Meta released Muse Spark 1.1 on July 9, 2026, and opened its first paid model API, pricing input/output at $1.25/$4.25 per million tokens—roughly one-quarter the cost of Claude Opus 4.8 or GPT-5.5. The model, built by Meta Superintelligence Labs under AI chief Alexandr Wang, targets coding and agentic workflows with a 1-million-token context window, native multimodal support (images, video, PDFs), and active context compaction that tracks state across extended sessions. Early partners including Replit, Cline, and Box praised tool-orchestration depth and cost-performance. The API is in public preview for US developers with $20 free credits per account, accessible via OpenAI-compatible and Anthropic-compatible SDKs (set base URL to api.meta.ai/v1, model name muse-spark-1.1).
Muse Spark 1.1 tops several specialist benchmarks: MedScribe and TaxEval (previously held by Fable 5), and Harvey's Legal Agent Bench (dethroning Grok 4.5). Internally, Meta trains a larger model codenamed Watermelon that reportedly matched GPT-5.5 on key benchmarks. The release marks Meta's philosophical shift from open-source Llama (permissive distribution) to proprietary, paid-API-only inference—a direct commercial bet against Anthropic and OpenAI on developer cost-sensitivity and tool-use reliability in agentic systems. Every new API account receives $20 in free credits; after that, token-based metering applies.
For ML engineers evaluating coding assistants: at ~25% of frontier closed-model list prices with claimed business-task parity on agentic evals, Muse Spark 1.1 lowers the price bar for multi-agent orchestration and long-context tool-use workloads. The SDK compatibility removes switching friction—teams can A/B models against existing Anthropic or OpenAI stacks without re-plumbing. However, production readiness depends on external benchmarks (internal evals often favor the host) and compatibility testing on your MCP servers, retrieval chains, and structured-output requirements.