Meta rarely moves fast in public. That changed this month. Muse Spark 1.1 launched on July 9, 2026, and it arrived with real weight behind it. Mark Zuckerberg posted about the release on X, his first post on the platform in three years. That alone told the industry something had shifted inside Meta’s AI labs. This is not a small patch. It is Meta’s second model from its Superintelligence Labs team, and it lands with real gains in coding, tool use, and agentic work.
This blog walks through what makes Muse Spark 1.1 worth the attention. We will cover how it handles agent tasks, what changed in coding and tool use, how its pricing stacks up against rivals, and where it still falls short. We will also look at why this New Meta AI model matters for the wider AI market and how Meta Muse Spark fits into Meta’s bigger plan for the rest of 2026.
A Big Week for Meta’s AI Push
This launch did not happen in isolation. Just days earlier, Meta released Muse Image, a separate model built for image generation. Alexandr Wang, Meta’s chief AI officer, called this New Meta AI model the lab’s strongest yet for agentic and coding work.
A few things line up around this launch:
- Meta shipped two major AI releases within the same week.
- Rival labs shipped their own major updates during the same stretch.
- The pileup shows just how fast competition has grown among major AI labs.
Meta is racing to close a gap that rivals opened months earlier, and this release is the clearest sign yet of how seriously the company is treating that gap.
What Is Muse Spark 1.1?
Muse Spark 1.1 is Meta’s newest multimodal reasoning model, built for agentic tasks. It follows the original Muse Spark model, internally called Avocado, which shipped in April under a closed partner program with no public API. This new version opens that access up in a real way, marking Meta’s first serious push to put a frontier-class model directly in developers’ hands rather than keeping it locked behind a limited partner program.
A few core facts define this release. It carries a 1-million-token context window, large enough to hold entire codebases or long conversation histories without losing earlier context. It runs in “Thinking” mode inside the Meta AI app and at meta.ai, and it is closed-weight and proprietary, unlike Meta’s open Llama family, a real shift in strategy for a company long associated with open-source AI. Furthermore, it also ships with a new Meta Model API in public preview, giving outside developers a real path to build on top of it.
Meta positions this New Meta AI model as a frontier-tier competitor to GPT-5.5, Claude Opus 4.8, and Gemini 3.1 Pro. That is a bold claim for a company that trailed rivals in this space for much of the past two years, and it signals just how much ground Meta believes it has closed with this single release.
How This New Meta AI Model Handles Agent Work
Agentic work means more than answering a single question well. It means planning, using outside tools, and managing several steps toward one goal. Muse Spark 1.1 was built around this exact kind of task, with a structure designed to hold up under real, multi-step projects rather than short, isolated requests.
The model can operate in two distinct roles. As the main agent, it gathers context, builds a plan, and delegates work across parallel subagents, acting more like a project lead than a single worker handling every step alone. As a subagent, it sticks to its assigned job, understands its available tools, and knows when to escalate a problem back to the main agent, rather than pushing forward blindly on a task it cannot fully solve on its own.
This structure lets Muse Spark 1.1 tackle complex projects faster than its predecessor. It splits large jobs into smaller pieces that run at the same time instead of one after another, cutting down on total wait time. It also zero-shot generalizes to new native tools, MCP servers, and custom skills, meaning it does not need special training to pick up a new tool it has never seen before, a real advantage for teams that add new tools to their workflow often.
Coding and Tool Use: The Real Upgrade
Coding sits at the centre of this release. Meta says the model handles real engineering work, not toy demos:
- It fixes bugs across real codebases, not simplified sample files built for a demo.
- It manages large code migrations that would normally take a team several days.
- It deploys new features inside real enterprise systems, where mistakes carry a real cost.
- It handles this work at a scale engineering teams face every day, not just in controlled tests.
- It aims to reduce the manual review time engineers spend on routine, repetitive coding tasks.
It also highlights a few specific gains in this update:
- Stronger multistep reasoning across long coding tasks so the model holds its plan together over many steps.
- Better handling of digital workflows that touch several systems at once, cutting down on handoff errors.
- Improved tool calling so the model picks the right tool for a given step more reliably.
- Smoother multimodal understanding, blending text, image, and video context in one session.
- More consistent output across repeated runs of the same task, which matters for teams that automate coding work.
Meta describes this combination as exceptional performance in personal agentic tasks that require planning and orchestration across a range of external apps and services. These claims sound strong on paper, but real adoption will come down to daily use, not launch-day marketing. A few things will decide if this holds up:
- How the model handles messy, real production codebases, full of legacy code and inconsistent style.
