Muse Spark 1.3: Pricing, Contributor Endpoint and Efficiency Claims

What Muse Spark 1.3 is

Muse Spark 1.3 was released on 2 September 2026 as Meta's model for long-running agentic, multi-agent and coding work, with a 1 million token context window and vision input. Meta frame the release around practical usability rather than raw capability: it was trained on more long-horizon coding tasks, takes fewer turns where they are not needed, is less verbose, and has a cleaner coding style. It is available with max reasoning through Muse Code and the Meta Model API. The weights are closed, though Meta list an open-weights Muse Spark release on their roadmap.

Efficiency claim

Meta compared 1.3 against Muse Spark 1.2: in comparisons by their own engineers, 1.3 used roughly 20% fewer tool calls and roughly 25% fewer tokens for the same work, and was significantly faster. That is a vendor figure from an internal comparison.

Two endpoints

Meta do not publish per-token rates on the launch post. The $1.25 per million input and $4.25 output recorded here are for the standard endpoint as reported by Artificial Analysis rather than read from Meta's own page. There is also a separate Contributor endpoint, reported at $0.10 per million input tokens and $0.20 output. That is roughly a twelfth of the standard input price and a twentieth of the output, and the consideration is explicit: your data is used to improve Meta's products.

LMArena

On LMArena's text leaderboard Muse Spark 1.3 has 1495 Elo (rank 11, max effort); by category, 1534 Elo in Coding (rank 17, max effort), 1508 Elo in Math (rank 11, max effort), 1522 Elo in Expert (rank 19, max effort) and 1464 Elo in Creative Writing (rank 25, max effort).

What Meta changed for long tasks

Meta state Muse Spark 1.3 was built to sustain longer-horizon work in a single long thread: given an open-ended objective it uses tools to gather context across conflicting sources, corrects gaps in its plan and tracks what it has learned. It is trained to ask clarifying questions when a prompt is ambiguous, ask the user for help when stuck, confirm before consequential actions, and either report frequently or work silently depending on the user's preference. Meta also describe better adherence to long, multi-step instructions, more accurate mapping of new prompts to the right task within one mixed thread, and better awareness of what it can and cannot do, so that it reports hurdles instead of hallucinating outcomes.

Availability

Muse Spark 1.3 with max reasoning is available on Muse Code and the Meta Model API, with access through dev.meta.ai, and Meta's evaluation report gives further detail. Meta describe the release as advancing their work toward personal superintelligence, and drew on months of use of Muse Code and the Meta Model API in designing the release, and trained it across a diverse set of harnesses to generalise to different agentic environments. Its demonstration tasks include drafting an engineering simulation report from CFD data and a CAD file, editing and mixing an audio track from stems, and preparing presentations and summaries from spreadsheets.

Official Resources