Meta has released Muse Spark 1.1, a new AI model built for coding and agentic tasks. Alongside it, the company opened a public preview of the Meta Model API, offering paid access to one of its own AI models for the first time. By leaning on aggressive pricing, Meta is making a serious push into the developer market that OpenAI and Anthropic already lead.

Entering the developer market with a first paid API

To coincide with the launch, Meta opened a public preview of its developer interface, the Meta Model API. After months of limited testing with a small group of partners, developers in the United States can now experiment with the model, compare its output against others, and build it into their own applications. New sign-ups start with 20 USD (about 3,200 yen) in free credits, after which usage shifts to pay-as-you-go billing.

This is the first time Meta has charged for access to one of its own AI models. For a company that has poured tens of billions into chips, research, and data centers, it marks an important step toward turning AI into revenue.

Built for agentic work and tool use

Muse Spark 1.1 is designed around agentic use: beyond writing and debugging code, it can call external tools and operate on-screen interfaces. According to Meta, the model is multimodal, handling text, images, and video, and it can work across desktop, mobile, and browser screens to carry long-running tasks through to completion.

The model supports a context window of up to 1 million tokens and can split work across multiple sub-agents running in parallel. In a post on X (formerly Twitter), Mark Zuckerberg highlighted agentic performance, tool use, and computer use as the model's biggest strengths. It is available in a "Thinking" mode through the Meta AI app and on the web.

Competing on price

Pricing is set at 1.25 USD (about 200 yen) per million input tokens and 4.25 USD (about 690 yen) per million output tokens. That is higher than OpenAI's GPT-5 mini and Anthropic's Claude Haiku 4.5, but lower than the higher-end Claude Sonnet 4.6, placing Muse Spark 1.1 squarely between the budget and premium tiers.

In its own comparisons, Meta lines the model up against GPT-5.5, Opus 4.8, and Gemini 3.1 Pro, arguing it is competitive on coding and agentic tasks. Those claims rest on internal evaluations, however, and independent verification is still to come. Some in the market suggest that if the model truly matches Claude and Codex on coding, Meta may finally have a clear path from AI models to paid developer tools.

*1 USD = 162 JPY (as of July 13, 2026)

Rolling out across Meta's services as a Llama successor

Muse Spark comes from the Superintelligence team Meta assembled last year, with the first version arriving in April. Muse Spark 1.1 is the second entry in that line. Going forward, Meta plans to gradually replace the Llama models that currently power WhatsApp, Instagram, Facebook, and its smart glasses with Muse Spark.

Two days before this launch, the same Superintelligence team unveiled Muse Image, Meta's first image-generation model. The company is assembling a broader family of models aimed not only at consumer chatbots but at developers, creators, and enterprises. Underlying it all is Zuckerberg's vision of "personal superintelligence."

Summary

The release of Muse Spark 1.1 and its paid API marks a turning point in Meta's serious entry into the developer AI market. With a long 1-million-token context and agentic capabilities as its strengths, the company is challenging OpenAI and Anthropic on price. Whether its performance claims hold up under independent testing, and how widely the model gets used in real development work, will be the key things to watch.