Meta Llama API

Meta's API, once called the Llama API, is now the Meta Model API and is built around the Muse models. On 2026-10-31 Muse Wonder costs $4 in and $20 out per million tokens, and Muse Spark costs $1.25 and $4.25. A Contributor mode cuts Spark to $0.10 and $0.20 if you let Meta train on your content.

Facts at a glance

Prices checked 2026-10-31. If you are looking for the old Llama API, the slug on this site is kept for that reason. The Llama weights themselves are covered on Llama models.

Models and prices

Meta bills text models per token, images per image, transcription by the hour of audio, and segmentation per image or video frame, with no minimums or upfront commitments.

Three notes from the pricing page. There is no long-context premium, so you pay the same rate whether the 1M window is empty or almost full. Cached input costs less, and the response reports cached_tokens so you can check the hit rate. Web search grounding costs $2.50 per 1,000 queries, added to the token cost. The Muse Wonder price matches Claude Opus 5.5's input rate closely, at $4 against $4, while output is $20 on both, which makes Wonder a direct price rival to Anthropic's Opus line. That is a price comparison, not a quality one, and we have no test results to cite.

Standard mode and Contributor mode

This is the most unusual feature of Meta's API, and it decides your data position.

Meta states it plainly: "In Standard mode, you pay standard prices and your content is not used to train Meta models." Contributor mode is described as a way to lower the barrier for prototyping, testing integrations and scaling experiments "where it's acceptable for Meta to use your content". In Muse Code, Meta's coding agent, subscriptions default to Contributor mode, and you can turn it off in settings.

The discount is large. Muse Spark in Contributor mode costs about 8 percent of the standard input price and under 5 percent of the standard output price. That makes it very attractive for public code and open-source work, and unsuitable for anything confidential. The risk to manage is the default in Muse Code. A developer who installs it and signs in with a subscription may be in Contributor mode without noticing, so check the setting before pointing it at a private repository.

Zero data retention is mentioned on the pricing page, for Muse Voice Transcribe at parity with Standard mode. We did not find a general retention period, or a statement of how long Standard-mode prompts are kept.

Rate limits

Limits apply per team, not per key: every key's requests, tokens and streams count towards the same limits. Each model has its own limits, so traffic to one does not use another's.

These are generous defaults for a new account. The one low number is 150 requests a minute for the Contributor variants, which matters if you run many parallel agent calls on the discounted model.

Which agents can use it, and how

Meta's docs cover coding agents in unusual detail. OpenAI-compatible agents connect to https://api.meta.ai/v1 with the Responses or Chat Completions API. Anthropic-format agents such as Claude Code connect through the Messages API at https://api.meta.ai. For Claude Code, Meta documents these variables:

export MODEL_API_KEY="<your-model-api-key>"
export ANTHROPIC_BASE_URL="https://api.meta.ai"
export ANTHROPIC_AUTH_TOKEN="$MODEL_API_KEY"
export ANTHROPIC_MODEL="muse-wonder-1.0"
export ANTHROPIC_DEFAULT_OPUS_MODEL="muse-wonder-1.0"
export ANTHROPIC_DEFAULT_SONNET_MODEL="muse-spark-1.3"
claude

So the Opus slot runs Wonder and the Sonnet slot runs Spark. See Claude Code. For OpenCode, Meta gives a provider block that uses the @ai-sdk/openai adapter against the Responses API with the base URL above, model muse-spark-1.3, reasoning enabled and a 1,048,576 context with 131,072 maximum output. It also documents Codex with the Responses API. See OpenCode and Codex CLI. Chat Completions is the universal fallback, because every OpenAI-compatible agent supports it.

Meta's own agent is Muse Code, a terminal and CI agent built on Muse Spark with approvals and an OS sandbox on from the first run. It installs with a shell script and starts with the muse command, and on first run you pick a browser sign-in or paste an API key. Muse Code subscriptions can include Muse Wonder Full and Contributor mode.

