# Boundless

> An inference platform for open-weight models, sold on prepaid credits. One service speaks three wire formats — OpenAI chat completions, the Codex-compatible Responses subset, and the Anthropic Messages API — over the same models on the same rate card. Every response carries exact token counts, so any charge can be recomputed from the published prices.

Operated by Boundless Networks, Inc. The service is invite-only while the first accounts are onboarded, and the model lineup, prices and base URL may change during that period. Support is a human at support@boundless.network.

## When to use this API

Reach for Boundless when:

- You already call an OpenAI or Anthropic chat API and want the same code to run open-weight models at a lower price — the wire formats are the ones your SDK already sends, so the change is a base URL and a key.
- You need the exact cost of each request, not a monthly total. Every response carries its token counts, and those counts against the published rate card reproduce the charge to the sub-cent.
- You need a hard spending ceiling. Credits are prepaid, requests stop at zero with no overage, and a single API key can carry its own budget below the team's.
- You are running a coding agent or harness — Codex CLI, Claude Code, omp, Hermes, or opencode — and want to point it at cheaper models without changing the harness.
- You need long context: the served models take 262,144 tokens.
- You want tool use and reasoning models on open weights, with reasoning effort you control per request.

Do not reach for it when:

- You need provider-native Responses features such as OpenAI-hosted tools or remote conversation storage. `/v1/responses` serves the Codex-compatible subset, not those hosted features.
- You need embeddings, legacy completions (`/v1/completions`), fine-tuning, or image or audio generation. None of these are served. Batch inference with `1h` and `24h` completion windows launches in the coming weeks; the discounted rates are on /docs/pricing.
- You need a specific proprietary model. Only open-weight models are served; Anthropic's `claude-*` and OpenAI's model names are not available here under any alias.
- You need instant self-service activation today. Sign-up and key creation are self-serve, but the service is invite-only and a human approves each new account while the first accounts are onboarded. A capped $0.01 agent key is mintable instantly for smoke tests; a human account is only needed for real spend.

### How to call it

Point an OpenAI SDK or Codex custom provider at `https://api.inference.boundless.network/v1`, or an Anthropic SDK at `https://api.inference.boundless.network` — the Anthropic SDKs append `/v1/messages` themselves. Codex uses `wire_api = "responses"`. Authenticate with `Authorization: Bearer $BOUNDLESS_API_KEY`, or with `x-api-key` if that is what your client already sends. Name a model from the catalog below.

```bash
curl https://api.inference.boundless.network/v1/chat/completions \
  -H "Authorization: Bearer $BOUNDLESS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "deepseek-v4.1-flash", "messages": [{"role": "user", "content": "Say hello"}]}'
```

Keys for real workloads are created in the console at https://inference.boundless.network/api-keys once an account is active. An anonymous smoke-test key can be minted instantly by `POST https://inference.boundless.network/api/agent-key` — no body, no credential — or by the `create_agent_key` MCP tool: $0.01 of credit, expires in 24 hours, limited to 6 requests and 30,000 tokens per minute. The response carries the key and the base URL to point an SDK at.

Every page listed below also serves its markdown source directly — send `Accept: text/markdown` and you get the source instead of the HTML, with no markup to strip. The whole documentation set concatenated is at https://inference.boundless.network/llms-full.txt.

### Models

Each identifier links to the model's own page, which carries its live rates.

