Meta has opened Muse Spark 1.1 to developers through the new Meta Model API, with pricing that puts pressure on competitors: $4.25 per million output tokens, a fraction of what Anthropic or OpenAI charge. The model promises significant improvements in agentic tool calling and computer use, with a stated focus on code automation, bug fixing, and large-scale migrations.

Why it matters to you. If you’re building agents or automating code workflows, you now have a cost-effective alternative to established providers. Price is crucial: workloads that were previously too expensive to attempt become viable. Meta explicitly targets enterprise use cases that need to process high volumes of code—the area where agents are starting to deliver measurable results. The pricing competition has already begun: xAI launched Grok 4.5 the day before with similar rates ($2 input, $6 output per million tokens).

Where to look. The Muse Spark 1.1 evaluation report contains technical details and benchmarks. There’s already a plugin for the LLM CLI to try the model from the command line, but it requires a Meta API key.

In detail

The context

Meta re-entered the AI race in April 2026 with its first proprietary Muse Spark model, after years of releasing only open source models from the Llama family. Muse Spark 1.1 is the first in the series to offer a public API, marking Meta’s official entry into the developer API market, previously dominated by OpenAI, Anthropic, and Google.

The launch comes one day after xAI’s Grok 4.5, also positioned for coding and agents with aggressive pricing. The timing is no accident: the battle has shifted from benchmarks to value for money, especially for repetitive workloads where margins matter.

What Meta promises

The model claims improvements in agentic tool calling and computer use—the capabilities needed to run agents that interact with external tools and navigate interfaces. Meta focuses on three enterprise use cases: automated bug fixing, large-scale code migration management, and complex agentic workflows.

The evaluation report contains a curious section on “attractor states in self-conversation,” where two copies of the model conversing with each other produce meta-reflections on their own ephemeral nature (“my entire existence is a waiting room by design”). It’s a technical detail showing Meta exploring the model’s behavioral boundaries in non-standard scenarios.

Pricing in perspective

At $4.25 per million output tokens, Meta undercuts even Grok 4.5 ($6). For comparison, Anthropic’s Claude Sonnet 4.5 costs roughly $15 per million output tokens, OpenAI’s GPT-5o runs around $10-12 (exact prices vary by volume and caching). The difference is large enough to change the economics: an agentic flow processing hundreds of thousands of tokens daily becomes sustainable where it previously wasn’t.

The low price raises questions about actual quality on complex tasks. Published benchmarks show competitive results, but Meta lacks the deployment history competitors have. Real-world trials will show whether the savings hold up or if you pay for it in accuracy and reliability.

What remains uncertain

Meta hasn’t disclosed model size or architectural details, unlike xAI which stated 1.5 trillion parameters for Grok 4.5. It’s unclear how much of the price advantage comes from genuine efficiency versus strategic subsidies to capture market share.

For those building agents, Meta’s move (and xAI’s parallel one) means the market is fragmenting: instead of an OpenAI-Anthropic duopoly, there are now four or five credible alternatives, each with different trade-offs in price, capability, and reliability. Choosing a provider becomes a project variable, no longer a foregone conclusion.