Muse Spark 1.3: What Meta's New Agentic Model Actually Does

Published 23 September 2026 by Jake Hissitt, Founder of Stob.AI.

Meta spent two years being the open-weight company. Muse Spark 1.3 is the clearest sign that this is no longer the whole strategy. It is hosted, closed-weight, priced in the middle of the market, and aimed squarely at long-running agents rather than chat.

Here is what changed from 1.2, what the numbers actually say, and where it belongs in a production stack.

What Muse Spark 1.3 is built for

Meta positions this release around three things:

• Long-horizon agentic workflows. The model tracks context and prior results across a long task, works through messy or conflicting inputs, and asks for input when it is genuinely stuck rather than guessing.
• Coding across many turns. It is tuned for multi-step coding work with fewer unnecessary turns and cleaner output, whether you are building coding agents or using it as a development partner.
• Native multimodal perception. Video, images and documents are handled natively, and visual reasoning runs through a real execution environment rather than scripted steps. You can hand it a screenshot or a clip and ask it to build from that.

The headline addition in 1.3 is a max reasoning effort setting for the hardest agentic and reasoning tasks, alongside general real-world usability improvements.

Specifications and pricing

The pricing trap worth knowing about

Meta publishes two SKUs for the same model:

The contributor rate is roughly twenty times cheaper, and it is the number you will see quoted in a lot of coverage. It is not the number you can use for client work, regulated data, or anything covered by a confidentiality clause. Budget against the standard rate, and check which SKU your API key bills to before anyone builds a business case on it.

What the benchmarks say

Meta's published comparison puts 1.3 ahead of 1.2 across every category it reports, and in the same conversation as GPT-5.6 Sol and Claude Opus 5.

Two things stand out. Opus 5 still leads on raw agentic Elo. But on AutomationBench — the benchmark closest to the work we actually build, end-to-end business workflows — Muse Spark 1.3 matches GPT-5.6 Sol and edges past Opus 5, at a fraction of the token price.

Treat vendor-published benchmarks as a starting hypothesis, not a verdict. Run your own evaluation set before you route production traffic.

A worked cost example

Take a research-and-outreach agent handling 30M input and 10M output tokens a month:

That is a four-to-tenfold difference on identical volume. It does not automatically make Muse Spark the right answer — if the cheaper model needs more turns or more human correction, the saving evaporates. The measure that matters is cost per accepted outcome, not cost per token. Run both against the same task set and compare the completed work, not the invoice.

You can model your own volumes in our cheapest production AI models calculator.

Where it fits in a production stack

• Good fit: long-running agents that grind through many steps, document and screenshot understanding, high-volume coding assistance, and workflow automation where per-token cost is the binding constraint.
• Think twice: the hardest single-shot reasoning, where Claude Opus 5.5 and GPT-6 Astra still lead; and anywhere your data governance rules out a newer hosted provider.
• Not relevant: if you chose Meta specifically for open weights and self-hosting, Muse is a different product. Llama 4 Maverick and Scout remain the open answer.

The sensible pattern is the one we use across client systems: route by task. Cheap, capable models carry the volume; a frontier model handles the small share of work that genuinely needs it. That is a routing decision, not a procurement one — see our full provider comparison for the wider picture, or compare all 200+ models side by side in the AI model comparison tool.

Frequently asked questions

How much does Muse Spark 1.3 cost?

$1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens on the standard commercial rate. A contributor SKU costs $0.10 input and $0.20 output but allows Meta to use your traffic to improve its products.

Is Muse Spark 1.3 open source?

No. Muse Spark is Meta's hosted, closed-weight line. Llama remains Meta's open-weight family.

How big is the context window?

1 million tokens, with up to 131,072 tokens of output.

Is it better than Claude Opus 5.5 or GPT-6 Astra?

Not on raw capability at the top end. It is competitive on end-to-end business-workflow benchmarks while costing several times less per token, which makes it a strong default for high-volume agent work rather than a frontier replacement.

What changed from Muse Spark 1.2?

A new max reasoning effort setting, better real-world usability, and improvements across the published agent, long-context and coding evaluations. Pricing is unchanged.

What to do with this

Pick one workflow you already run on an expensive model. Build a test set of twenty representative tasks with known good outcomes, run both models against it, and compare completed outcomes and total cost. If Muse Spark 1.3 clears your quality bar, the saving is real and repeatable. If it does not, you have learned that for the price of an afternoon.

Compare other providers: Best ChatGPT Model 2026 · Best Claude Model 2026 · Best Gemini Model 2026 · Best Llama Model 2026