Skip to main content
Product2026-09-03•6 min read

Muse Spark 1.3 Is Now on Multos — Meta's Agentic Workhorse, 20% Fewer Tool Calls

Zuckerberg called it Meta's biggest jump yet in coding and agentic work. Spark 1.3 uses roughly 20% fewer tool calls than 1.2, carries a 1M context, and powers Muse Code. Live on Multos Starter today.

O
Oghenekaro Oboido

Muse Spark 1.3 is the model behind Meta's Muse Code agent, and it is on Multos Starter and above. It costs $1.25 per million input and $4.25 per million output, carries a 1M context window, and uses roughly 20% fewer tool calls and 25% fewer tokens than Spark 1.2 to finish the same work. That last number is the one that matters when you are building an agent.

Fewer Tool Calls Is the Headline

An agent's cost is not mostly the prompt. It is the loop: read a file, run a command, check the result, decide what next, repeat. Every one of those steps is a round trip with tokens attached. A model that needs 20% fewer tool calls to reach the same outcome is 20% cheaper and meaningfully faster, and it is less likely to wander off. For anyone building something that runs unattended, that is a bigger deal than a benchmark point. It is also the difference between an agent that finishes inside a timeout and one that does not.

The Range of Builds It Fits

Spark is a general agentic model, so it suits a broad spread rather than a niche. A multi-file refactor tool. A code review bot. A test-writing agent. A deployment assistant that reads config and fixes drift. A documentation generator that walks a repo. A migration helper. A scaffolding tool that stands up a new service from a spec. A bug triage app that reads an issue and points at the likely file. These are builds where the model is doing structured work against a codebase rather than open-ended reasoning, which is exactly where Spark's tool-call efficiency pays off.

What Improved Over 1.2

The 1.3 release is a developer-oriented upgrade rather than a general capability jump. Meta put substantially more compute into training on programming and agentic tasks, and the result is better long-horizon collaboration, better multitasking, and better instruction following across a long session. Zuckerberg described it as Meta's biggest jump yet in coding and agentic work, which is the kind of claim to discount, but the tool-call and token reductions are measurable rather than rhetorical. If you tried Spark 1.2 and found it competent but wasteful, 1.3 addresses the wasteful part.

Where It Sits in the Meta Lineup

Meta now has three models in this catalog and they are aimed at different things. Ember is the cheap worker for bounded, repetitive steps. Spark is the agentic workhorse for structured work against a codebase. Wonder is the frontier model for long-running projects that need memory and sustained focus. Spark is the middle of that range and the one most builds should start with. It is capable enough for real agent work and cheap enough that iterating on your agent's loop does not feel like burning money.

What It Costs on Multos

Spark is a standard-tier model on Multos, available on Starter, Pro, and Elite. At $1.25 input, $4.25 output, and $0.15 cached, it bills at the standard rate for your plan: a 0.7x multiplier on Starter, 0.9x on Pro, and 1.04x on Elite. That tiering is deliberate. Starter gets the cheapest standard rate because that is the plan where budget is tightest, and the multiplier rises with the larger token pools on higher plans. In practice Spark is one of the better value models on Starter, where a 1M-context agentic model at 0.7x is unusual.

When to Pick It

Pick Spark when you are building an agent that works against a codebase and you care about how many round trips it takes to finish. It is the right default for refactor tools, review bots, test generators, and deployment assistants. Reach for Ember instead when the work is small and repetitive and you want the cheapest possible step. Reach for Wonder when the build spans weeks and needs to remember what happened last time. Spark is the one in the middle, and for most agentic builds on Multos it is where you should start.

Ready to Get Started?

Muse Spark 1.3 is live on Multos Starter and above. 1M context, 20% fewer tool calls than 1.2, and the model behind Meta's Muse Code agent. Open the model dropdown, find it under Meta, and point it at the agent you have been meaning to build.

Start Building Free