Releases

Meta's Muse Spark 1.3: what's new, what it costs, and the catch

Meta released Muse Spark 1.3 on 2 September with better coding and agent skills at the same price. Here is what changed, the cheap tier that comes with a data trade-off, and who should try it.

Lines of HTML and JavaScript code on a dark screen in a code editor.

Muse Spark 1.3

2026-09-02

Meta's latest model, with stronger coding and agent skills at the same price as version 1.2.

  • Better at coding and multi-step agent tasks
  • Same price as Muse Spark 1.2
  • Available through Muse Code and the Meta Model API
  • Cheaper contributor tier, in exchange for letting Meta use your data

Source

Meta released Muse Spark 1.3 on 2 September 2026. It is the newest version of the model behind Meta AI, and Meta says it is better at coding and agentic work (tasks where the AI works through several steps on its own). The price did not change.

Meta has been shipping fast. Muse Spark 1.1 arrived in July and 1.2 in August, so 1.3 is the third version in under two months. If you are choosing an AI model for your business, here is what is different this time.

What's new in 1.3

Meta's own claims, as reported by Yahoo Tech:

  • Better at long, multi-step work. It follows complex instructions more closely and produces cleaner, less wordy code.
  • Less waste. Meta says it used about 20% fewer tool calls and about 25% fewer tokens (the small chunks of text AI models are billed by) than before.
  • Asks before it acts. It is better at asking clarifying questions and at checking with you before big or irreversible changes.
  • Tougher against attacks. Meta reports improved resistance to attempts to trick the model.

Meta AI chief Alexandr Wang called it very competitive with frontier models and said it will support personal agents that work around the clock. That is a company claim, not an independent test.

Where you can use it

Muse Spark 1.3 is available through Muse Code, Meta's coding agent for the terminal, and through Meta's Model API. The API is still in limited preview, so not everyone can sign up yet. The model is proprietary: there are no downloadable weights.

A "max reasoning" version is planned but is waiting on more safety testing.

Prices

According to OrcaRouter's model guide, per million tokens:

TierInputOutput
Standard$1.25$4.25
Contributor$0.10$0.20

The model can read up to about one million tokens in one go (roughly several full-length books), and accepts text, images, video, audio and PDFs, with text-only output.

The contributor tier trade-off

The contributor tier is far cheaper, but you pay with your data. Meta uses your prompts and responses to improve its models on that tier. On the standard tier, Meta does not retain what you submit.

Limits and caveats

  • Audio is not fully working in 1.3. OrcaRouter says Meta recommends staying on version 1.2 for audio jobs.
  • Limited API access means you may be on a waiting list.
  • Benchmarks are Meta's own. Run your own tests before you switch.
  • The model changes quickly. A new version every month or so means you should expect prompts to behave a little differently each time.

Related Meta releases

Meta also released Muse Code (5 August), a terminal coding agent that can run several sub-agents at once, and Muse Glimmer (10 August), a 30-billion-parameter model with open weights under the Apache 2.0 licence. Meta says Glimmer can run fully offline on one 24 GB consumer graphics card. That one is aimed at developers who want to host a model themselves.

Who should try it

  • Developers and agencies building tools or automations can test Muse Spark 1.3 against the model they use now. Cost per task is the number to compare.
  • Teams working with sensitive data should stay on the standard tier, or choose a different provider.
  • Everyday users will meet Muse Spark inside Meta AI on Meta's apps without needing to do anything.

Sources