measured throughput, plus the reserved grain — ttft p95,error_rate, samples, per region. The answer an agent can route on without a human in the loop.

Where the free line is

It's a line of access mode, not a paywall on facts. A human reading the efficient frontier or atier page in a browser — including the Pareto chart and the headline p50 latency — stays free, forever. Anagent pulling the front, the tail latency and the routing decision on every call, at scale is what this service is for. The headline (price, throughput, ttft p50) stays in the free dataset as a citation; the decision grain (p95, error rate, sample counts) is what the tool reserves.

What `recommend_endpoint` answers

“I've chosen deepseek-v4-flash — where do I buy it?” → the host with the best measured throughput per dollar, the full non-dominated front behind it, and the reserved grain for each candidate. Concretely, per recommended endpoint:

Price — blended, input and output USD / 1M tokens.
Measured speed — tokens/s and ttft p50 (headline), plus reserved ttft p95, error_rate, samplesand n_runs — first-party benchmarks, not vendor claims.
The Pareto front — every endpoint that isn't beaten on both price and speed at once, so you see the real trade-off, not one cherry-picked row.
Region-aware — eu-paris or us-east; throughput and latency are measured per region, not averaged into a fiction.

Objectives: balanced (throughput per dollar on the front, default),cheapest, fastest, or pareto (returns the whole front). Hard constraints: max_price (blended $/1M),min_tokens_per_s. Coverage v1: the Chinese flash andflagship tiers; the annex and price history follow.

Connect it

A remote MCP server over HTTP. Every call needs a valid key in theAuthorization header — the buyer configures it once in their client.

claude mcp add --transport http apipriceindex-pro \
  https://mcp.apipriceindex.com/mcp \
  --header "Authorization: Bearer <your-key>"

Endpoint: https://mcp.apipriceindex.com/mcp · headerAuthorization: Bearer <key>. Works from any MCP client (Claude, Cursor, or your own agent). No key, or a revoked key → an explicit refusal, never a silent empty answer.

Get a key — free during beta

Keys are issued by hand while the service is in beta, and they're free: the point right now is to see who routes on what before any billing exists. Tell us what you'd call the tool from and which models or tiers you care about, and we'll send a key.

Request a beta key →

Each call is logged (key, tool, params, timestamp) to measure usage before billing is wired — a revoked key stops working but the trace is kept. Fair-use during beta; the data served is first-party and measured, covered by theterms & disclaimer.

Free tier vs Pro

	Free (CC BY)	Pro (keyed MCP)
Prices (input/output/blended)	Yes — all.json	Yes
Throughput & ttft p50	Yes (headline)	Yes
Pareto chart	Yes — on the pages	Yes — as data, per call
ttft p95 · error rate · samples	—	Yes (reserved grain)
Routing decision as a tool	—	Yes — `recommend_endpoint`
Access	Browser / open dataset	Agent, on every call

Looking for the free prices in your agent instead? The openMCP server and dataset serve verified prices under CC BY — this Pro tool is only for the measured routing decision.