Morning Edition · Monday, August 10, 2026Published at 2:21 AM EDT · New York
Muse Spark 1.1 carries a one-million-token context window and sells at $1.25 per million input tokens, putting Meta into the metered API business it avoided for years.

Meta Superintelligence Labs has moved its Muse family behind a paid, metered interface. Muse Spark 1.1, released on July 9, is a multimodal reasoning model aimed at agentic work, with a one-million-token context window and improvements the company claims in tool use, computer use and coding. It is served through the new Meta Model API in public preview at $1.25 per million input tokens and $4.25 per million output tokens, with structured output, parallel tool calling, a files endpoint and prompt caching. That pricing sits below the headline rates of the top closed frontier tiers, which is the point.
On the media side, Muse Image shipped July 7, with direct-manipulation editing, multi-reference composition and agentic tool use. Muse Video was previewed on the same pretraining base, with native audio and no announced release date. CNBC reported that Meta's internal comparisons put Muse Image behind OpenAI's GPT Image 2 while ahead of Nano Banana 2 on single-image and multi-image editing. On Arena's human-preference rankings, Muse Image currently holds second place for text-to-image and for both editing categories.
There is no published model card or third-party benchmark suite for the image models yet, so the ranking is the most independent signal available, and internal comparisons are not.
The strategic shift matters more than any single score. Meta spent years distributing weights for free. It now sells tokens.
Meta, which converts a free distribution asset into metered revenue and undercuts incumbent API pricing, and price-sensitive developers who gain a cheaper mid-tier option as switching costs fall.
The $1.25 and $4.25 per million token rates and the second-place Arena standing are independently confirmed, with , but Arena measures crowd preference rather than editing fidelity, Meta's own comparisons remain unaudited, and Mark Zuckerberg's characterization of the price as roughly a quarter of rival rates is a seller's comparison, not a like-for-like benchmark.
Part of a tracked trend
Hyperscalers Enter the Frontier API Market
Large platform companies increasingly convert in-house frontier models into metered, incumbent-compatible APIs, competing on switching cost and distribution rather than pure capability.
Start a discussion in Townsquare.
More from this edition
An open-source-intelligence read of how likely this story is true with its real nuance, not a judgment of any outlet. It assesses the claim, weighing independent and adversarial reporting. How we label confidence.
What this means
Meta converting in-house models into a metered, OpenAI-compatible-style API turns a distribution asset into a direct competitor for developer spending, and it competes on price and context length rather than on top-of-leaderboard capability. The exposure is to mid-tier API revenue at OpenAI, Anthropic and Google, where buyers are price-sensitive and switching cost is low because the interface conventions are converging. It also changes Meta's own incentives on open weights: a company selling tokens has a reason to keep its best model closed, which weakens the assumption that Meta remains the reliable supplier of frontier-adjacent downloadable weights.
What to watch
Observations to monitor, not financial advice.
Comments
0No comments yet.