Meta released a new AI model this week, and both halves of the story matter. Muse Spark 1.1 is stronger at the kind of multi-step work businesses actually want to automate, and it costs a fraction of what comparable models charge. For anyone weighing agentic AI for business use, that combination is the part worth paying attention to.
On July 9, Meta Superintelligence Labs launched Muse Spark 1.1, a multimodal reasoning model built for agentic tasks like tool use, computer operation, and software coding. Meta positions it as a major step up from the original Muse Spark, with real gains in how it plans multi-step work, orchestrates other AI agents, and handles large, messy codebases. Developers can access it through a public preview of the new Meta Model API, and it is already live in Thinking mode inside the Meta AI app.
Two things are worth understanding here: what the model can do, and what it costs. Both change the calculation for businesses building with AI, so we will take them in order.
What Muse Spark 1.1 Can Do
The capability gains are not cosmetic. Four of them stand out for the kind of work our clients are automating.
A one-million-token context window that it manages itself. Muse Spark 1.1 can hold roughly a million tokens at once, enough to keep an entire codebase or a long project history in view. What makes it useful is not the size but the management. The model compacts its own context as it works, keeping the actions and structure that matter for later steps and discarding the noise from early in a session. For long-running workflows that would normally lose the thread, that means fewer dropped details and less hand-holding.
Computer use and scripting. Rather than clicking through every interface by rote, Muse Spark 1.1 decides how to get a job done. When writing and running a script is the faster path, it scripts. If an interface needs direct interaction, it clicks. It can also batch actions to move through multi-app workflows more efficiently, which is the difference between an assistant that follows instructions and one that finds the shortest route.
Parallel multi-agent orchestration. The model can operate as a lead agent that plans and delegates, or as a subagent taking direction from another. Give it a broad goal, such as pulling a quarter of support tickets, categorizing them, and drafting a summary, and it can spin up and manage subagents that split the work in parallel. For teams, that turns a sequential grind into something closer to a small project crew working at once.
Visual troubleshooting and coding. Its coding strength goes past writing new code. Muse Spark 1.1 can take an automated screenshot of a broken interface, trace the visual problem back to the source code, deploy a fix, and then check its own work to confirm the change actually held. That closes a loop that usually needs a human at every step.
Capabilities like these are what make agentic workflows worth running in the first place. Whether you can afford to run them at scale depends on the second half of the story.
The Price That Changes the Math
Meta priced Muse Spark 1.1 at $1.25 per million input tokens and $4.25 per million output tokens, which puts it at roughly a tenth of the cost of comparable models from Anthropic and a quarter of comparable offerings from OpenAI. Mark Zuckerberg broke a three-year silence on his own social platform to promote the launch, calling it a highly capable agent and coding model priced well below the market, and arguing that frontier-level intelligence does not need to come with a frontier-level bill.
That framing matters more to your business than any benchmark score. Enterprise leaders have spent much of this year raising concerns about ballooning AI spend, and several major tech executives have noted that adoption stalls when token costs climb too high relative to the value delivered. Meta is betting that cutting the price of agentic AI by an order of magnitude will pull hesitant enterprises off the sidelines. Competing labs are responding in kind. OpenAI released GPT-5.6 the same week promising more output per token, and xAI shipped Grok 4.5 touting similar efficiency gains. The industry’s center of gravity is moving from which model is smartest to which model delivers the most usable intelligence per dollar.
What This Means If You’re Building With AI
For the businesses we work with, most of whom are evaluating AI not as a novelty but as infrastructure, the combination of stronger capability and lower cost changes the math on a few fronts.
Agentic workflows become financially viable at smaller scale. Multi-step agents that plan, delegate to subagents, and operate software on a person’s behalf have mostly lived in enterprise pilot budgets. Lower per-token costs, paired with orchestration that actually works, mean a mid-market company or a fast-growing startup can run these workflows continuously instead of rationing them.
The build-versus-buy calculus shifts again. As base model costs fall, the value in a custom AI product increasingly lives in the workflow design, the data integration, and the guardrails wrapped around the model, not in the model itself. That is exactly where a strategic build partner earns its place.
Vendor lock-in becomes riskier to ignore. Meta’s aggressive pricing is deliberately designed to pull developers into its own API and ecosystem. Businesses that build AI capabilities dependent on a single provider’s pricing and roadmap take on real exposure if that provider changes terms later. Custom, provider-independent AI architecture is becoming less of a preference and more of a hedge.
Caution Before Speed
Rock-bottom pricing from a company with Meta’s advertising profits to subsidize it is not the same as a stable long-term price floor for the industry. Pricing wars compress margins for everyone, including the smaller AI vendors your business might already depend on, and today’s bargain rate is not guaranteed to hold. The smarter move is treating this moment as a signal that agentic AI is getting more accessible, not as a reason to build your entire strategy around one vendor’s current price sheet.
The bigger trend underneath the headline is the one worth watching closely. The cost of building genuinely capable AI agents is falling fast, and the capability itself is climbing. The businesses that figure out how to use that shift, without becoming dependent on any single lab’s pricing decisions, will be the ones who benefit most.
If you are weighing how agentic AI fits into your product or operations, that is a conversation worth having now, while the landscape is still moving.
