Meta Enters the Coding Agent Race with Muse Spark 1.1

Meta announced the launch of Muse Spark 1.1 on July 9, a multimodal language model built for agentic and coding tasks. It marks the first time Mark Zuckerberg's company has offered paid API access to one of its models, a clear shift away from its longstanding focus on free, open-weight models in the Llama family.

The model went straight into public testing in the United States, accessible both through the Meta AI app's Thinking mode and via the new Meta Model API for developers. Every new API account starts with $20 in free credits, meant to let engineering teams test the model without financial risk.

Pricing Undercuts Rivals

The most striking part of the launch is the price list. By most comparisons, Muse Spark 1.1's rates run roughly four times lower than comparable offerings from Anthropic and OpenAI. Alexandr Wang, Meta's chief AI officer, described the pricing as deliberately aggressive toward competitors.

Very aggressive and attractive - Alexandr Wang, chief AI officer at Meta, on Muse Spark 1.1's pricing

Meta is betting that lower rates will draw in startups and engineering teams that had previously held back from heavy use of agentic models due to scaling costs. The million-token context window lets the model work across many files in large codebases at once, useful for code migrations or fixing complex bugs in enterprise systems.

What the Model Can Do

Muse Spark 1.1 handles multi-step task planning, delegates work to subagents, and automatically compresses context during long sessions. The model also supports computer use, deciding on its own whether it's faster to automate a task with a script or interact directly with an application's interface. On top of that, it offers extended multimodal reasoning, including generating code from images and describing video content.

Amjad Masad, CEO of Replit, one of the first platforms to integrate the new model, praised the completeness of its feature set.

A complete agentic foundation - Amjad Masad, CEO of Replit, on Muse Spark 1.1

Masad pointed to the combination of a large context window, full multimodal support, built-in search with source citations, strong reasoning, and solid frontend coding performance, all bundled into one OpenAI API-compatible package.

An Infrastructure Arms Race

The launch fits into Meta's broader strategy, which aims to boost its data center compute capacity to 14 gigawatts by 2027. The company is also developing its own MTIA400 chips, expected to enter mass production in September 2026 and reportedly deliver roughly 400 percent higher performance than the previous generation. In-house hardware is meant to eventually lower the cost per query and let Meta sustain its aggressive pricing for longer.

For Polish software companies and teams building developer tools, this adds another, cheaper option to an increasingly crowded field of agentic models, alongside Anthropic's Claude, OpenAI's GPT-5.6, and xAI's Grok 4.5. Price pressure among providers translates directly into lower deployment costs for coding agents at Polish software houses, which are increasingly billing clients for AI-generated work rather than just programmer hours.

Meta has not yet disclosed plans to expand Muse Spark 1.1 access beyond the United States, meaning European and Polish developers can for now mainly reach the model through third-party platform integrations like Replit or Cline, provided those platforms make it available to users outside the US.

Sources: Introducing Muse Spark 1.1 (ai.meta.com), Meta jumps into AI coding market in effort to chase Anthropic and OpenAI (cnbc.com), Meta launches flagship Muse Spark 1.1 model with multi-agent upgrades (siliconangle.com), Meta Launches Muse Spark 1.1: A Lower-Cost AI Model for Coding Agents (eweek.com)