Meta Muse Spark 1.3: Verified Context, Pricing, Features, and API Notes

Editorial update — September 27, 2026: An earlier version of this article included unsupported benchmark results, a two-million-token context claim, made-up architecture details, outdated prices, and an unverified open-weights date. Those claims have been removed. This replacement checks Meta's published developer documentation and is not an independent benchmark of Muse Spark.

Meta lists Muse Spark 1.3 (muse-spark-1.3) as its current Muse Spark version for agentic and coding work. Meta describes it as tuned for multi-step tool use, browser tasks, and long-horizon workflows, with coding improvements over version 1.2. These are the provider's product descriptions, not results measured by Accelerated Logic. The official model guide is the source to check for current model IDs and limits.

The model guide lists a context window of 1,048,576 tokens. It supports text, images, video, audio, and PDF as input, with text output. Meta marks audio understanding in version 1.3 as not fully supported and warns that quality may be degraded for audio requests; for audio understanding it points developers to version 1.2 or its dedicated transcription product. A maximum context size describes the service limit, not a guarantee that every detail inside a very long prompt will be retrieved correctly.

Meta documents chat, image and video understanding, tool calling, structured output, and search grounding for Muse Spark. The current API documentation also describes a Responses API and OpenAI SDK-compatible setup. Check the relevant endpoint documentation before choosing a client library: compatibility does not mean that every SDK feature or request field is identical.

Meta's pricing and rate-limit page lists these rates per million tokens for Muse Spark 1.3:

Standard tier: $0.15 per million cached input tokens, $1.25 per million input tokens, and $4.25 per million output tokens. Meta says Standard prompts and completions are not used to train Meta models.

Contributor tier: $0.002 per million cached input tokens, $0.10 per million input tokens, and $0.20 per million output tokens. This discounted tier grants Meta permission to use prompts and completions to train future models.

The low Contributor rates have a data-use tradeoff. Review the current tier terms before sending sensitive or confidential material. Meta separately lists search grounding at $2.50 per 1,000 search queries, in addition to the token cost. Prices can change, so use Meta's live pricing page rather than carrying these figures into a budget without checking them again.

For a simple Standard-tier estimate, 100,000 input tokens at $1.25 per million cost $0.125; 10,000 output tokens at $4.25 per million cost $0.0425. The combined token charge is $0.1675 before any separately priced search requests. Real usage varies with prompt length, output size, caching, and enabled tools.

Meta's API overview lists https://api.meta.ai/v1 as the base URL and muse-spark-1.3 as the model identifier. The quickstart walks through authentication and a first request. Use an API key issued by Meta and follow its current request-format examples; do not paste a real credential into source code, a public prompt, or a shared screenshot.

The built-in model catalog in this repository lists Muse Spark 1.1 and 1.2, but not 1.3. This article describes Meta's API; it does not claim that Muse Spark 1.3 is selectable from Accelerated Logic's default model picker. Check the current workspace settings and model catalog for supported configuration options.

No shared benchmark run or hardware test is reported here. The earlier percentages, speed measurements, competitor rankings, price comparison, architectural explanation, and promised weight-release timeline were not supported by the cited material and have been removed. Meta's capability descriptions and published limits should be labeled as provider-reported; an independent comparison would require a documented model version, prompt set, harness, run settings, and results.

Muse Spark 1.3 may be worth evaluating when an application needs Meta's hosted multimodal API or multi-step tool workflows. The practical decision still depends on your task, request format, latency, pricing tier, data-use terms, and whether the model works with the tools you need. Start from the primary documentation, test with representative prompts, and keep your own measurements separate from vendor claims.

- Meta Model API: models and context limits - Meta Model API: pricing and rate limits - Meta Model API: overview and base URL - Meta Model API: quickstart