- [`dsv4`](https://inference.boundless.network/models/dsv4): Cheap 1M-context text agents and long-prefix cache work. 1,048,576-token context, supports text, reasoning, tool-use. Not for: Image, audio, or video — this identifier is text only. Choose DeepSeek V4.1 Flash when you need native images, or V4 Pro when a main coding agent needs more depth.
- [`deepseek-v4.1-flash`](https://inference.boundless.network/models/deepseek-v4.1-flash): Fast, low-cost agentic coding and visual analysis. 1,000,000-token context, supports text, reasoning, tool-use, image input, vision. Not for: Audio or video input. Not V4 Pro — keep Pro on the main agent and this Flash on sub-agents. Not V4 Flash 0731 (`dsv4`) when you want the cheaper text-only path.
- [`glm-5.2`](https://inference.boundless.network/models/glm-5.2): Hours-long agentic engineering on text. 262,144-token context, supports text, reasoning, tool-use. Not for: Image, video, or audio. Prefer GLM-5.3 Flash for native vision, video, and cheaper 1M-context serving.
- [`nemotron3-super`](https://inference.boundless.network/models/nemotron3-super): Multi-agent coordination and planning on an open 120B hybrid MoE. 262,144-token context, supports text, reasoning, tool-use. Not for: Image, audio, or video. Not the high-volume routine worker — pair it with a lighter model for that. Context served here is 262,144 tokens, not a million.
- [`qwen3.6`](https://inference.boundless.network/models/qwen3.6): Lightweight multimodal work at the low end of the rate card. 262,144-token context, supports text, reasoning, tool-use, image input, vision. Not for: Audio or video. Not Qwen3.8 Flash when you need a 1M-token window, and not a frontier-scale coder.
- [`qwen3.8-flash`](https://inference.boundless.network/models/qwen3.8-flash): Fast 1M-context visual analysis on compact Qwen3.8-27B. 1,000,000-token context, supports text, reasoning, tool-use, image input, vision. Not for: Audio input. Not the pick for frontier-scale coding; this is the compact visual Qwen, not Kimi K3 or a DeepSeek Pro main agent.
- [`kimi-k3`](https://inference.boundless.network/models/kimi-k3): Agentic coding, tool use, and long-horizon reasoning. 1,000,000-token context, supports text, reasoning, tool-use, image input, vision. Not for: Audio or video. Not the cheapest Flash-class option for short chat or classification.
- [`glm-5.3-flash`](https://inference.boundless.network/models/glm-5.3-flash): Cheap 1M-context multimodal work at speed. 1,048,576-token context, supports text, reasoning, tool-use, image input, vision, video input. Not for: Audio input. Not GLM-5.2 — this is the cheaper, natively visual 5.3 Flash sibling, not the text-only 5.2 flagship.
- [`deepseek-v4-pro-0813`](https://inference.boundless.network/models/deepseek-v4-pro-0813): Open-weight DeepSeek V4 Pro as the main agent in a Flash-and-Pro stack. 1,048,576-token context, supports text, reasoning, tool-use. Not for: Image, audio, or video. Not the latency or cost pick — use V4 Flash (`dsv4`) or V4.1 Flash for sub-agents and high-volume steps.
- [`mimo-v2.5`](https://inference.boundless.network/models/mimo-v2.5): Omnimodal understanding across text, image, audio and video. 1,000,000-token context, supports text, reasoning, tool-use, image input, vision, audio input, video input. Not for: Text-only, cost-sensitive work — a Flash-class text model is cheaper when you do not need audio or video.
- [`minimax-m3`](https://inference.boundless.network/models/minimax-m3): Paid MiniMax M3 for long-horizon agents over text, image and video. 1,048,576-token context, supports text, reasoning, tool-use, image input, vision, video input. Not for: Audio input. Use `minimax-m3-free` to evaluate the same weights without a rate card; this identifier is the billed lineup entry.
- [`minimax-m3-free`](https://inference.boundless.network/models/minimax-m3-free): Unpaid evaluation of MiniMax M3 on text, image and video. 1,048,576-token context, supports text, reasoning, tool-use, image input, vision, video input. Not for: Billed production traffic — switch to `minimax-m3`. Same weights; this identifier is the free-tier alias. No audio input.
- [`gpt-5.6-luna`](https://inference.boundless.network/models/gpt-5.6-luna): High-volume GPT-5.6 chat, classification, and light agents. 262,144-token context, supports text, reasoning, tool-use. Not for: Image, audio, or video. Not GPT-5.6 Sol — use Sol for multi-step coding and long-horizon reasoning. Call it as `gpt-5.6-luna`.
- [`gpt-5.6-sol`](https://inference.boundless.network/models/gpt-5.6-sol): GPT-5.6 flagship for CLI-heavy coding and long-horizon problem solving. 262,144-token context, supports text, reasoning, tool-use. Not for: Image, audio, or video. Not the cheap high-volume pick — use Luna for chat and classification. Call it as `gpt-5.6-sol`.
- [`claude-opus-5`](https://inference.boundless.network/models/claude-opus-5): Anthropic Opus 5 for demanding coding, review, and long-horizon agents. 262,144-token context, supports text, reasoning, tool-use. Not for: Image, audio, or video input on this route — text only. Not Fable 5.1 when you want the cheaper, more concise coding and knowledge-work upgrade. Call it as `claude-opus-5`.
- [`claude-fable-5.1`](https://inference.boundless.network/models/claude-fable-5.1): Claude Fable 5.1 for agentic coding, front-end generation, and knowledge work. 262,144-token context, supports text, reasoning, tool-use. Not for: Image, audio, or video input on this route — text only. Not Opus 5 when the job is the heaviest reasoning or parallel-subagent coordination. Call it as `claude-fable-5.1`.

The whole lineup, with rates, is [Models](https://inference.boundless.network/models); the rate card on its own is [Pricing](https://inference.boundless.network/docs/pricing). Both serve their tables as markdown to a client that asks for `text/markdown`. Prices are per million tokens and are deliberately not restated here: a price copied into a second place is a price that can be wrong.

