Meta Muse Spark 1.1: Meta's First Paid Agent Model

On July 9, 2026, Meta did two things at once: it shipped Muse Spark 1.1, its most capable model yet for real-world coding and agentic tasks, and it started charging developers to use its own model for the first time. Both matter. The model is a genuine step up on agentic work; the paid API is Meta planting a flag in a market Anthropic and OpenAI have owned.

The honest headline is not that Muse Spark 1.1 tops every leaderboard — it does not. On pure coding and multimodal reasoning, Opus 4.8 and GPT-5.5 beat it. What it does win is a narrower, more interesting race: professional and scaled tool use — the ability to plan, call tools, and orchestrate work across apps and MCP servers. For a firm that builds agents to run real marketing tools, that is the number worth reading.

This guide maps what Meta actually shipped, the spec sheet an engineering team plans around, the benchmark standings read honestly, why tool use is the wedge that matters for agentic marketing, where Muse Spark 1.1 fits alongside Fable 5, Opus 4.8, and GPT-5.5, and the caveats to flag first — starting with the fact that the public preview is US-only.

Muse Spark 1.1 comes out of Meta Superintelligence Labs — the same group that shipped Muse Image, Meta's first image-generation model. It is a multimodal reasoning model built for agentic tasks, and Meta frames it as a clear upgrade over the original Muse Spark on tool use, computer interaction, coding, and multimodal understanding. The gap between the two is not subtle: on several agentic evals the base Muse Spark scores in the teens where 1.1 clears the fifties, so read "1.1" as a different weight class, not a point release.

The bigger story is commercial. Until now, Meta's models reached developers mainly as open weights. Muse Spark 1.1 launches on the paid Meta Model API in public preview, which means Meta is now charging for access to its own frontier model the way Anthropic and OpenAI do — a strategic pivot as much as a product launch. It is also available in the Meta AI app's "Thinking" mode and on meta.ai for consumers.

For developers, the friction is deliberately low. Meta first introduced Muse Spark in April; the public-preview release is the answer to a steady developer ask for a way to build with it. The Meta Model API is self-serve and speaks both the OpenAI SDK (Chat Completions and Responses formats) and the Anthropic Messages format, so pointing an existing agent at Muse Spark is a base-URL and key change rather than a rewrite: set the base URL to api.meta.ai/v1, pass your key, and name the model muse-spark-1.1. That dual compatibility is a quiet but important distribution choice — it lets teams A/B a new model against their current one without re-plumbing the stack. For the broader field Meta is now competing in, our Q2 2026 agentic-coding platform matrix maps twenty tools side by side.

Strip away the positioning and here is what an engineering team plans around. Per Meta's own developer blog, pay-as-you-go pricing starts at $1.25 per million input tokens and $4.25 per million output tokens, with a one-time $20 in free credits on every new account. That lands above OpenAI's entry-level small models and Claude Haiku 4.5, but below a mid-tier model like Claude Sonnet 4.6 — mid-market, not rock-bottom, but cheap for the agentic performance it posts.

Muse Spark 1.1 is a reasoning model: it thinks before it answers, those reasoning tokens are billed as output, and a reasoning_effort parameter runs from minimal to xhigh so you match depth to the task. The context window is 1,000,000 tokens, and Meta says the model actively manages and compacts that context — dropping stale detail while keeping what a long-running workflow needs to continue. It reads images, video, and documents in a single call, and exposes native tool integration, structured output, and parallel tool calling. The Model API also ships built-in web-search grounding: add a web_searchtool to a Responses API call and the model returns real-time, cited answers with no retrieval stack to build. Meta reports lower hallucination and stronger resistance to jailbreaks and prompt injection than the original Muse Spark, and evaluates the model against Chemical & Biological, Cybersecurity, and Loss-of-Control risk categories under its Advanced AI Scaling Framework.

