Published: 3 September 2026
Meta released Muse Spark 1.3 on 2 September 2026 and began rolling it out through Muse Code and the Meta Model API the same day. The new checkpoint targets longer agentic workflows and coding tasks, with Meta reporting fewer tool calls and lower token use than Muse Spark 1.2.
For software teams already evaluating Muse Code, the practical change is a newer hosted model without a migration to a different agent product. Open weights are still planned for a later release.
What is Muse Spark 1.3?
Muse Spark 1.3 is Meta’s closed-weights reasoning model for long agent runs and coding. It reads text, images and video, and holds a million tokens of context. Meta trained it to behave more carefully: it asks clarifying questions, calls for help when stuck, and confirms before doing anything it cannot undo.
- Where to get it: Muse Code and the Meta Model API.
- Standard price: $1.25 per million input tokens, $4.25 per million output.
- Contributor price: roughly $0.10 and $0.20, if Meta may train on your traffic.
- Context window: 1 million tokens.
- Weights: closed. No open-weights release was announced for 1.3.
- Reasoning modes: xhigh ships today, max is held back for safety testing.
Muse Spark 1.3 benchmarks
Read the rows rather than the headline. Meta wins clearly on coding and long context, and finishes second on most agent tasks.
| What is measured | Muse Spark 1.3 | Claude Opus 5 |
|---|---|---|
| DeepSWE v1.1 (agentic coding) | 75.4 | 74.0 |
| SWEAtlas CodeBase QnA (reading a codebase) | 59.4 | 52.7 |
| MRCR 512K to 1M (recall in very long input) | 98.1 | no figure published |
| GDPval-AA v2 (knowledge work, Elo) | 1754 | 1824 |
| OSWorld 2.0 (operating a computer) | 66.9 | 68.3 |
| AutomationBench (end-to-end workflows) | 49.4 | 50.3 |
These are Meta’s own numbers from its launch scorecard. They were produced with the max reasoning mode, which has not shipped. Artificial Analysis rates the available xhigh mode at 61 on its intelligence index, one point below max.
How much does Muse Spark 1.3 cost?
Meta sells the same model through two endpoints, and the difference is your data. The standard endpoint keeps your traffic private and is priced like a frontier model. The contributor endpoint costs roughly a tenth as much, and in exchange Meta uses what you send to improve its products. No rival offers that trade so openly.
That is a sharper choice than it looks. You are not really buying tokens at two prices, you are paying a premium for Meta not to keep your source code and customer conversations. For a hobby project the discount is close to free money. For a regulated business it is a compliance question, not a procurement one.
What Meta is not saying
The scorecard compares Muse Spark 1.3 at max reasoning against Muse Spark 1.2 at xhigh. Those are different effort settings, so part of the jump between versions is not a jump at all. The OSWorld gain looks dramatic, and a like-for-like run would narrow it.
The bigger cost is verbosity. Artificial Analysis measured 120 million output tokens to finish its full test suite, against a field median of 72 million. Community testing puts the model at roughly triple the token use of Muse Spark 1.2. A low rate card does not survive that.
What this means
Test this now if you feed models very large inputs. Nothing else comes close on recall across a full million-token window, and Claude Opus 5 did not even post a number. A team indexing an entire codebase, or lawyers reading a complete case file in one pass, gets a real new capability here rather than a discount.
The contributor endpoint deserves a firm no for anything sensitive. It is far cheaper precisely because Meta trains on what you send, so it belongs on side projects, not on customer data or private source code. Also wait on the headline scores. The mode that earned them is not something you can call today.
For more information, visit the official announcement of Muse Spark 1.3 on the Meta blog.