## Documentation

- [Boundless Platform](https://inference.boundless.network/docs): Boundless Platform: authentication, endpoints, examples, limits and agent integrations
- [Quickstart](https://inference.boundless.network/docs/quickstart): From API key to first streamed token
- [Coding Agents & Harnesses](https://inference.boundless.network/docs/coding-agents-and-harnesses): Point Claude Code, omp, Hermes and other harnesses at this API
- [API Compatibility](https://inference.boundless.network/docs/api-compatibility): The endpoints we serve — OpenAI chat completions and the Anthropic Messages API — and the ones we do not
- [Credits & Billing](https://inference.boundless.network/docs/billing): Buying credits, automatic top-ups, per-key budgets, and the hard stop at zero
- [Pricing](https://inference.boundless.network/docs/pricing): The models we serve and the live rate card
- [Rate Limits](https://inference.boundless.network/docs/rate-limits): The throughput cap on your team, and how to raise it
- [Teams](https://inference.boundless.network/docs/teams): Invite colleagues, share credits, and manage roles
- [FAQ](https://inference.boundless.network/docs/faq): When to use Boundless, when not to, wire formats, credits, and what is not served

## Machine-readable

- [llms.txt](https://inference.boundless.network/llms.txt): This file
- [llms-full.txt](https://inference.boundless.network/llms-full.txt): Every public page's markdown in one file — one request instead of one per page
- [docs/llms.txt](https://inference.boundless.network/docs/llms.txt): The documentation section on its own, for an agent that only needs to find a page
- [index.md](https://inference.boundless.network/index.md): This origin's home page as markdown
- [pricing.md](https://inference.boundless.network/pricing.md): The model catalog and the live rate card as markdown, at the path a machine looks for prices
- [OpenAPI specification](https://inference.boundless.network/openapi.json): OpenAPI 3.1 description of the inference API
- [auth.md](https://inference.boundless.network/auth.md): How an agent obtains a credential here, in the structure the auth.md convention prescribes
- [Agent key mint](https://inference.boundless.network/api/agent-key): GET describes the offer; POST mints an anonymous trial key with no credential — $0.01 of credit, 24h expiry
- [Sitemap](https://inference.boundless.network/sitemap.xml): Every indexable URL on this origin
- [Schema map](https://inference.boundless.network/schemamap.xml): The structured-data feeds this origin publishes
- [MCP manifest](https://inference.boundless.network/.well-known/mcp.json): MCP server manifest, Streamable HTTP transport
- [MCP Server Card](https://inference.boundless.network/.well-known/mcp/server-card.json): Branded MCP Server Card for the product inference endpoint
- [MCP endpoint](https://inference.boundless.network/mcp): Product MCP endpoint — the live catalog, the rate card and cost estimates, over Streamable HTTP
- [Documentation MCP endpoint](https://inference.boundless.network/docs/mcp): Documentation MCP endpoint — search, read and list the published pages, over Streamable HTTP
- [AI catalog](https://inference.boundless.network/.well-known/ai-catalog.json): Agentic Resource Discovery catalog — every agentic resource this origin publishes
- [API catalog](https://inference.boundless.network/.well-known/api-catalog): RFC 9727 API catalog, as an RFC 9264 linkset
- [Agent card](https://inference.boundless.network/.well-known/agent-card.json): A2A agent card, v1.0 — what this service can be asked and where to ask it
- [A2A endpoint](https://inference.boundless.network/a2a/v1): A2A endpoint, JSON-RPC binding
- [Documentation MCP manifest](https://inference.boundless.network/docs/mcp.json): Server card for the documentation MCP endpoint
- [Agent skills](https://inference.boundless.network/.well-known/agent-skills/index.json): Agent Skills index — self-contained skills for calling this service
- [Ask](https://inference.boundless.network/ask): NLWeb natural-language query endpoint, with SSE streaming
- [Model catalog](https://inference.boundless.network/api/catalog): The model catalog and the live rate card as JSON, with no credential
- [Health](https://inference.boundless.network/api/health): Console liveness probe

Two MCP servers are available over Streamable HTTP, neither needing authentication; every tool is read-only except `create_agent_key`, which mints the capped trial key described above. https://inference.boundless.network/mcp is the product: the live model catalog, the rate card and cost estimates, described at https://inference.boundless.network/.well-known/mcp.json. https://inference.boundless.network/docs/mcp is the documentation: search it, read a page in full, or list what exists.

## Company

- [Privacy](https://inference.boundless.network/privacy): What this service records, what it never records, and the governing policy
- [Support](https://inference.boundless.network/support): How to reach us, and what to send with a question

## Optional

- [Terms of Service](https://inference.boundless.network/terms): the Boundless Platform Services Terms, in full
- [Boundless Networks](https://boundless.network): the company behind this service