Muse Spark 1.1: Meta's Agentic Reasoning Model and the New Meta Model API

Meta released Muse Spark 1.1 on July 9, 2026, its second model from Meta Superintelligence Labs and an upgrade from the original Muse Spark that debuted in April. The release also opened the Meta Model API to developers in public preview, putting Meta into the same paid-API business model as Anthropic and OpenAI.

The launch comes just two days after Meta rolled out Muse Image, the first image-generation model from the Superintelligence Labs.

Muse Spark 1.1 is a multimodal reasoning model built for agentic tasks, with a 1M-token context window and gains in tool use, computer use, coding, and multimodal understanding.

It runs in "Thinking" mode inside the Meta AI app and at meta.ai, and Meta is positioning it as a frontier-tier competitor to GPT-5.5, Claude Opus 4.8, and Gemini 3.1 Pro. API pricing at $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits to start.

In this article, I'll cover everything new with Muse Spark 1.1, looking at the new agentic features, exploring the benchmarks, and proposing hands-on tests for your own evaluation. For related coverage, see our writeup of Sakana Fugu vs. Claude Fable 5 and our tutorial on running MiniMax M3 locally.

TL;DR

• Muse Spark 1.1 is Meta's second Superintelligence Labs model, released July 9, 2026, built for agentic tasks with a 1M-token context window.
• It's proprietary and closed-weight, unlike the open Llama family, with free consumer access at meta.ai and a new paid Meta Model API in public preview.
• API pricing is $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits.
• Strengths: Focus on agentic tasks, with gains in tool and computer use, coding, and multimodal understanding  
• Weakness: Long-horizon agentic work is still weak compared to GPT-5.5 and Opus 4.8.  

What is Muse Spark 1.1?

Muse Spark 1.1 is Meta's latest multimodal reasoning model, designed to plan and orchestrate work across external apps and services rather than just answer single-turn questions. It sits above the open-weight Llama family, Meta's previous major line before the Muse Spark models arrived.

Unlike Llama, Muse Spark is proprietary and closed-weight, with no local deployment or community fine-tuning.

The headline change from the original Muse Spark is agentic capability.

Meta says 1.1 tackles complex projects faster because it is trained to orchestrate multi-agent systems, acting as a main agent that plans and delegates, or as a subagent that sticks to its job and escalates when needed.

It also actively manages its 1M-token context window, compacting earlier work while keeping the steps it needs later.

On coding, Meta reports that 1.1 significantly improves on the first model on its internal Meta Internal Coding Bench and is competitive with leading alternatives. That framing matters because independent benchmarks tell a more mixed story, which I'll get to in the benchmarks section.

What's New With Muse Spark 1.1

The upgrade centers on doing things across apps and interfaces, not just generating text. Here are the capabilities Meta emphasizes for this release.

Orchestrate multi-agent projects

Muse Spark 1.1 can run as the lead agent that gathers context, makes a plan, and delegates execution across parallel subagents.

As a subagent, it understands its available tools and knows when to hand control back to the main agent.

Meta says this parallel structure is what lets it complete complex projects faster than the original Muse Spark by reducing end-to-end latency.

For a practitioner, this means you can point it at a project like "pull last quarter's support tickets, categorize them, and draft a summary deck" and have it split the retrieval, tagging, and drafting across subagents.

Meta also claims it zero-shot generalizes to new native tools, MCP servers, and custom skills, so you don't have to fine-tune it for each integration.

Operate a computer across multiple apps

Muse Spark 1.1 handles computer-use workflows that span several applications where information changes mid-task.

Meta trained it to decide when to write a script for speed, when to click through an interface directly, and to generate batches of actions per step rather than reasoning one click at a time.

The company's own example is an agent placing a dinner-party order that notices new context and updates the order without user intervention.

A closer example for a data team might be an agent that scrapes a vendor pricing page, notices a mid-session price change, and revises a comparison spreadsheet before exporting it. 

The claim is that it maintains context across extended sessions and adapts to shifting requirements with minimal human input.

Fix bugs and migrate code in large codebases

Coding is where Meta puts a lot of the marketing weight, citing gains on real-world tasks like diagnosing complex bugs, implementing features in enterprise systems, and running large code migrations.

