Mark Zuckerberg returned to X after three years to push Meta's new Muse Spark 1.1 model, but the real story is the price Meta is willing to charge for agentic AI.
Zuckerberg hadn't posted on Elon Musk's platform since 2023. On July 9, he came back for a very Meta reason: to announce that the company was putting Muse Spark 1.1 behind a paid developer API. The model is not being sold as another chatbot with a fresh coat of benchmark paint. It's built to use tools, operate computer interfaces, and keep working on long tasks without a person checking every step.
That is the pivot. As The Verge reported, Muse Spark 1.1 is now available through the Meta Model API in public preview for U.S. developers, and Meta is also pushing the model through Meta AI across its own apps. The release follows the first Muse Spark launch in April from Meta Superintelligence Labs, the unit Zuckerberg built around Alexandr Wang after Meta's expensive effort to catch up with OpenAI, Anthropic, and Google.
The benchmark claims are pointed at one specific market. Meta says Muse Spark 1.1 scored 88.1 on MCP Atlas, which tests extended tool use, while Anthropic's Opus 4.8 and OpenAI's GPT-5.5 landed around 80. On JobBench, a professional tool-use benchmark, Meta says Muse Spark 1.1 reached 54.7, ahead of 48.4 for Opus 4.8 and 38.3 for GPT-5.5. On coding and multimodal reasoning, Meta's own positioning is more modest. It still trails the strongest rivals in places. That matters less than it sounds if the product is meant to do work across apps rather than win every model leaderboard.
Meta Wants the Agent Workload #
Muse Spark 1.1 carries a 1 million token context window, which is enough to hold a large codebase, a long document set, or a messy chain of tool calls without losing the thread too quickly. It is trained to operate desktop and mobile interfaces directly, including inside the browser, so it can click through a flow or fill in a form the way a human assistant would, moving between pieces of software as it goes.
Here is the practical part. The model can act as a lead agent, plan a job, hand parts of that job to sub-agents running in parallel, and pull the results back before deciding what to do next. If you're building agent software, that is not a side feature. It is the job. Nobody wants an agent that talks beautifully and then stalls when it has to use the tools sitting in front of it.
Business Insider reported that Zuckerberg described the pricing as very low, and Alexandr Wang has framed the offer around large-scale usage. That is the right pressure point. Developers don't only compare models by intelligence anymore. They compare how many calls a task will burn, and whether the final bill makes the product impossible to ship. How often the model fails matters too.
The Price Changes the Argument #
The Meta Model API prices Muse Spark 1.1 at $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits for new accounts. Barron's noted that this is far below OpenAI's higher-end GPT-5.5 pricing of $5 per million input tokens and $30 per million output tokens. That gap is not cosmetic. For an agent that may call a model dozens of times to finish one task, token price becomes product strategy.
Frankly, the price is the headline. Meta has spent two years taking criticism for the scale of its AI infrastructure bill. Reuters reported this week that the company is aiming to double compute capacity from 7 gigawatts this year to 14 gigawatts in 2027, and Barron's reported that Meta expects $125 billion to $145 billion in capital investments this year. You don't spend like that and then politely follow the API pricing table set by your rivals.
This is also new territory for Meta. The company has usually given models away, folded them into Meta AI, or used them to make Facebook, Instagram, WhatsApp, and its glasses more useful. Charging developers per token is a different business. It asks the market to treat Meta not only as a consumer AI distributor, but as a serious supplier of model infrastructure.
There is still a hard question underneath the launch. Cheap does not automatically mean better. If Muse Spark 1.1 loses too many coding or reasoning fights, developers will use it only for the jobs where price matters more than precision. But if Meta is even close on agentic workloads, OpenAI and Anthropic have a problem they can't brush off with a benchmark chart. A model that is good enough and much cheaper tends to find users fast.
Zuckerberg's return to X gave the launch its easy headline, especially because Threads exists and the old Musk rivalry still follows both companies around. Don't get distracted by that. The more important move is that Meta has put a public price on its AI work and made that price aggressive. Now developers get to decide whether Muse Spark 1.1 is only a loud entrance, or the first sign that the API market is about to get much less comfortable.
Also read: Elliott builds a stake in CCC Intelligent Solutions as a sale looms • SK Hynix's CEO Says the Memory Chip Crunch Will Outlast This Decade • What Is an AI Agent Marketplace and How Agents Get Bought and Sold