How much does Muse Spark 1.3 cost?

Meta charges $1.25 per million input tokens, $0.15 per million cached input tokens, and $4.25 per million output tokens for muse-spark-1.3 on the standard Meta Model API tier. A separate muse-spark-1.3-contributor tier costs $0.10 input, $0.002 cached and $0.20 output per million tokens, in exchange for letting Meta use your data to improve its products.

Is Muse Spark 1.3 open source?

No. Muse Spark 1.3 is proprietary and closed-weight, reachable only through Muse Code and the Meta Model API. Meta does say an open-weights Muse Spark release is on its roadmap, alongside bigger models, but has not given a date or said which model in the family would be published.

What changed between Muse Spark 1.2 and Muse Spark 1.3?

In side-by-side work by Meta engineers, Muse Spark 1.3 was significantly faster and more efficient than Muse Spark 1.2, using about 20% fewer tool calls and about 25% fewer tokens. Meta also describes better multitasking inside a single long thread, more reliable following of long-form instructions, and a clearer sense of what the model can and cannot do.

Can I use max reasoning with Muse Spark 1.3 today?

Not yet. Meta says the reasoning modes that were already available carry over to Muse Spark 1.3 from day one, but max reasoning arrives shortly after Meta finishes extra safety testing. Everything else — Muse Code and the Meta Model API — is live as of the 2 September 2026 announcement.

Is Muse Spark 1.3 safer to run unattended than earlier versions?

Meta reports three safety improvements in Muse Spark 1.3 aimed at agent use: stronger resistance to adversarial inputs, better resistance to prompt injection, and better judgment about which actions are irreversible so the model pauses instead of proceeding. Meta frames these as better discretion in long-horizon agentic tasks rather than a formal safety guarantee.

Meta · 2026-09-02 · major

Muse Spark 1.3 — Meta's flagship model uses 25% fewer tokens on coding

Muse Spark 1.3 is Meta's updated flagship reasoning model, out today in Muse Code and the Meta Model API. Meta engineers found it used about 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 on coding work.

Layered blue paths and arrows curving across a light abstract field, the Muse Spark 1.3 announcement artwork — Meta AI Research

Meta's Muse Spark 1.3 keeps the same price and 1M-token window while doing coding work with fewer tool calls and fewer tokens.

Quick facts

Maker	Meta
Version	Muse Spark 1.3
Availability	Muse Code + Meta Model API
Context window	1M tokens
License	Proprietary, closed weights
Price (input / output)	$1.25 / $4.25 per 1M tokens
Efficiency vs 1.2	~20% fewer tool calls, ~25% fewer tokens

Pricing

Input · standard tier	$1.25 / 1M tokens
Cached input · standard tier	$0.15 / 1M tokens
Output · standard tier	$4.25 / 1M tokens
Contributor input · your data is used to improve Meta's products	$0.10 / 1M tokens
Contributor output · your data is used to improve Meta's products	$0.20 / 1M tokens

source ↗

What is it?

Fewer wasted turns is the headline of Muse Spark 1.3, Meta's updated flagship reasoning model, released on 2 September 2026. Meta says it is less verbose than Muse Spark 1.2, takes fewer turns where they are not needed, and writes in a cleaner coding style. It rolled out the same day in Muse Code and the Meta Model API, at the same $1.25 / $4.25 per million tokens and the same 1M-token context window as 1.2.

How does it work?

Meta trained Muse Spark 1.3 across a range of agent harnesses so the behaviour carries over between environments. Given an open-ended goal, the model uses tools to build its own context out of messy and conflicting sources, fixes gaps in its plan as it goes, and keeps track of what it has already learned. Meta also trained it to ask clarifying questions when a prompt is ambiguous, ask the user for help when stuck, and check before doing something consequential.

Why does it matter?

Cost is the practical reason to care about Muse Spark 1.3. Because Meta bills hidden reasoning at the output rate, spending ~25% fewer tokens on the same job lowers the real bill without any price cut. The reliability changes matter for long agent runs too: Meta reports better calibration on which actions are irreversible and stronger resistance to prompt injection, both of which decide whether an agent can be left alone with a repository.

Who is it for?

developers running long agent and coding sessions

Frequently asked questions

How much does Muse Spark 1.3 cost?: Meta charges $1.25 per million input tokens, $0.15 per million cached input tokens, and $4.25 per million output tokens for muse-spark-1.3 on the standard Meta Model API tier. A separate muse-spark-1.3-contributor tier costs $0.10 input, $0.002 cached and $0.20 output per million tokens, in exchange for letting Meta use your data to improve its products.
Is Muse Spark 1.3 open source?: No. Muse Spark 1.3 is proprietary and closed-weight, reachable only through Muse Code and the Meta Model API. Meta does say an open-weights Muse Spark release is on its roadmap, alongside bigger models, but has not given a date or said which model in the family would be published.
What changed between Muse Spark 1.2 and Muse Spark 1.3?: In side-by-side work by Meta engineers, Muse Spark 1.3 was significantly faster and more efficient than Muse Spark 1.2, using about 20% fewer tool calls and about 25% fewer tokens. Meta also describes better multitasking inside a single long thread, more reliable following of long-form instructions, and a clearer sense of what the model can and cannot do.
Can I use max reasoning with Muse Spark 1.3 today?: Not yet. Meta says the reasoning modes that were already available carry over to Muse Spark 1.3 from day one, but max reasoning arrives shortly after Meta finishes extra safety testing. Everything else — Muse Code and the Meta Model API — is live as of the 2 September 2026 announcement.
Is Muse Spark 1.3 safer to run unattended than earlier versions?: Meta reports three safety improvements in Muse Spark 1.3 aimed at agent use: stronger resistance to adversarial inputs, better resistance to prompt injection, and better judgment about which actions are irreversible so the model pauses instead of proceeding. Meta frames these as better discretion in long-horizon agentic tasks rather than a formal safety guarantee.

Try it

Model id muse-spark-1.3 on the Meta Model API