05·Gateway API

One key. Every model, every host, every effort.

The gateway speaks the OpenAI Chat Completions dialect, so existing SDKs work by changing the base URL. What's new: the model id addresses a specific serving (host + quantization), and reasoning_effort is a first-class routing parameter — the two dimensions the analysis pages measure.

Base URL

https://models.fyi/api/v1

local dev: http://localhost:3000/api/v1

Auth

Bearer mfyi_…

create keys on /dashboard

Upstream status

mock placeholder

env-switchable to any OpenAI-compatible upstream

A·Quickstart

Three requests to fluency

1 · chat completion (host-pinned)

curl https://models.fyi/api/v1/chat/completions \
  -H "Authorization: Bearer $MFYI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5.2@fireworks",
    "reasoning_effort": "medium",
    "messages": [{"role": "user", "content": "Summarize this diff..."}]
  }'

2 · streaming (SSE passthrough)

curl -N https://models.fyi/api/v1/chat/completions \
  -H "Authorization: Bearer $MFYI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-5.5", "stream": true,
       "messages": [{"role": "user", "content": "hi"}]}'

openai sdk · javascript

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://models.fyi/api/v1",
  apiKey: process.env.MFYI_API_KEY,
});

const res = await client.chat.completions.create({
  model: "claude-fable-5",
  reasoning_effort: "high",
  messages: [{ role: "user", content: "Refactor this function..." }],
});

openai sdk · python

from openai import OpenAI

client = OpenAI(
    base_url="https://models.fyi/api/v1",
    api_key=os.environ["MFYI_API_KEY"],
)

res = client.chat.completions.create(
    model="deepseek-v3.2",
    reasoning_effort="low",   # the sweet spot for everyday tasks
    messages=[{"role": "user", "content": "Draft a reply..."}],
)

B·Model addressing

Ids that name a serving, not a vibe

Id	Meaning
glm-5.2	Reference serving (official host) of GLM-5.2
glm-5.2@fireworks	Same weights served by Fireworks (fp8) — different scores, price, speed
kimi-k2-thinking@groq	Groq serving — 720 tok/s, −3.5% quality; the trade-off is yours to make

Full catalog with per-variant pricing: GET /api/v1/models (no auth required). Every id and its measured scores appear on the model pages.

C·Effort routing

reasoning_effort, resolved honestly

Tiers: minimal · low · medium · high · xhigh. If a model doesn't support the tier you asked for, the gateway resolves to the nearest lower supported tier — the same rule the leaderboard uses.
Defaults are per-model (each model's page shows its default and its sweet spots). Omitting the field uses the default, never silently xhigh.
Usage records store the resolved effort, so your dashboard shows spend per tier — the bill finally matches the quality dial.

error shape (openai-compatible)

{
  "error": {
    "message": "Model 'glm-9' not found. List ids via GET /api/v1/models ...",
    "type": "invalid_request_error",
    "code": "model_not_found"
  }
}

D·Upstreams

Official providers, routed per model

Each model routes to its official provider's API — set a provider key and its models go live (translation to native dialects, e.g. the Anthropic Messages API, happens inside the gateway; the client contract never changes). GET /api/v1/models reports upstream: "live" | "mock" per model. Models without a key answer from the clearly-labeled built-in mock, so the full contract runs end-to-end either way.

environment

OPENAI_API_KEY=sk-...        # gpt-5.5
ANTHROPIC_API_KEY=sk-ant-... # claude-fable-5 · claude-opus-4.8 · claude-sonnet-5
GOOGLE_API_KEY=AIza...       # gemini-3-pro

# optional global fallback for unrouted models (any OpenAI-compatible endpoint)
UPSTREAM_BASE_URL=https://openrouter.ai/api/v1
UPSTREAM_API_KEY=sk-...