# Inference APIs

> OpenAI-compatible hosted inference for open-weight models (DeepSeek V4, GLM 5.3, Kimi K2.7, MiniMax M3, GPT-OSS 120B, Qwen3-VL), plus embeddings, rerank, text-to-speech and transcription, on one API key. Models run on GPUs in the United States with zero data retention. Prepaid, billed per token, no subscription, no per-minute or per-day quotas.

Base URL: https://api.inferenceapis.com/v1
Auth: Authorization: Bearer <api_key>   (x-api-key also accepted)
Endpoints: POST /chat/completions, POST /responses (OpenAI Responses API, Codex CLI works incl. MCP tools and sub-agents), POST /messages (Anthropic Messages API), POST /embeddings, POST /rerank, POST /audio/speech, POST /audio/transcriptions, GET /models, POST /signup
MCP server (Streamable HTTP; signup, list_models, account, usage, create_api_key, test_completion, docs): https://api.inferenceapis.com/mcp · setup skill for agents: https://inferenceapis.com/skill/SKILL.md
Machine-readable: https://inferenceapis.com/openapi.json · errors: https://inferenceapis.com/docs/errors · full text: https://inferenceapis.com/llms-full.txt

## Get an account and a key without a browser

POST https://api.inferenceapis.com/v1/signup  {"email": "person@example.com", "name": "optional", "password": "optional, 8+ chars"}
-> 201 {"api_key": "ia-...", "base_url": ..., "balance": 0, "starter_credit": {"amount": 1.00, "status": "pending_email_confirmation"}, "top_up_url": ...}
The key works at once. The $1 starter credit is added when the person opens the confirmation email; until then requests return 402 insufficient_balance. Adding funds is done by a person at https://inferenceapis.com/billing. One account per email; throwaway domains are refused; 5 signups per IP per day. Accounts without a password sign in with "Email me a sign-in link" on https://inferenceapis.com/login.

## Models (live now; GET /v1/models returns the same with prices, context and capabilities)

### Chat (also vision-capable where noted)
- openai/gpt-oss-120b: $0.2 in / $0.8 out per 1M tokens · context 131K · reasoning (cannot be switched off; reasoning_effort "none" maps to low) · https://inferenceapis.com/models/openai/gpt-oss-120b
- deepseek-ai/DeepSeek-V4-Flash: $0.19 in / $0.38 out / $0.04 cached in per 1M tokens · context 1M · reasoning (switch off with reasoning_effort: "none") · https://inferenceapis.com/models/deepseek-ai/DeepSeek-V4-Flash
- deepseek-ai/DeepSeek-V4.1-Flash: $0.4 in / $1.6 out / $0.01 cached in per 1M tokens · context 1M · reasoning (switch off with reasoning_effort: "none") · https://inferenceapis.com/models/deepseek-ai/DeepSeek-V4.1-Flash
- zai-org/GLM-5.3-Flash: $0.2 in / $0.66 out / $0.04 cached in per 1M tokens · context 1M · reasoning (switch off with reasoning_effort: "none") · https://inferenceapis.com/models/zai-org/GLM-5.3-Flash
- meta-llama/Llama-3.3-70B-Instruct-Turbo: $1.35 in / $1.35 out per 1M tokens · context 131K · https://inferenceapis.com/models/meta-llama/Llama-3.3-70B-Instruct-Turbo
- zai-org/GLM-5.3: $1.82 in / $5.72 out / $0.34 cached in per 1M tokens · context 1M · reasoning (switch off with reasoning_effort: "none") · https://inferenceapis.com/models/zai-org/GLM-5.3
- deepseek-ai/DeepSeek-V4-Pro: $1.72 in / $5.15 out / $0.17 cached in per 1M tokens · context 1M · reasoning (switch off with reasoning_effort: "none") · https://inferenceapis.com/models/deepseek-ai/DeepSeek-V4-Pro
- MiniMaxAI/MiniMax-M3: $0.39 in / $1.56 out / $0.08 cached in per 1M tokens · context 512K · reasoning (switch off with reasoning_effort: "none") · https://inferenceapis.com/models/MiniMaxAI/MiniMax-M3
- moonshotai/Kimi-K2.7-Code: $0.89 in / $4.42 out / $0.18 cached in per 1M tokens · context 262K · reasoning (switch off with reasoning_effort: "none") · https://inferenceapis.com/models/moonshotai/Kimi-K2.7-Code
- Qwen/Qwen3-VL-235B-A22B-Instruct: $0.26 in / $1.15 out per 1M tokens · context 262K · vision · https://inferenceapis.com/models/Qwen/Qwen3-VL-235B-A22B-Instruct