Open weights

Meta also publishes weights. Muse Spark Open is the open-weight release of Muse Spark, with 816 billion parameters and 42 billion active, under the Muse Model License, a gated release with manual approval on Hugging Face. Its licence asks every product that uses the model or a derivative to display "Built with Muse Spark Open". Muse Glimmer is described as open-weight and built for local and edge deployment, under Apache 2.0. Third-party hosts serve Muse Spark Open too: Together AI and Fireworks AI both list it at $1.25 in and $4.25 out, the same as Meta's Standard price for Spark. The older Llama 4 weights are from April 2025.

Wonder or Spark

Meta's docs steer you between the two models, and the choice drives most of your bill.

Wonder is about three times the input price and nearly five times the output price of Spark. Both take the same requests, so Meta says any example that uses Spark also runs on Wonder by changing the model id to muse-wonder-1.0. That makes a routing design easy: send ordinary steps to Spark and escalate hard ones to Wonder. The price gap is the argument for doing so, and your own task test is the argument against assuming it works.

What changed from the Llama API

If you used Meta's earlier Llama API, expect four changes. The brand and docs now lead with Muse, and Llama sits under a documentation menu. The base URL is https://api.meta.ai/v1. The API exposes three surfaces, Responses, Chat Completions and an Anthropic-style Messages endpoint, instead of one. And the data position is more explicit, with Standard and Contributor modes. If you built on a Llama model through Meta, check which Llama models are still served and plan a move to a Muse model or to self-hosted weights, since Meta's focus moved. We did not confirm which Llama models the hosted API still serves.

Moving to the Meta Model API

1. Get a key from the Model API dashboard and export it as MODEL_API_KEY.
2. Pick the surface your agent uses. Chat Completions is the safe default, Responses adds server-managed conversation state and video input, and Messages serves Anthropic-format clients.
3. Pick the mode. Standard for private work, Contributor only for data you are happy to share.
4. Use the cache. Keep stable prefixes at the front and watch cached_tokens.
5. Test tools. The docs say tool calling, streaming and multi-step reasoning work through the main agents, and your own schemas still need a run.
6. Check team limits. One team shares one limit across keys, so a CI job and a developer laptop compete for the same pool.

Moving away is mostly a base URL change, and the Messages and Chat Completions surfaces make that easy. The one part to plan for is any state you keep in server-managed Responses conversations, which you should export before leaving.

Alternatives

A worked cost example

Take 10 million input tokens and 2 million output tokens. The arithmetic is ours, from the Meta price list, with no caching.

The Contributor row costs less than a fiftieth of Wonder on this mix. It is the cost of letting Meta train on the data. If 8 million of the 10 million input tokens are cache hits, Spark in Standard mode costs $1.20 for the cached part plus $2.50 for the other 2 million, so $3.70 of input and $12.20 in total.

Who should pick Meta's API

A pre-launch checklist

1. Choose Standard mode for anything private, and check the Muse Code setting.
2. Use the -contributor model ids only for data you are happy to share.
3. Check cached_tokens in responses to confirm the cache works.
4. Budget web search at $2.50 per 1,000 queries.
5. Remember limits are per team, not per key.
6. Get a retention statement and region terms in writing.
7. Show the "Built with Muse Spark Open" notice if you ship the open weights.

Our take

Meta has moved from a Llama API to a full model platform, and the documents are better than most: a clear price list, honest limits, a long-context price with no premium, and setup guides for the main coding agents. The Contributor mode is the notable idea. It puts a price on your data and lets you choose. The cost of the clarity is that the data terms outside the training question, such as retention and region, were not on the pages we read, and that the default in Muse Code leans toward sharing. For public work it is a bargain. For private work use Standard mode and ask for the retention terms.

We have not run tests for this page. It reflects Meta's published documentation on 2026-10-31. Meta does not pay for placement here.

Related pages

Llama models, LLM API prices, open source LLM, Claude Code, OpenCode.