Muse Spark 1.3 keeps its API price, with its top 'max' mode added after launch

Muse Spark 1.3 is in Meta's API at an unchanged $1.25/$4.25 per million tokens with 1M context; Meta held back its top 'max' reasoning mode at launch for more safety testing and now says it is available.

Key points

• Model IDs muse-spark-1.3 and muse-spark-1.3-contributor, each with a 1M-token context and native multimodal input covering video, images and documents, per Meta's developer page, which names three access routes: Meta Model API, Muse Code and OpenRouter.
• Standard pricing is unchanged from 1.2: $1.25 input, $0.15 cached input, $4.25 output per 1M tokens. The developer page lists the contributor tier at $0.10/$0.002/$0.20, and Meta may use those prompts and completions for training.
• Meta's own claim, from its engineers' internal comparisons: roughly 20% fewer tool calls and about 25% fewer tokens than Muse Spark 1.2 on coding work. Not independently verified.
• Meta's evaluation report says 1.3 was scored at max reasoning effort, with 1.2 at xhigh.
• At launch, Meta's post held max back for more safety testing (archived copy, September 4). VentureBeat's September 3 report relayed that, adding that Artificial Analysis had tested max in a limited partner preview and listed no API provider for it. Meta has since updated the post, which still carries its September 2 date, to say max is available in Muse Code and the Meta Model API; the new wording first shows up in an archived copy from September 5. Artificial Analysis's page for max now lists one API provider, Meta's.
• Meta-reported figures relayed by VentureBeat, xhigh vs max: GDPval-AA v2 1,709 vs 1,754 Elo; OSWorld 2.0 57.2 vs 66.9 (Meta's report now puts xhigh at 59.0); JobBench 61.2 vs 64.9; and on Terminal-Bench 2.1 Meta has xhigh slightly ahead, 89.2 to 88.8. Artificial Analysis's September 2 write-up put xhigh at 61 and max at 62 on its Intelligence Index, and its own Terminal-Bench 2.1 run had max slightly ahead, 86% to 85%.
• The Muse Code terminal tool installs on macOS or Linux, per Meta's launch post. None of the three routes on the developer page is a hyperscaler (AWS, Azure, Google Cloud). Meta's roadmap mentions a Muse Spark open-weights release but gives no version, date, size or licence; VentureBeat notes Zuckerberg said in August that 1.2's weights would be opened within weeks.

“frontier performance almost too cheap to meter.” — Mark Zuckerberg, on X, as quoted by VentureBeat

Why it matters

For a studio shipping agent workflows, the numbers that matter are the unchanged $1.25/$4.25 standard rate, the 1M context, and a contributor tier that drops input cost about 12x provided you can accept Meta training on that traffic, which rules out most client codebases. Read the benchmark table with care: Meta scored 1.3 at max, a mode it held back at launch and only later said was available, and xhigh trails max on several of Meta's own figures. The efficiency gains over 1.2 are Meta's own internal measurements. If you were holding out for Muse Spark open weights to self-host, there is still no version, date or licence.

This is an AI-written summary of the reporting credited above and the other sources linked in the text, read and edited by Nicholas before publishing. The facts and any quote belong to those sources; the wording is ours. Read the original.