### Embeddings
- BAAI/bge-m3: $0.02 per 1M tokens · context 8K · https://inferenceapis.com/models/BAAI/bge-m3
- Qwen/Qwen3-Embedding-8B: $0.02 per 1M tokens · context 32K · https://inferenceapis.com/models/Qwen/Qwen3-Embedding-8B

### Rerank
- Qwen/Qwen3-Reranker-8B: $0.07 per 1M tokens · context 32K · https://inferenceapis.com/models/Qwen/Qwen3-Reranker-8B

### Text to speech
- hexgrad/Kokoro-82M: $5.2 per 1M characters · https://inferenceapis.com/models/hexgrad/Kokoro-82M
- canopylabs/orpheus-3b-0.1-ft: $19.5 per 1M characters · https://inferenceapis.com/models/canopylabs/orpheus-3b-0.1-ft

### Speech to text
- openai/whisper-large-v3: $0.002 per audio minute · https://inferenceapis.com/models/openai/whisper-large-v3
- nvidia/parakeet-tdt-0.6b-v3: $0.002 per audio minute · https://inferenceapis.com/models/nvidia/parakeet-tdt-0.6b-v3
- mistralai/Voxtral-Small-24B-2507: $0.004 per audio minute · no word timestamps · https://inferenceapis.com/models/mistralai/Voxtral-Small-24B-2507
- mistralai/Voxtral-Mini-3B-2507: $0.0013 per audio minute · no word timestamps · https://inferenceapis.com/models/mistralai/Voxtral-Mini-3B-2507

Aliases: common ids from other providers are accepted (deepseek-chat, deepseek-reasoner, whisper-1, tts-1, llama-3.3-70b-versatile, gpt-oss-120b). GET /v1/models lists them under "aliases".

## Configure a tool

Every guide has a tested config block: https://inferenceapis.com/integrations/
- Claude Code: https://inferenceapis.com/integrations/claude-code — zai-org/GLM-5.3 was the fastest large model in our run and deepseek-ai/DeepSeek-V4.1-Flash the best value. moonshotai/Kimi-K2.7-Code and deepseek-ai/DeepSeek-V4-Pro are the other serious choices for agent work.
- Codex CLI: https://inferenceapis.com/integrations/codex-cli — For real coding work use a large model: zai-org/GLM-5.3, deepseek-ai/DeepSeek-V4-Pro or moonshotai/Kimi-K2.7-Code. deepseek-ai/DeepSeek-V4-Flash is the cheap option for small edits.
- opencode: https://inferenceapis.com/integrations/opencode — A large model for agent work: zai-org/GLM-5.3, deepseek-ai/DeepSeek-V4-Pro or moonshotai/Kimi-K2.7-Code. DeepSeek and GLM here run in the United States, so opencode's "models hosted in China" switch is not needed. If you came here from OpenCode Go's 403 RegionError, that error is explained here.
- VS Code Copilot Chat (bring your own key): https://inferenceapis.com/integrations/vscode-copilot-byok — Agent mode needs reliable tool calling: zai-org/GLM-5.3 or deepseek-ai/DeepSeek-V4-Pro. For quick chat, deepseek-ai/DeepSeek-V4.1-Flash.
- OpenAI Python SDK: https://inferenceapis.com/integrations/openai-python — Start with openai/gpt-oss-120b; switch to deepseek-ai/DeepSeek-V4-Flash when cost matters more than latency.
- OpenAI Node.js SDK: https://inferenceapis.com/integrations/openai-node — Start with openai/gpt-oss-120b; for a streamed UI, deepseek-ai/DeepSeek-V4.1-Flash or meta-llama/Llama-3.3-70B-Instruct-Turbo reach the first answer word soonest; reasoning models think first unless you turn that down.
- LiteLLM: https://inferenceapis.com/integrations/litellm — As a fallback for Groq, use the same id on both sides: openai/gpt-oss-120b.
- LangChain: https://inferenceapis.com/integrations/langchain — For agents and LangGraph, openai/gpt-oss-120b; for high-volume chains, deepseek-ai/DeepSeek-V4-Flash.
- LlamaIndex: https://inferenceapis.com/integrations/llamaindex — For RAG over long documents, a 1M-context model such as deepseek-ai/DeepSeek-V4-Flash (set context_window=1000000).
- Vercel AI SDK: https://inferenceapis.com/integrations/vercel-ai-sdk — For streamed UI responses, deepseek-ai/DeepSeek-V4.1-Flash, with reasoning_effort or the reasoning switch turned down so the answer starts sooner.
- Continue (VS Code / JetBrains): https://inferenceapis.com/integrations/continue — Chat and edit: openai/gpt-oss-120b. Autocomplete needs speed: meta-llama/Llama-3.3-70B-Instruct-Turbo does not think before answering.
- Cline (VS Code): https://inferenceapis.com/integrations/cline — openai/gpt-oss-120b. Cline needs reliable tool calling and long system prompts.
- Open WebUI: https://inferenceapis.com/integrations/open-webui — Chat: any model from the list. Speech-to-text: check accuracy for your language first.
- n8n: https://inferenceapis.com/integrations/n8n — For classification and extraction steps, deepseek-ai/DeepSeek-V4-Flash with reasoning disabled keeps cost and latency down.
- Aider: https://inferenceapis.com/integrations/aider — openai/openai/gpt-oss-120b for edits; aider sends whole files, so watch token usage.

