Muse Spark 1.3

Meta Superintelligence Labs' latest Muse Spark, trained for the agentic builds developers actually ship: long-running, multi-agent workflows where the model has to track prior results, work through conflicting inputs and ask when it is genuinely stuck. Meta reports it taking roughly 20 percent fewer tool calls and 25 percent fewer tokens than 1.2 while scoring higher, and calls it less verbose and less prone to unnecessary turns. It reads video, images and documents alongside text across a million-token window, with visual reasoning that runs in a real execution environment rather than a scripted one. Effort runs from minimal through max.

Muse Spark 1.3 is a release about restraint, which is not how model launches usually read. Meta's framing is that it is tuned for the agentic builds developers actually ship - long-running, multi-agent workflows - and the improvements it puts forward are as much about what the model stops doing as what it starts doing. It takes fewer turns where turns are not needed, it is less verbose, and in Meta's own engineers' comparisons it used roughly 20% fewer tool calls and 25% fewer tokens than 1.2 while scoring higher.

That combination is the whole argument. On a long-horizon agentic task the cost is dominated by turns that did not need to happen and context that did not need to be re-read, so a model that reaches the same place in fewer steps is cheaper twice over: fewer tokens billed, and less time spent. Meta reports the quality moving in the right direction at the same time, with DeepSWE v1.1 - its long-horizon agentic coding evaluation - rising to 1754 from 1615 for 1.2.

What the model does with a long task has changed in kind, not just in efficiency. Meta describes it tracking context and prior results across a workflow, working through messy or conflicting inputs rather than picking one and proceeding, and asking for input when it genuinely needs it instead of guessing. It pairs that with better awareness of its own capabilities and better calibration around irreversible actions - which is the property that decides whether an agent can be left alone with anything that writes.

The input side is unchanged in breadth and remains unusually wide: video, images and documents alongside text, across a one-million-token combined window, with visual reasoning that Meta says runs through a real execution environment rather than a scripted pipeline. Feed it a screenshot or a clip and it builds from what is actually there. Effort runs from minimal through max, with max added for the hardest reasoning and agentic work.

Pricing is identical to 1.2 - $1.25 per million input tokens, $0.15 cached, $4.25 output - which, combined with the token and tool-call reductions, means the same work costs less on 1.3 than it did on 1.2 at the same posted rate. Note that Meta also publishes a contributor tier of this model at a much lower price which trains on submitted prompts; that is a different model id and is never routed here.