Inference API

The Critique Inference API is token-in/token-out chat for sidecars, eval harnesses, CritiqueCode, and internal tools. Completions bill at published $/1M input and $/1M output rates. They do not consume the PR review credit pool. The same crt_ keys still authenticate.

Marketing overview, rate cards, and an API key gate live at /inference-api. Long-form background: Why we built the Inference API.

Authentication

write:builder also authorizes chat completions if your key predates inference scopes.

Endpoints

Non-streaming responses include X-Critique-Estimated-Usd, X-Critique-Input-Usd, and X-Critique-Output-Usd. Inference does not charge review credits.

Voice transcription

POST /api/v1/audio/transcriptions is speech-to-text for CritiqueCode /voice. It is not a chat model. Send base64 audio; the only hosted id is qwen/qwen3-asr-0.6b (Qwen3 ASR 0.6B). Billing is $0.000003 per second of audio, not $/1M tokens. Session commands: CritiqueCode voice mode. Product notes: Talk to code.

curl https://critique.sh/api/v1/audio/transcriptions \
  -H "Authorization: Bearer crt_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen/qwen3-asr-0.6b",
    "input_audio": { "data": "BASE64_WAV", "format": "wav" }
  }'

Coding auto router

openrouter/auto ranks models by seven-day market spend. That is a weak prior for coding agents.

critique/auto is a session coding selector over the Critique-resold Inference catalogue. It does not search OpenRouter's full model list.

Two rungs. A new session starts on the efficient rung (Seed 2.0 Code, DeepSeek V4 Flash 0731, DeepSeek V4.1 Flash, Gemini 3.8 Flash, Qwen3.8 Flash, Muse Glimmer, Qwen3.8 27B, Hy3, Ling 3.0 Flash). critique/auto-best starts on the frontier rung (Kimi K2.7 Code, GLM-5.3, Qwen3.8 2.4T, KAT-Coder Pro, MiniMax M3, Trinity, Kimi K2.6, Kimi K3, Fugu Ultra, Grok 4.6).

Sticky. The first turn picks a specialist. Later turns in the same session reuse it so the prompt cache stays warm. Send X-Critique-Router-Session from the previous response. If you omit it, Auto mints a new session.

Escalate once. Repeated tool failures move an efficient session onto the frontier rung. Auto does not re-pick every message, and the router-brain cannot jump rungs on a healthy session.

Vetoes. Screenshots cannot land on text-only models. Oversized prompts cannot land on a 128K specialist. Weak tool-callers cannot land on an agent loop.

Token rates follow the routed model. The JSON model field stays critique/auto*.

Response headers: X-Critique-Routed-Model, X-Critique-Router-Session, X-Critique-Router-Rung, X-Critique-Router-Escalated, X-Critique-Router-Task, X-Critique-Router-Confidence.

curl https://critique.sh/api/v1/chat/completions \
  -H "Authorization: Bearer crt_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -H "X-Critique-Router-Session: crs_…" \
  -d '{
    "model": "critique/auto",
    "messages": [
      { "role": "user", "content": "implement the retry queue in src/billing.ts" }
    ],
    "tools": [{ "type": "function", "function": { "name": "read_file" } }]
  }'

Marketing overview: /inference-api#auto. Essay: How Critique Auto picks a coding model.

Hosted models

These are the Critique-resold Inference offers (rate cards, training opt-in, marketing on /inference-api).

The chat-completions proxy also accepts any model id from the /pricing review catalog (for example anthropic/claude-opus-5, openai/gpt-5.6-terra). Those passthrough models are not Critique resell offers: they are not featured on the Inference API page, and usage is metered from upstream cost at token rates.

All launch (resold) models route through Western sweetener servers. Critique does not train on Inference API payloads by default.

Log retention opt-in (75% off)

Eligible Inference API models (including DeepSeek V4 Flash 0731, DeepSeek V4.1 Flash, Qwen3.8 Flash, Qwen3.8 2.4T A95B, Qwen3.8 27B, Seed 2.0 Code, Muse Glimmer 30B, Muse Spark 1.3, Laguna S 2.1, Ling 3.0 Flash, Hy3, Nemotron, MiniMax M3, KAT-Coder Pro, Kimi K2.7 Code, Kimi K2.6, GLM-5.3, and Trinity Large Thinking) keep the private tier at full rates, or opt in to short-term log retention at 25% of list price (75% off). Grok 4.6, Kimi K3, Gemini 3.8 Flash, and Fugu Ultra are list-price only — they are not on this deal.

Western hosting applies either way. This deal does not apply to PR review runs or Builder.

Quickstart

curl https://critique.sh/api/v1/chat/completions \
  -H "Authorization: Bearer crt_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek/deepseek-v4-flash-0731",
    "messages": [
      { "role": "user", "content": "Summarize this retry policy in three bullets." }
    ]
  }'

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.CRITIQUE_API_KEY,
  baseURL: "https://critique.sh/api/v1",
});

const res = await client.chat.completions.create({
  model: "deepseek/deepseek-v4-flash-0731",
  messages: [{ role: "user", content: "Draft a TypeScript type for idempotent enqueue." }],
});

Usage dashboard and limits

Signed-in operators get a full dashboard at /inference-dashboard:

• USD and tokens over time (7 / 30 / 90 days)
• Model and API-key attribution
• Limit status vs caps
• Paginated activity log

Settings → Connections includes a mini Inference API panel:

Settings API: GET / PATCH /api/settings/inference-api. Dashboard data: GET /api/dashboard/inference-api (?days=30 or ?view=activity).

Privacy and acceptable use

• Payloads are processed to return completions and meter token spend — not sold and not used to train foundation models unless you explicitly enable log-retention pricing on eligible models.
• Traffic stays on Western-hosted capacity; Critique does not mirror customer prompts to non-Western training pipelines.
• Short acceptable-use copy lives on /inference-api#policy.

Billing

Inference bills $/1M input and $/1M output on the published rate card. Prepaid credits can be added from the Inference dashboard; each top-up adds a 2.5% fee. Payment checkout runs through Autumn, which owns the customer balance, hosted Stripe checkout, and payment confirmation. Configure AUTUMN_SECRET_KEY, AUTUMN_INFERENCE_TOPUP_PRODUCT_ID, and AUTUMN_INFERENCE_FEATURE_ID on the server. The feature quantity is USD cents, so the Autumn one-off plan should grant the selected quantity of inference_usd_cents. No client redirect or custom payment webhook is trusted to credit an account.

Send X-Critique-Billing: byok to use saved OpenAI, Anthropic, OpenRouter, Crof, or LLM Gateway keys. OpenAI model ids route to api.openai.com. BYOK does not debit prepaid Critique Inference credits.

Insufficient spend against an Inference USD cap returns 402 with spend_cap. Per-user caps return errors when limits block the request.

Related

• Merge Gate API — PR review queue, structured findings, webhooks
• Connections & Platform API — crt_ keys, scopes, MCP, REST v1
• Coding Agent API — full sandbox agent runs
• Models — review catalog including Hy3 and Nemotron
• Ship log — operator release notes (v6.0 Platform API and lifecycle webhooks; v5.2 Inference API)