Muse Spark 1.3: what changed, what it costs, and where to use it
Meta’s September model update replaced Muse Spark 1.2 inside Muse Code without raising prices. The headline benchmarks come with a caveat — and the efficiency gains may matter more to your bill.
What shipped on September 2
Meta released Muse Spark 1.3 on September 2, 2026, four weeks after 1.2. It rolled out the same day as the model behind Muse Code and on the Meta Model API, with paid developer access opening on September 3. On the API it comes in the two familiar variants: muse-spark-1.3, whose data is never used to improve Meta’s products, and muse-spark-1.3-contributor, which is far cheaper because it is. The context window stays at one million tokens, and the model takes text, images, and video as input and returns text.
If you use Muse Code, you don’t need to do anything: the agent moved to 1.3 on its own.
What actually changed
Meta frames 1.3 around real-world usability rather than raw capability:
- Longer threads. It sustains longer-horizon work across multiple workflows in a single thread — a good fit for the Workflows feature Muse Code gained when it left beta.
- It asks instead of guessing. It asks clarifying questions, flags when it’s stuck, and is better calibrated about its own limits instead of reporting outcomes it didn’t reach.
- Leaner runs. In Meta engineers’ comparisons, it finished the same coding work with about 20% fewer tool calls and 25% fewer tokens than 1.2.
That last point is the one your invoice will notice. Prices didn’t move, so fewer tokens per task is a straight discount: if the reduction holds on your workload, a job that cost $1.00 in tokens on 1.2 lands closer to $0.75 on 1.3.
The benchmark caveat
Meta’s launch scorecard is strong — but its best numbers come from a max reasoning configuration that was still finishing safety testing at launch and available only to a limited set of partners. For that configuration, Meta reported:
| Benchmark | Muse Spark 1.3 (max) |
|---|---|
| DeepSWE 1.1 | 75.4% |
| Terminal-Bench 2.1 | 88.8% |
| SWEAtlas CodeBase QnA | 59.4% |
| Long-context retrieval (MRCR, 256K–512K) | 98.5% |
On DeepSWE, that put it ahead of GPT-5.6 Sol (73.0) and Claude Opus 5 (74.0) in DataCamp’s summary of the scorecard. The configuration most developers get at launch scores lower. VentureBeat’s comparison put the shipping model at 61 on the Artificial Analysis Intelligence Index — level with GPT-5.6 Sol (max) and Claude Opus 5 (high) — and showed gaps to max on agent tasks such as OSWorld 2.0 (57.2 vs. 66.9) and JobBench (61.2 vs. 64.9). Treat the headline table as a ceiling, not as what your first session will feel like.
What it costs
Nothing changed on the price sheet:
| Per 1M tokens | Standard (muse-spark-1.3) |
Contributor (muse-spark-1.3-contributor) |
|---|---|---|
| Input | $1.25 | $0.10 |
| Cached input | $0.15 | $0.002 |
| Output | $4.25 | $0.20 |
Inside Muse Code you can also skip token billing entirely with a monthly plan from $5; the pricing page compares the options.
Where to use it
- Muse Code — the default model;
/modelswitches between variants mid-session. - Meta Model API — drop-in compatible with the OpenAI and Anthropic SDKs; change the model ID and keep your code.
- OpenRouter — both
muse-spark-1.3andmuse-spark-1.3-contributorare listed. - Cursor — team and enterprise admins enable the Meta provider first; Cursor offers a 300K-token context, or the full 1M in Max Mode, at the same per-token rates.
Is Muse Spark open source?
No. Muse Spark is available only through Meta’s products and API. Meta’s open-weight model is Muse Glimmer, released on August 10, 2026: a 30-billion-parameter agentic model distilled from Muse Spark, under the Apache 2.0 license, small enough to run on a single consumer GPU once quantized (about 20 GB). The weights are on Hugging Face as meta-models/Muse-Glimmer-30B, and it runs in Ollama and LM Studio. Meta hasn’t documented Glimmer as a model for Muse Code, so think of it as a separate, local option rather than a free version of the same agent.
Should you switch?
If you’re on Muse Code, you already have. If you call Muse Spark 1.2 through the API, switching is a one-line model-ID change — run your own prompts before and after, since “fewer tokens” is an average and your workload is not. For how the Spark line got here, see our write-up of how Muse Spark 1.2 was trained; for where it sits against rival agents, see the comparison page.