Meta Muse Spark 1.1 launch: API pricing undercuts rivals

Meta has released Muse Spark 1.1 and opened a public preview of the Meta Model API, offering a one-million-token context window and aggressive developer pricing that places output tokens at $4.25 per million. The model appears in the Meta AI app in "Thinking" mode and is built for agent orchestration, coding, multimodal understanding, and computer-use workflows.

What does Muse Spark 1.1 do and how does it perform?

Muse Spark 1.1 is a multimodal reasoning model trained to orchestrate multi-agent systems, to act as both a main agent that gathers context and delegates to parallel subagents, and as a subagent that stays on task and escalates when needed. Meta says the model actively manages a one-million-token context window, remembers actions, and retrieves and compresses earlier work without losing critical steps.

Meta highlights improved coding on large codebases, saying the model can diagnose complex bugs, add features to enterprise systems, and handle large-scale code migrations. On benchmarks, Muse Spark 1.1 leads four of twelve tests including MCP Atlas at 88.1 and Humanity's Last Exam at 62.1. On SWE-Bench Pro Muse scores 61.5 while Anthropic's Opus 4.8 scores 69.2. In the independent VALS-AI benchmark Muse Spark 1.1 ranks fourth and jumped 36 places on the Vibe Code Bench versus its predecessor.

How much does the Meta Model API cost and who does it squeeze?

Meta set developer pricing at $1.25 per million input tokens, $4.25 per million output tokens, $0.15 for cached input, and Web Search Grounding at $2.50 per 1,000 queries. Those rates undercut xAI's Grok 4.5, which had just become the price leader, and sit far below Anthropic, OpenAI, and Fable, which charge between $25 and $50 per million output tokens.

The pricing creates a new U.S. price floor among major providers, according to Meta. The company is moving away from the open-weights approach that characterized its earlier Llama strategy and is making Muse Spark 1.1 available through its own API while keeping image generation from Muse Image out of the initial API offering.

Chinese open-source models are also driving prices down from another direction. Meta cites examples where Snowflake has shown GLM 5.2 costs a fraction of what Opus 4.8 charges, and notes that companies such as Coinbase and Lindy have cut AI spending by switching to Chinese models. That dynamic squeezes pure-play AI labs between well-funded platform owners and low-cost open-source alternatives.

Why it matters

Meta can absorb low prices in ways smaller labs cannot. The company posts more than $60 billion in annual profit and can run APIs as gateways into its ecosystem without needing immediate profit on them. OpenAI and Anthropic, which the article says are burning through billions, rely on higher token margins to cover losses and justify valuations. Meta's low pricing forces those labs to justify margins on cost and token efficiency rather than headline model capability alone.

What to watch

Two practical signals will show if Meta's move sticks: real-world token efficiency per task and production performance versus the benchmarks. Databricks has shown that token burn per task can materially shift costs, and Meta notes it ran evaluations under the Advanced AI Scaling Framework with a detailed security analysis available. If Muse Spark 1.1's benchmark performance holds up in production while maintaining low token usage, competitors will face sustained pricing pressure.