Muse Spark 1.1: Meta Model API, Coding & Pricing

A metered endpoint, not a download

Meta Superintelligence Labs described Muse Spark 1.1 as a multimodal reasoning model built for agentic tasks: diagnosing bugs and executing code migrations in enterprise systems, automating multi-step work across separate applications, and delegating pieces of a job to parallel subagents. Meta reported a one million token context window and said the model resists jailbreak and prompt injection attempts. The model reached general availability on July 9, 2026 through a new Meta Model API, launched the same day in public preview, and through a Thinking mode in the Meta AI app and at meta.ai. Meta did not release the model's weights. [1]

Meta's own comparisons were against what it called the Meta Internal Coding Bench, an evaluation it built itself, where it reported Muse Spark 1.1 performing competitively against other agentic coding models on tasks drawn from large, real codebases. It reported no results from an independent benchmark. [1]

Atlas interpretation: The absence of independent numbers means the only verifiable claim about capability is the one about access: a model reachable solely through Meta's own metered API, with no weights to download and no independent lab free to reproduce the internal benchmark that is the only evidence offered for it. [1]

The turn away from open weights happened three months earlier

Muse Spark 1.1 was not Meta Superintelligence Labs' first closed model. The lab's original Muse Spark, released April 8, 2026 after nine months of rebuilding Meta's model organization, had already shipped without published weights, breaking with the pattern Meta had followed since Llama. Muse Spark 1.1 extended that decision into a second model and, for the first time, into a public API business built around it. [1][2]

Atlas interpretation: That sequencing matters for how the event is read. The identity break, a lab known for giving weights away choosing not to, was April's news. July's news is that Meta turned the same closed model line into a metered product instead of reversing course, which is a different and smaller kind of decision: having already closed the weights, selling access to them was the more conventional half of the choice. [1]

Priced next to Anthropic and OpenAI's cheap tiers, launched next to OpenAI's flagship

Meta priced the Model API at $1.25 per million input tokens and $4.25 per million output tokens, which TechCrunch reported sits close to Anthropic's Claude Haiku 4.5 and OpenAI's GPT-5.6 Luna, Meta's rate running slightly higher than both. GPT-5.6 itself, in its Sol, Terra and Luna tiers, became publicly available the same day, a coincidence of timing rather than a response, since Meta's release had been planned independently. [2]

Mark Zuckerberg posted about the release on X, his first post on the platform since July 2023, around when it was still rebranding from Twitter. He called Muse Spark "a strong agentic and coding model at a very low price," pointing to its agentic performance, tool use, and computer use. [2]

Atlas interpretation: Zuckerberg's pitch, like Meta's later framing of Muse Code a month later, is about price rather than a claimed edge in quality. Setting the rate just above two products already read as cheap, Claude Haiku 4.5 and GPT-5.6 Luna, rather than undercutting them, is a smaller ambition than the coding-agent pricing that followed it: Meta was matching the market it had just entered, not trying to move it. [2]