Muse Spark 1.1 is Meta’s newest AI model, built to do work rather than chat. It runs teams of copies of itself, drives a browser the way a person does, and costs a quarter of Claude Opus 4.8 on input tokens and a sixth on output.
What is Muse Spark 1.1?
The name decodes the same way Claude’s does. Muse is the family, Spark is the model line inside it, and 1.1 is the update to 1.0, which shipped in April. It is fully multimodal with a one million token context window, matching Opus 4.8 and OpenAI’s GPT-5.5. Meta has not published the parameter count or the architecture. You can use it free in the Meta AI app today, and over the next few weeks it is expected to replace the older models running the AI inside WhatsApp, Instagram and Facebook. Meta also launched the Meta model API, the first time outside developers can pay Meta to build on one of its models.
Two things stand out in how it works.
First, it delegates to itself. A main copy plans the big job and hands pieces to parallel copies called sub-agents, the way a lead engineer splits the day’s calls across the vans and keeps the whole schedule in his head.
Second, it operates a computer the way you do, clicking interfaces and typing into browsers on desktop and mobile, and it decides for itself whether a script or the screen is faster.
Meta showed one demo worth watching. You film a short video of something you want to sell, an old chair or a bike. The model watches the video, pulls the best frames, works out what the product is and what it is worth, then opens your browser and builds the Facebook Marketplace listing, photos and title and description and price included. Nothing about that sequence is specific to Facebook Marketplace.
Why did Llama 4 fail, and what Meta spent to recover
Llama, Meta’s open model line, ran for years under Yann LeCun, hired in 2013 to build the FAIR lab on the condition that whatever he built stayed open and free. Llama 4 was meant to put Meta in front. It flopped in April 2025 because the team was caught inflating its own benchmark numbers, submitting one model to public leaderboards and shipping a different one to developers. On the independent Intelligence Index, Llama 4’s best version scored 18 while the leaders sat in the high 50s.
Meta then paid $14.3 billion for a 49% non-voting stake in Scale AI, mostly to get its 28-year-old founder Alexandr Wang as chief AI officer, and raided OpenAI, Anthropic and Google DeepMind at $100 million to $300 million a head. Some researchers ignored Zuckerberg’s DMs because they assumed they were fake. LeCun left, over 600 people lost their jobs, and in April 2026 the new lab shipped Muse Spark 1.0 at a score of 52 on the index where Llama 4 scored 18.
How much does Muse Spark 1.1 cost?
| Model | Input, per million tokens | Output, per million tokens |
|---|---|---|
| Muse Spark 1.1 | $1.25 | $4.25 |
| Claude Opus 4.8 | $5 | $25 |
| Fable 5 | $10 | $50 |
Input is what an agent reads, output is what it writes, and both are billed per million tokens. Reading is where the volume sits. An intake agent that goes through every inbound job request, every photo caption and every previous note on the customer burns input tokens all day and writes back three lines. At $1.25 against $5 that day of reading costs a quarter as much, and on output the gap is wider, $4.25 against $25.
That is the whole cost story, and it only shows up on a usage bill rather than a subscription. If your automation runs on a per-seat plan, the table above changes nothing for you this month.
Muse Spark 1.1 vs Opus 4.8 and GPT-5.5: the benchmarks
| Model | Humanity’s Last Exam, with tools | SWE-Bench Pro |
|---|---|---|
| Muse Spark 1.1 | 62.1 | 61.5 |
| Claude Opus 4.8 | 57.9 | 69.2 |
| GPT-5.5 | not given | behind 61.5 |
| Fable 5 | not given | 80.3 |
| GLM 5.2 | not given | 62.1 |
Humanity’s Last Exam with tools is a brutal reasoning test where the model gets search and code, and Muse Spark 1.1 beats Opus 4.8 on it. It also takes the top spot on the two big tool-use benchmarks, MCP Atlas and Job Bench. Hard coding is weaker. Its 61.5 on SWE-Bench Pro ties GLM 5.2, the open Chinese model you can download and run yourself for nothing, and on terminal work it comes third behind GPT-5.5 and Opus. Best at orchestrating tools and running errands, a step behind at writing hard code.
Every number above comes from Meta’s own launch table, and Meta chose which modes to show its rivals in. This is the company that admitted juicing Llama 4’s benchmarks. The eval report is public and independent testers will have their own figures within days, but as of today there is no independent score for 1.1. The full set Meta published, including the safety numbers, is in the technical breakdown of Muse Spark 1.1.
Should you switch to Muse Spark 1.1?
Not this quarter. Watch the computer use rather than the price.
Take a warranty claim on a compressor that failed inside its cover. Someone in your office logs into the manufacturer’s portal, types the serial number off a photo the engineer sent from the roof, picks a fault code from a dropdown, uploads the invoice, submits, then pastes the claim number back into the job so the customer can be told. That portal has no API and never will, which is why every automation quote you have had either skips that job or prices a person to sit and do it. A model that clicks interfaces and reads photos runs the whole sequence, and it is the Marketplace demo with a different website in front of it. Supplier reordering, utility interconnection paperwork and the insurer’s claims site all look the same from the model’s side.
So the useful work this week is a list. Write down every job that stalls because the software involved has no API and someone has to type into a screen. That list is what an agent with computer use is worth to you, and it does not change if the benchmark table turns out to be generous. Then ask whoever built your automation what it spent on tokens last month, and what that same month costs at a quarter of the input rate.
Two things are true about Meta regardless. Wang says a model trained on ten times the compute behind Muse Spark 1.0 is already matching GPT-5.5, with an Opus-level coding model coming soon. And Meta ships this model as the default brain inside apps that more than 3 billion people already use, so your customers will be talking to it inside WhatsApp whether or not you ever call the API. Write the list, get last month’s token bill, and check the independent scores when they land in a few days.
