Meta's Muse Spark 1.1 Opens a Public API to Challenge OpenAI and Anthropic
Meta Superintelligence Labs releases Muse Spark 1.1, a multimodal agentic model with a 1M-token context, now accessible via the new Meta Model API at $1.25/$4.25 per million tokens.
- New model: Meta Superintelligence Labs releases Muse Spark 1.1, a multimodal agentic reasoning model with a 1M-token context window.
- First public API: The Meta Model API launches in public preview for US developers, with $20 in free credits on signup.
- Pricing: $1.25 per million input tokens and $4.25 per million output tokens -- mid-tier vs. competitors.
- Benchmark highlights: Leads on JobBench (54.7 vs. Opus 4.8's 48.4) and MCP Atlas (88.1), with a massive coding leap on Vibe Code Bench (72.2 vs. 19.7 for the original).
- Key capabilities: Parallel subagent orchestration, adaptive computer use (scripting vs. UI clicks), and visual-to-code generation from video input.
- What's next: A more powerful successor model, code-named Watermelon, is currently in training at Meta.
Muse Spark 1.1 is Meta Superintelligence Labs' most capable model yet, and it comes with a milestone that the original Muse Spark never had: a public API. For the first time, developers can build on top of Meta's frontier model, putting it in direct competition with Anthropic and OpenAI in the enterprise AI market. The model is live now in "Thinking" mode on meta.ai and the Meta AI app, and the Meta Model API is in public preview for US-based developers.
A new model, a new business
Muse Spark 1.1 is the second model from Meta Superintelligence Labs (MSL), the research division led by Alexandr Wang. It represents a complete departure from Meta's Llama model family -- not Llama 5, but an entirely new model architecture designed to compete at the very top of the frontier AI landscape. Unlike the open-weight Llama models that preceded it, Muse Spark is closed source -- a significant strategic shift for Meta, which built its AI reputation on open-source releases.
Meta is now charging for use of its AI, pitting it directly against the business models of Anthropic and OpenAI. The initial Muse Spark model released earlier this year was only available to "select partners" who could access the technology via a "private API preview." The 1.1 release tears that wall down.
What's actually new
Muse Spark 1.1 is a multimodal reasoning model built for agentic tasks, with major gains in tool and computer use, coding, and multimodal understanding. The headline capability is its approach to multi-agent orchestration -- the ability to coordinate fleets of AI subagents working in parallel on a single task.
- 1M-token context window. The model can actively manage its context window of 1 million tokens, remembering actions, retrieving information from much earlier work, and compacting in a way that keeps the critical steps needed for later work.
- Dual-role agent. The defining architecture of Muse Spark 1.1 is its ability to operate natively within multi-agent networks -- it can function simultaneously as a "primary agent" and a "subagent," allowing it to break down large tasks and delegate them across parallel subagents to drastically reduce processing time.
- Adaptive computer use. Rather than reasoning through every desktop step one click at a time, Muse Spark 1.1 understands when to automate and when to use the interface directly. Meta trained the model to write scripts when automation is faster, click when direct interaction is simpler, and generate batches of actions at each step.
- Multimodal perception. The model excels in perception, multimodal reasoning, and tool use, with strengths in visual-to-code artifact generation, ultra-descriptive image and video captioning, and agentic workflow execution for multimodal use cases.
The numbers
Meta published benchmark results across several agentic and coding evaluations. The picture is strong for agentic tasks, more competitive (but not dominant) for coding.
- JobBench (agentic job completion): Muse Spark 1.1 scores 54.7, ahead of Opus 4.8 max (48.4), GPT-5.5 (38.3), and the original Muse Spark (17.0).
- MCP Atlas (tool orchestration): 88.1, edging out Opus 4.8 (82.2) and Gemini 3.1 Pro (78.2).
- Vibe Code Bench v1.1 (end-to-end coding): 72.2 vs. the original Muse Spark's 19.7 -- a massive leap.
- SWE Atlas Codebase QnA: 42.0 vs. 24.2 for the original model.
- DeepSearchQA: 84.9, just behind GPT-5.5 (87.8) and ahead of Opus 4.8 (84.3).
On Meta's primary internal coding evaluation (Meta Internal Coding Bench), Muse Spark 1.1 significantly improves upon Muse Spark and is competitive with leading alternatives. Researchers are now also automating model development and evaluation tasks by leveraging Muse Spark 1.1 in their workflows.
How it's built
Meta tracks Muse Spark's scaling properties along three axes: pretraining, reinforcement learning, and test-time reasoning. The pretraining phase is where Muse Spark acquires its core multimodal understanding, reasoning, and coding abilities. Over the last nine months, Meta rebuilt its pretraining stack with improvements to model architecture, optimization, and data curation -- advancements that increase the capability extracted from every unit of compute.
MSL trained Muse Spark 1.1 to excel in coding-related tasks because that ultimately improves the capabilities of AI agents. "You kind of have to build coding capabilities as part of that in service of overall agentic capabilities," said Wang. Meta trained Muse Spark 1.1 "to be able to work well with all of the most popular harnesses that developers use today," with the goal of maximizing adoption.
A key safety note: across all frontier risk categories -- Chemical and Biological, Cybersecurity, and Loss of Control -- Meta's evaluations show Muse Spark 1.1 operates within safe margins, demonstrating strong resistance to direct jailbreaks, indirect attacks from untrusted data, prompt injection, and developer-prompt attacks.
Pricing and access
Developers in the United States can now access Muse Spark in public preview on the Meta Model API, letting them test prompts, compare outputs, and prototype integrations. Those who sign up receive $20 in free credits before switching to pay-as-you-go pricing.
The API is priced at $1.25 per million input tokens and $4.25 per million output tokens -- above OpenAI's entry-level GPT-5 mini and Anthropic's Claude Haiku 4.5, but below Anthropic's higher-end Claude Sonnet 4.6. For now, Meta is limiting API access to its own properties rather than making it available on third-party platforms like the popular OpenRouter marketplace.
Wang confirmed that Meta is currently training a more powerful AI model, code-named Watermelon, but did not say when it would be released. Muse Spark 1.1's own codename was Avocado.
Who's already using it
A trio of prominent early API partners -- Replit, Cline, and Box -- headlined the announcement, signaling strong early adoption in the developer tooling space. Their feedback points to a consistent theme: the combination of a million-token context, multimodal support, and competitive pricing is what makes this model compelling for production agentic workloads -- not any single capability in isolation.
Muse Spark 1.1 is also expected to replace existing Llama models powering chatbots on WhatsApp, Instagram, Facebook, and Meta's collection of smart glasses. That deployment scale, combined with the new paid API, signals that Meta is serious about turning its AI investment into a revenue line -- not just a research showcase.