Meta's own launch table tells a split-decision story, and the split is the point. Group the benchmarks by what they test and a clean pattern falls out: Muse Spark 1.1 leads the agent and tool-use evals, and trails on pure coding and multimodal. The table below shows the vendor-reported standings against the strongest modes of each rival; Muse Spark's column is highlighted, and the leading score in each row is bold.

Read the pattern, not the highlight. On the four agent and tool-use rows — MCP Atlas, JobBench, Humanity's Last Exam, and Finance Agent v2 (57.2, not shown above) — Muse Spark 1.1 is first. On computer use (OSWorld) and personal tool use (Toolathlon) it slips behind Opus 4.8. And across coding (SWE-Bench Pro, DeepSWE 1.1, Terminal-Bench 2.1) and multimodal (CharXiv, BabyVision) it is a solid third — competitive, but not the leader. This is a model that wins the tool-orchestration race and concedes the raw-accuracy one.

Here is why the split matters for marketing rather than for a leaderboard. The work an agentic marketing system actually does is not one-shot code generation — it is orchestration: read a CRM, query an ads API, update a spreadsheet, call a research tool, hand off to a subagent, and keep the thread straight across a long session. Those are tool-use and planning skills, and they are exactly what JobBench and MCP Atlas measure. JobBench — professional tool use — is the clearest tell.

The capability set behind that score reads like a spec for building agents. Meta reports zero-shot generalization to new tools and MCP servers — the model can pick up a tool it has never seen and use it correctly — plus multi-agent coordination with subagent delegation, native parallel tool calling, and structured output. It is even trained to choose between writing a script to automate a task and interacting with an interface directly, which is the kind of judgment that separates a demo agent from a dependable one. If you are standing up an agent stack on the Model Context Protocol, our guide to building an MCP server from scratch and the agent-protocol ecosystem map are the natural companions to a tool-use-first model.

Meta's own developer recipes show the depth in practice. In its computer-use walkthrough, Muse Spark drives a real Linux desktop from a single plain-language goal — no coordinates, no click-by-click script — by taking a screenshot, reasoning about what it sees, acting, and looking again, all inside a throwaway sandbox. In its multi-agent recipe, one model fills four roles — product manager, backend, frontend, and technical writer — that coordinate only through durable comments on a shared Kanban board, with the product manager as sole arbiter and the whole run replayable and auditable. That auditable, role-scoped orchestration is precisely the shape of a well-governed marketing agent team.

The pricing makes the wedge practical. Tool-use agents are chatty — they loop, they re-plan, they burn tokens on intermediate steps — so a model that is strong on tool use but expensive per token can be uneconomical to run at scale. At $1.25 in and $4.25 out per million, with a million-token window to hold a long session, Muse Spark 1.1 is priced to run the high-volume orchestration work rather than just demo it. For a real-world example of what tool-first agents unlock for an agency, see our MCP server rollout case study.

The move is not to crown a new default — it is to route. In a multi-model stack, Muse Spark 1.1 earns the tool-orchestration and high-volume agentic jobs, and you keep your top-accuracy models for the work where a wrong answer is expensive. The matrix below is the practical read.

The competitive context is the other half of the story. Muse Spark 1.1 lands the same week as SpaceXAI's Grok 4.5 and OpenAI's GPT-5.6 — a frontier now clearly splitting into tiers by job rather than one model to rule them all. Meta's entry is notable less for topping a chart than for where it aims: agentic tool use, at a price that makes agents economical to run. For a structured way to weigh price against performance across the field, our performance-vs-price efficient-frontier analysis is the framework we use, and Meta's enterprise business-agent push is the strategy this API extends.

Four things to weigh before you wire Muse Spark 1.1 into anything that matters. The first is regional, and for an EU-based team it is the big one.

The first caveat deserves the emphasis. Because the preview is US-only, EU-based teams cannot production-test Muse Spark 1.1 on Meta's own API yet, and the most reliable path is to wait for the official regional rollout rather than route around access controls, which can violate terms of service. Plan any serious evaluation for whenever EU access opens, and treat the pricing and benchmark numbers above as a launch-day snapshot to re-verify — not a settled spec.