- Whether Meta Muse Spark keeps its edge outside clean demo scenarios built to show its best side.
- How developers rate it after weeks of daily use, not just launch week, once the initial excitement fades.
Pricing That Undercuts Rivals
Price is where this release gets genuinely aggressive. This marks the first time Meta has charged businesses for access to one of its AI models, and the numbers came in lower than most expected.
Here is how the pricing breaks down:
- 1 dollar 25 cents per million input tokens.
- 4 dollars 25 cents per million output tokens.
- 20 dollars in free credits for new developer accounts.
That pricing sits at roughly a quarter of what OpenAI and Anthropic charge for their top-tier models, according to Meta’s own framing. Zuckerberg described the strategy in a few clear points:
- This is not an open-source model, so it needed a real, serious paid API.
- The pricing is meant to be aggressive and attractive from day one.
- This marks a real shift in how Meta Muse Spark generates revenue for the company.
Meta Muse Spark vs the Competition
Meta claims Meta Muse Spark beats Google’s Gemini 3.1 Pro on several benchmarks and shows real advantages over models from Anthropic and OpenAI, including speed and multimodal reasoning. These are strong claims, and they deserve scrutiny rather than blind acceptance.
A few things are worth noting about these comparisons:
- Meta did not include comparisons against the very newest releases from some rivals.
- Meta has faced past accusations of training on test sets for a prior model, which it has denied.
- Independent, third-party testing will matter more than any single company’s own benchmark chart.
None of this means the claims are false. It does mean a careful developer should run their own tests before betting production on any single benchmark result.
Where This Model Still Falls Short
No launch is without weak spots, and Meta has been fairly open about at least one of them:
- Long-horizon agentic work still lags behind top rivals like GPT-5.5 and Claude Opus 4.8.
- This gap spans many steps over a long stretch of time, not quick single tasks.
- A team pushing Muse Spark 1.1 through a long, sprawling project may hit rougher edges sooner than expected.
This gap matters most for complex, multi-day projects rather than quick coding tasks. A team leaning on Muse Spark 1.1 for short bursts of agent work may see strong results instead.
Why Choose Us
Picking a strong model is only the first step. Knowing how to route tasks, test claims, and catch weak spots before they reach production is the real work. This is where we help:
- Not a job board – we are a curated creative network built for serious professionals.
- Home to the world’s best designers, directors, and creative technologists.
- We cover the latest tools, including the Muse Spark 1.1, so your skills always stay relevant.
- We connect you with work that actually matches your craft and ambition.
- Every feature on our platform exists to help serious creative professionals grow and move forward
Final Thoughts
Meta has spent the past two years chasing rivals in the AI race, and Muse Spark 1.1 is the clearest sign yet of how seriously that chase has become. Strong gains in coding, tool use, and agentic reasoning back up Zuckerberg’s rare public post, and the aggressive pricing shows a company willing to compete hard on cost, not just capability.
The model is not perfect. Long-horizon agentic work still trails top rivals, and Meta’s own benchmark claims deserve independent testing before anyone treats them as settled fact. Even with those caveats, this release marks a real turning point for Meta’s place in the wider AI market, and it is worth close attention as more releases follow through the rest of 2026. Want to apply or have a query? Reach out to Working Not Working on WhatsApp and follow us on LinkedIn and Facebook.
FAQs
1. What is Muse Spark 1.1?
Muse Spark 1.1 is Meta’s newest multimodal reasoning model, built for agentic tasks, coding, tool use, and computer use, with a 1 million token context window. It stands as Meta’s boldest New Meta AI model to date.
2. Is Muse Spark 1.1 free to use?
It is free inside the Meta AI app and at meta.ai in “Thinking” mode. A paid developer API for Meta Muse Spark is also available, with 20 dollars in free credits for new accounts.
3. How much does the Muse Spark 1.1 API cost?
The Muse Spark 1.1 API costs 1 dollar 25 cents per million input tokens and 4 dollars 25 cents per million output tokens, roughly a quarter of top rival pricing according to Meta.
4. Is this a New Meta AI model built on Llama?
No. Meta Muse Spark sits above the open-weight Llama family and is closed-weight and proprietary, marking a real shift from Meta’s earlier open-source approach.
5. What is the biggest weakness of Muse Spark 1.1?
Long-horizon agentic work, tasks that span many steps over a long stretch, still lags behind top rivals like GPT-5.5 and Claude Opus 4.8. Even so, this New Meta AI model holds up well on shorter, focused Muse Spark 1.1 tasks.