## Facts an agent should know

- Stateless Responses API: send the whole conversation in input each turn; previous_response_id and store are not supported. Codex CLI's namespace-wrapped MCP tools and agent_message hand-offs are translated, so MCP servers and sub-agents work on every chat model.
- Cached input: repeated prompt prefixes bill at the cached-input rate automatically; nothing to configure.
- Thinking: reasoning_effort "none" switches reasoning off on any reasoning model (Responses/Codex: effort "minimal").
- Limits: no rate limits on your requests or tokens; a 429 means the backend is momentarily saturated, retry after Retry-After. Speech input 20,000 chars; audio upload 100 MB; embeddings 1,024 texts; rerank 1,000 documents.
- Hosting: Together AI (United States); DeepInfra (United States); Our own GPU server (United States). Zero data retention is on with the GPU providers; the gateway stores token counts for billing, never content. https://inferenceapis.com/trust
- Status: https://inferenceapis.com/status · changelog: https://inferenceapis.com/changelog · contact: support@inferenceapis.com

## Errors

OpenAI envelope; every error carries docs_url -> https://inferenceapis.com/docs/errors#<code>.
- 401 missing_api_key: No Bearer token was sent. Retry: No.
- 401 invalid_api_key: A token was sent but it does not match any key. Retry: No.
- 402 insufficient_balance: The prepaid balance is zero or below the cost of the request. Retry: After topping up.
- 404 model_not_found: The model id is not in the catalog. Ids are case-insensitive; common aliases such as deepseek-chat or whisper-1 are accepted. Retry: No.
- 400 wrong_endpoint: The model exists but belongs to another endpoint, for example a speech model sent to /v1/chat/completions. Retry: No.
- 503 model_unavailable: The model is listed but is not being served right now, either disabled by us or refused by the backend. Retry: Yes, or use another model.
- 400 unsupported_parameter: A Responses API field that needs server-side state was sent: previous_response_id, conversation or background. Retry: No.
- 413 input_too_long: Speech input over 20,000 characters, or an embedding or rerank request over the per-request limits (1,024 texts; 1,000 documents). Retry: No.
- 413 file_too_large: An audio upload over 100 MB. Retry: No.
- 400 timestamps_unsupported: Word timestamps, SRT or VTT were requested from a transcription model that does not produce them. Retry: No.
- 429 upstream_rate_limited: The inference backend is saturated for that model. This is not a quota: your requests and tokens are not metered per minute or per day. Retry: Yes, after Retry-After.
- 502 upstream_error: The inference backend failed or timed out. The message carries what it said. Retry: Yes, with backoff.
- 502 upstream_auth: The backend refused our credentials. This is our problem, not yours, and it is logged and alerted on our side. Retry: Yes, later.
- 404 unknown_endpoint: The path is not one we serve. Files, Batch, fine-tuning, moderation, image generation and audio translation are not offered. Retry: No.
- 400 invalid_request: The body is not valid JSON, a required field is missing, or a value is out of range. Retry: No.
- 400 signup_rejected: POST /v1/signup refused the email address: malformed, a throwaway domain, already registered, or too many signups from this address today. Retry: No.

Generated 2026-09-18 from the live catalog. Prices are USD.