Meta sells Muse Spark 1.3 as two API variants that differ in one documented respect: what Meta may do with the traffic a developer sends. The contributor variant costs $0.10 per million input tokens and $0.20 per million output tokens: 12.5 times less on input and more than 20 times less on output than the standard variant’s $1.25 and $4.25. That price applies only if the developer lets Meta use the traffic “to improve our products.” The rest of the launch, an agentic pitch and a benchmark chart, sits on top of the same trade.
Meta released Muse Spark 1.3 on 2 September 2026 in its Muse Code coding agent, for macOS and Linux, and in the Meta Model API. Both variants share a 1M-token context window and are also reachable through OpenRouter. Meta positions the model for longer-horizon agentic work: it asks clarifying questions when a prompt is ambiguous, “confirms before taking consequential actions,” and tracks several tasks inside one thread. Meta also says an open-weights release of Muse Spark is on its roadmap, without a date.
A discount that is paid for with data
The two API variants, muse-spark-1.3 and muse-spark-1.3-contributor, are listed on Meta’s model page with one label each: “Not used to improve our products” and “Used to improve our products.” That label is the entire difference the page documents, and it comes with a price gap large enough to change which tier a team defaults to. Agent workloads are the case where the gap matters most: an agent sends whole repositories, long tool-call transcripts and multi-turn plans, which is exactly the material a lab needs to train the next generation of agents. Choosing the $0.10 tier means choosing it with that material. Meta’s pages do not say how long the traffic is retained, whether it trains future models directly, or whether a developer can opt out after enrolling.
A chart built from each model’s best result
Muse Spark 1.3 is compared against Muse Spark 1.2, Claude Opus 5 and GPT-5.6 Sol, all at their highest reasoning setting (1.2 at xhigh), across professional work, computer use, web research, software engineering, long context and an internal instruction-following index, according to the four-page evaluation methodology Meta published alongside the model. The scores themselves are rendered only as charts on the blog; the methodology document explains how they were built rather than listing the numbers. Its rule: “for each model, we report the highest comparable primary-metric value available from our evaluation, the official leaderboard, or the model provider’s self-reported results.” Meta adds that its own runs of rival models are “best-effort” and “may not reflect their best provider-optimized performance,” and that refusals score zero and stay in the denominator. Taking the best available number per model fills gaps in a landscape where nobody can run every competitor optimally, but it also mixes three different measurement setups under one bar per model, by Meta’s own admission untuned for the rivals. A chart built that way shows which direction a model leans; it is a poor stand-in for a ranking a reader could defend line by line.
The agentic pitch has one comparison, and it is Meta’s own
Meta says Muse Spark 1.3 used about 20% fewer tool calls and 25% fewer tokens than 1.2 in coding work, and that it is better calibrated about what it can and cannot do. Both figures come from Meta’s own engineers’ comparison, with no published method behind either number. The company also says it “trained the model across a diverse set of harnesses to generalize to various agentic environments,” a claim aimed at exactly the cross-tool reliability that decides whether an agent finishes a task or stalls partway through it.
What the two choices add up to
A data-for-discount tier and a best-of-three benchmark chart, paired with an open-weights promise that still has no date, they describe a launch built around the two things Meta actually controls this week, the price of data and the framing of a comparison, while model size, training data and independent verification of the agentic gains stay outside what Meta discloses.