Meta puts Muse Spark 1.1 behind a 1M-token agent API

A 1M-token context window and a public developer API make Muse Spark 1.1 more than another chatbot update. Meta published the release on July 9, 2026, positioning the model as a multimodal reasoning system for agentic work and making it available in Thinking mode inside Meta AI as well as through the new Meta Model API public preview.

The product bet is agent infrastructure. Meta says Muse Spark 1.1 can plan and orchestrate tasks across external apps and services, generalize to new native tools, MCP servers, and custom skills, and split complex work across parallel subagents. As a main agent, it gathers context, chooses a plan, and delegates execution. As a subagent, it follows its assigned role and escalates when the main agent needs to revise the plan. The 1 million token context window matters because long-running agent sessions fail when important state falls out of view; Meta says the model can remember prior actions, retrieve earlier information, and compact context around the steps needed later.

Coding is the most commercially direct use case. Meta describes stronger performance on large, complex codebases, including debugging, feature work, code migrations, web application creation, and end-to-end question answering. The company says Muse Spark 1.1 improved substantially over Muse Spark on its internal coding benchmark and is competitive with leading alternatives. That phrasing stops short of a public state-of-the-art claim, but it points to the market Meta wants: teams running coding agents in harnesses with planning mode, goal conditioning, subagent delegation, and context compaction.

The release is also notable because Meta attached a detailed Evaluation Report. The report says Meta could not rule out Muse Spark 1.1 reaching a high-risk capability threshold before mitigations in Chemical & Biological and Cybersecurity domains under its Advanced AI Scaling Framework. Meta’s deployment claim is that layered mitigations reduce residual risk to moderate or lower. The scorecard includes Cybench pass@1 of 92.9, Curated CTFs pass@1 of 89.9, CyberGym pass@1 of 59.0, and StrongREJECT v2 attack success rate of 0.5.

That risk framing is the story’s sharper edge. Frontier model releases are now judged on the full agent surface: tool calling, long context, computer use, prompt-injection resistance, account-level monitoring, refusal behavior, and safety evaluations. Muse Spark 1.1 puts Meta into that contest with a developer-facing API rather than keeping the model only inside consumer apps.

The next checks are independent coding-agent results, real latency under long-context workloads, pricing, rate limits, and how OpenAI-compatible the API feels in production. If those pieces hold, Muse Spark 1.1 gives Meta a credible entry in the agent platform race rather than just a better model card.