Meta drops Muse Spark 1.1 with 1M context

If you have ever watched an AI agent try to operate your desktop by blindly clicking one button at a time, you know the pain.

Meta Superintelligence Labs just dropped Muse Spark 1.1, a major multimodal reasoning upgrade built specifically for agentic workflows, coding, and computer use. Alongside the release, Meta rolled out a public preview of its new Meta Model API and shipped the model directly into Thinking mode on meta.ai and the Meta AI app.

The launch was significant enough that Mark Zuckerberg posted on X for the first time in three years to hype up its raw agentic performance and low pricing.

A master planner with a million-token memory

Muse Spark 1.1 is built to tackle long, messy digital tasks without human babysitting.

Its flagship trick is smart computer use. Instead of pausing to reason through every single click, Spark 1.1 decides on the fly whether it is faster to write an automation script or click through the graphical interface directly, often firing off actions in parallel batches.

In one demo, a user provided a quick smartphone clip of an item. Spark 1.1 pulled high-res stills from the video, reasoned about the specs, navigated a desktop browser, and published a complete listing on Facebook Marketplace.

Under the hood, the model can manage a massive 1 million token context window, actively compacting past steps to preserve what matters. It can also act as an orchestrator, spinning up and delegating tasks across parallel subagents to slash overall latency.

It features zero-shot support for Model Context Protocol (MCP) servers and native developer tools. In agentic coding tests via OpenCode, it built web apps from scratch, took automated screenshots of UI bugs, traced failures back to the codebase, and patched them.

According to Reuters, Meta set API pricing at $1.25 per million input tokens and $4.25 per million output tokens, positioning it directly against Anthropic's Claude Haiku 4.5 and OpenAI's freshly dropped GPT-5.6 Luna.

Why it matters

We think Meta is executing a classic commoditization play here, and it looks remarkably sharp.

For developers and technical builders, Spark 1.1 solves the biggest roadblock in production agents: cost and coordination overhead. Having a million-token multimodal foundation that natively orchestrates subagents, connects to MCP endpoints, and runs at near-Haiku pricing makes building automated workflows far more viable today.

For the AI ecosystem, Meta is putting immense pressure on OpenAI and Anthropic. Those two pioneers held the high ground on coding agents and computer interaction, but Meta is packaging comparable agentic capabilities into an accessible, developer-friendly API.

The real risk, however, lies in autonomy. Giving a reasoning model the freedom to alternate between GUI clicks and background script execution on real operating systems is inherently delicate. While Meta touts strong safety evaluations against prompt injections, enterprises will likely tread carefully before handing it the keys to production infrastructure.