Muse Spark 1.1

Meta built its AI identity on giving models away: download the Llama weights, run them yourself, pay nothing. Muse Spark 1.1 ends that as the default. It ships only through a new paid, metered developer API — $1.25 per million words of input, $4.25 per million out, with $20 of free credits to start — the first time Meta has sold access to one of its models the way OpenAI and Anthropic do. The model itself is not the news; the meter is.

The company that made open weights its whole argument is now competing on price, not capability.

The pricing is the whole pitch. Zuckerberg put it at roughly a quarter of what comparable frontier models cost — Anthropic's Claude Opus 4.8 runs $5 in, $25 out; GPT-5.5 is dearer still. And Meta made switching almost frictionless: the API answers to both the OpenAI and the Anthropic software kits, so a developer already wired into either can point at Meta's endpoint and change little else. This is a migration play aimed straight at the two incumbents' installed base.

The candor is in the benchmarks. Muse Spark 1.1 — a multimodal reasoning model that uses tools, writes code, and holds a million words in context — tops the scaled tool-use tests but loses on coding accuracy to both Opus 4.8 and GPT-5.5. So the offer is not 'best,' it's 'good enough at agentic work, and far cheaper.' For a company that spent years arguing open weights would win the platform, charging by the token — and competing on price rather than capability — is the strategic tell worth watching.