Meta Superintelligence Labs released Muse Spark 1.1 on July 9, an upgraded multimodal reasoning model aimed at agentic and coding work, alongside a public preview of a new Meta Model API for developers. It is Meta's first paid frontier-model API offering, arriving about three months after the unit — led by chief AI officer Alexandr Wang — shipped its first Muse Spark release.

What changed in 1.1

Meta says Muse Spark 1.1 can actively manage a one-million-token context window: retaining actions from earlier in a session, retrieving information from much earlier in a long task, and compacting its own history in a way meant to preserve the steps still needed for later work. The model adds native support for orchestrating a primary agent alongside subagents, plus support for the Model Context Protocol (MCP) and custom skills, and Meta reports substantial gains on real-world coding tasks — diagnosing and fixing bugs, implementing features, and executing migrations in large, complex codebases — over the original Muse Spark.

Pricing and benchmarks

The Meta Model API preview, limited to US developers at launch, prices Muse Spark 1.1 at $1.25 per million input tokens and $4.25 per million output tokens, with cached input at $0.15 and $20 in free credits for new accounts — well under the list prices OpenAI and Anthropic charge for their comparable coding-and-agent-tier models. On MCP Atlas, Meta's tool-use benchmark, the company reports a score of 88.1, ahead of the high-70s-to-low-80s range it attributes to Claude Opus 4.8 and GPT-5.5. On Terminal-Bench 2.1, Meta reports 80.0, trailing GPT-5.5 (83.4) and Claude Opus 4.8 (82.7); independent benchmarking firm Vals AI has measured a lower score on that same test, a gap of more than ten points that underscores the usual caution around vendor-reported numbers.

Why it matters

Muse Spark 1.1 is Meta's clearest bid yet to compete directly with OpenAI and Anthropic on agentic coding and tool use rather than only on open-weight model releases, and its pricing is aggressive enough to pressure both rivals' API rates. But the gap between Meta's self-reported Terminal-Bench score and Vals AI's independent measurement is a reminder that agentic benchmarks remain contested ground: strong tool-use numbers don't automatically translate into equally strong raw coding performance, and buyers evaluating Muse Spark 1.1 against Claude or GPT-5.6 will want results from more than one source before switching.