Muse Spark 1.1 supports common agentic-coding features, including planning mode, goal conditioning, subagent delegation, and context compaction, and Meta says it adapts to diverse harnesses like OpenCode.

Early partners back this up. Saoud Rizwan, CEO of Cline, said Meta is "clearly building for serious agentic coding, strong tool use at a price point that makes it viable to run real coding workloads at scale." Amjad Masad, CEO of Replit, praised its "top-tier coding abilities (particularly frontend and design)" and OpenAI-compatible packaging.

Source

Combine perception and action in one workflow

Muse Spark 1.1's multimodal strengths cover perception, visual-to-code generation, ultra-descriptive image and video captioning, and multimodal agentic execution. What Meta stresses here is doing perception and action together: inspecting visual or audio input, holding those details across a long workflow, and using them while operating a computer for you.

Meta's demonstration used smartphone video to extract product photos and reason about an item, then operate a browser to create a Facebook Marketplace listing on the user's behalf. Note the modality constraint that several third-party guides flag: inputs cover text, images, and audio, but outputs are currently text-only, with no direct image or video generation from this model.

Muse Spark 1.1 Benchmarks

Meta's official post leans on internal evaluations like Meta Internal Coding Bench, which are not independently reproducible.

The most useful public numbers come from third-party comparisons, and they show Muse Spark 1.1 trading strong token efficiency for a gap on the hardest coding tasks.

Terminal-Bench 2.1 (agentic terminal coding)

Terminal-Bench 2.1 measures how well a model handles multi-step coding and agent tasks in a terminal.

Meta posted a score of 80.0 for Muse Spark 1.1, slightly behind the 82.7 for Claude Opus 4.8 and the leading 83.4 of GPT-5.5. 

MCP Atlas (scaled tool use)

This index measures scaled tool use that requires planning and orchestration across a range of external apps and services. 

Muse Spark 1.1 takes the lead here, with a score of 88.1 vs 82.2 for Opus 4.8 and 75.3 for GPT-5.5. Meta's new model also tops the JobBench benchmarks, which focuses on professional tool use, showing that agentic tool use is very much the priority for Muse Spark 1.1 

Evaluation awareness (Apollo Research)

Apollo Research's evaluation-awareness metric measures whether a model appears to recognize it is being tested. Third-party summaries repeat the claim that Muse Spark showed the highest evaluation-awareness rates among the models evaluated. Take this one with caution: it has not been widely reproduced independently, and some commentators worry it could be over-spun without more replication.

Muse Spark 1.1 Pricing and Availability

Consumer access is free. Muse Spark 1.1 runs in Thinking mode inside the Meta AI app and at meta.ai, with a Meta login required and likely server-side rate limits for heavy use. It accepts text, image, and audio input but returns text-only output.

For developers, the new Meta Model API is in public preview. US developers can now access it to test prompts, compare outputs, and prototype integrations, with pricing set at:

• $1.25 per million input tokens
• $4.25 per million output tokens
• $20 in free credits before switching to pay-as-you-go

Source

Reuters positions that pricing above OpenAI's entry-level GPT-5 mini and Anthropic's Claude Haiku 4.5, but below Anthropic's higher-end Claude Sonnet 4.8.

The model is expected to replace existing Llama models powering chatbots on WhatsApp, Instagram, Facebook, and Meta's smart glasses. it's worth noting that several independent guides note the API is documented sparsely, with no detailed model card and no official, precise confirmation of some core specs.

Final Thoughts

The clearest signal from this launch is that Meta has switched business models. By charging per token through the Meta Model API, Meta is now competing on Anthropic and OpenAI's turf, not just shipping open weights the way it did with Llama.

That is a big shift for a company whose reputation in AI was built on openness.

My read is that Muse Spark 1.1 is genuinely interesting for two reasons: free frontier-level consumer access at meta.ai, and strong token efficiency.

The bigger caveat for practitioners is transparency. There is no open-weight release, so if your workflow needs local deployment or fine-tuning, Muse Spark is not the model for you, and the open Llama line or an open-weight alternative like MiniMax M3 remains the better fit.

If you want to build the agentic and multi-tool workflows Muse Spark 1.1 is designed for, I recommend our Developing AI Systems with the OpenAI API course, since the Meta Model API ships OpenAI-compatible.