Meta quietly became a serious player in the agent model race in July 2026. Muse Spark 1.1 is its first paid agentic model API, and unlike most "we added agents too" launches, the architecture is actually built around the idea from the ground up.
What Muse Spark 1.1 is built to do
This isn't a chatbot with tool access bolted on. Muse Spark 1.1 is a multimodal reasoning model designed specifically for agentic performance, tool use, and computer use, and Meta's own benchmarks lean hardest into those three areas rather than general knowledge tests.
The model ships with a 1,048,576-token context window, rounded to "1 million" in the marketing copy, and it can actively manage that context instead of just filling it up. It remembers earlier actions, retrieves information from much earlier in a session, and compacts the context to keep the steps that matter for later work. That matters a lot for anything that runs for more than a few minutes.
Parallel subagents and computer use
The standout feature is delegation. As the main agent, Muse Spark 1.1 can gather context, build a plan, and hand off pieces of execution to parallel subagents running at the same time, which is how it handles complex, multi-part projects faster than the original Muse Spark.
It's also trained to drive real interfaces: desktop, mobile, and browser. Give it a goal in plain language and it can navigate a page, click through a flow, read what's on screen, and report back, alongside standard tool and function calling, structured outputs, web-search grounding, and multimodal input across images, video, and PDFs.
What it costs
Meta's pricing undercuts the two biggest players by a wide margin:
| Muse Spark 1.1 | Typical frontier pricing | |
|---|---|---|
| Input tokens | $1.25 per million | $2 to $3+ per million |
| Output tokens | $4.25 per million | $10 to $15+ per million |
| Cached input | $0.15 per million | Varies |
New Meta Model API accounts start with $20 in one-time credits. One catch worth flagging: because Muse Spark 1.1 is a reasoning model, its internal chain-of-thought tokens are billed at the full output rate rather than a discounted one, so reasoning-heavy workloads cost more than the headline numbers suggest.
How it compares
Meta is pricing Muse Spark 1.1 at roughly a quarter of what Anthropic and OpenAI charge for comparable agentic models, which is the real story here. If you've read our breakdown of [Claude Sonnet 5's pricing](https://www.questloops.com/blog/claude-sonnet-5-explained-what-changed-what-it-costs-and-who-should-switch) or the [GPT-5.6 model lineup](https://www.questloops.com/blog/openai-s-gpt-5-6-explained-sol-vs-terra-vs-luna-and-which-one-to-use), Muse Spark 1.1 is Meta's answer to both, undercutting them on cost while matching the agentic feature set that's become table stakes this year.
The takeaway
Muse Spark 1.1 is worth testing if you're building anything that needs a model to actually operate software rather than just talk about it: browser automation, multi-step research, or long-running coding agents. The $20 starting credit makes it cheap to try, and the pricing holds up well against Claude and GPT alternatives as long as you're not running reasoning-heavy workloads that eat into the savings.
