Meta has launched Muse Spark 1.1, a new multimodal reasoning model from Meta Superintelligence Labs, alongside the public preview of the Meta Model API.
The launch is important for two reasons. First, Meta is pitching Muse Spark 1.1 as a serious agentic model for coding, computer use, long-context work, and multimodal workflows. Second, Meta is now charging developers for direct model access, putting the company into the same API market as OpenAI, Anthropic, Google, xAI, and the newer low-cost model providers.
Meta says Muse Spark 1.1 is available in Thinking mode in the Meta AI app and on meta.ai. Developers can also begin building with the model through the new Meta Model API, which is currently in public preview.
What Muse Spark 1.1 is
Muse Spark 1.1 is not an image-only model. It is Meta's updated multimodal agent model, designed to reason across text, images, video, documents, tools, code, and browser or computer-use tasks.
Meta says the model is a major upgrade over the original Muse Spark, especially in four areas:
- agentic task planning and tool use
- computer-use workflows across apps and interfaces
- coding in large, complex codebases
- multimodal perception and reasoning
The model also supports a 1 million token context window, which matters for long-running agents that need to preserve instructions, project history, files, earlier tool outputs, and user corrections across a large task.
The clearest product direction is multi-agent work. Meta says Muse Spark 1.1 can act as a main agent that gathers context, plans the task, and delegates work to parallel subagents. It can also act as a subagent that follows a narrower assignment and escalates when needed.
That puts Muse Spark 1.1 in the same strategic category as newer coding and workplace agents: models are no longer being judged only on chat quality, but on whether they can execute multi-step work inside real software.
Pricing
Meta is using price as the headline commercial move.
Through the Meta Model API, reported pricing for Muse Spark 1.1 is:
| Item | Price |
|---|---|
| Input tokens | $1.25 per 1 million tokens |
| Output tokens | $4.25 per 1 million tokens |
| Cached input | $0.15 per 1 million tokens |
| Web Search Grounding | $2.50 per 1,000 queries |
| New account credits | $20 free credits |
That is aggressive for a near-frontier agent model. The important comparison is output pricing: $4.25 per million output tokens is far below the flagship pricing used by several high-end models from OpenAI and Anthropic.
The practical caveat is token efficiency. A cheaper model is not always cheaper per finished task if it burns more tokens, needs more retries, or requires more scaffolding. But if Muse Spark 1.1 performs close to Meta's benchmark claims in real workflows, the API price will put pressure on other premium model providers.
Benchmark claims
Meta's own evaluation report and launch materials position Muse Spark 1.1 as competitive with frontier peers, especially on agent, coding, and multimodal work.
The most notable benchmark numbers being reported from Meta's launch materials include:
| Benchmark | Muse Spark 1.1 result | What it measures |
|---|---|---|
| MCP Atlas | 88.1 | Multi-step tool use across MCP servers and tools |
| Humanity's Last Exam with tools | 62.1 | Hard reasoning tasks with tool access |
| SWE-Bench Pro | 61.5 | Agentic coding tasks in complex repositories |
| Finance Agent v2 | Reported as a category lead | Financial-analysis agent tasks |
The Decoder, summarizing Meta's benchmark chart, reports that Muse Spark 1.1 leads four of twelve displayed benchmarks: MCP Atlas, JobBench, Humanity's Last Exam, and Finance Agent v2. It also notes that Claude Opus 4.8 leads five of the twelve and GPT-5.5 leads three, so Meta is not claiming a clean sweep.
For coding specifically, the picture is competitive rather than dominant. Muse Spark 1.1's reported 61.5 on SWE-Bench Pro trails Claude Opus 4.8's reported 69.2, but sits ahead of several other frontier comparisons in Meta's chart. Meta also says the model performs strongly in agentic coding setups that include planning, goal conditioning, subagent delegation, and context compaction.
The strongest story is agentic work, not pure benchmark bragging. Meta is emphasizing tasks where the model has to use tools, manage long context, interact with computers, and coordinate multiple steps rather than simply answer a prompt.
Availability
Muse Spark 1.1 is available now inside Meta AI's Thinking mode and on meta.ai. The Meta Model API is in public preview, with The Verge reporting that API access is available to US developers at launch.
Meta's timing is notable. Muse Spark 1.1 follows the launch of Muse Image, Meta's new image-generation model, and continues a broader shift away from the old Llama-only identity. Muse Spark 1.1 is proprietary and API-accessible, not an open-weight model.
That matters because Meta is trying to monetize AI more directly. Llama helped Meta gain developer mindshare. Muse Spark 1.1 is aimed at paid production workloads.
Why it matters
Meta has three advantages that make this launch worth watching.
First, it can subsidize AI pricing with a massive advertising business. That gives Meta room to make the API cheap while it fights for developer adoption.
Second, Meta owns consumer surfaces where the Muse family can be distributed quickly: Meta AI, Instagram, WhatsApp, Facebook, Messenger, and smart glasses.
Third, the model is being built around agentic workflows from the start. If Muse Spark 1.1 becomes good enough for coding agents, business workflows, and multimodal computer use, Meta can compete in API infrastructure without needing to beat every model on every benchmark.
The risk is trust. Developers will want independent benchmarks, stable API behavior, predictable rate limits, enterprise controls, and clear data policies before moving serious workloads. Meta's launch benchmarks are useful, but production results will matter more.
Our take
Muse Spark 1.1 is Meta's clearest move yet into the paid model API market. The model is positioned for the right category: not generic chat, but agentic work where tools, code, browsers, files, and long context all matter.
The pricing is the sharp edge. At $1.25 input and $4.25 output per million tokens, Meta is telling developers that agentic workloads should be cheaper to run at scale. That could be attractive for coding agents and workflow automation products where token volume grows fast.
The benchmark story is promising but still needs outside confirmation. Muse Spark 1.1 appears strongest on tool use, agent orchestration, long-context workflows, and multimodal computer use. If those strengths hold up in real apps, Meta has gone from open-model distributor to serious paid API competitor in one launch.
Sources: Meta launch post, Meta Model API, Meta Muse Spark 1.1 Evaluation Report, The Verge, The Decoder