Meta Superintelligence Labs, Meta's research organization, released the Muse Spark 1.3 AI model on September 2, 2026 (US time). It is available from day one in the developer coding environment Muse Code and through the Meta Model API, and the focus of this release is practical usability in long-running agentic work and coding. In Meta's internal comparisons, the model used about 20 percent fewer tool calls and about 25 percent fewer tokens than its predecessor, Muse Spark 1.2.

Holding Long Jobs Together in a Single Thread

What Meta emphasizes this time is less raw intelligence and more the question of whether the model is easy to use in real settings. The company says it fed back what it learned from months of broad adoption of Muse Code and the Meta Model API into the model's behavior.

Concretely, Muse Spark 1.3 was designed to sustain long-horizon work, juggling several workflows in a single long thread. Even when given only an open-ended objective, it is expected to use tools to build its own context from messy and conflicting sources, fill gaps in its plan on its own, retain what it has learned along the way, and carry the job through to a final deliverable. It was trained across a diverse set of execution harnesses, which Meta says makes it less dependent on any particular tool configuration.

Trained Not to Charge Ahead on Its Own

A recurring problem with agentic models is that they push forward on their own judgment when instructions are ambiguous. Muse Spark 1.3 was trained to ask clarifying questions when a prompt is unclear, to call on the user for help when stuck, and to confirm before taking consequential actions. It also adapts its reporting style to user preference, either providing frequent progress updates or working quietly in the background.

Adherence to complex, long-form instructions has improved as well. Across multi-step tasks, the model is less likely to drop detailed constraints partway through or drift away from the requested workflow. In addition, when several topics coexist in one thread, it more accurately maps an incoming instruction to the right task, distinguishing whether the user is steering a past request or interrupting the one in progress.

Another notable point is that the model was taught the limits of its own abilities. Meta says it trained Muse Spark 1.3 to recognize what it can and cannot do, what it knows and does not know, and to acknowledge when it hits a wall rather than fabricate an outcome. As a countermeasure to agents that pretend to have succeeded, it is an unglamorous improvement that should matter in practice.

Coding: Shorter, Quieter, Faster

On the coding side, the model was trained on more long-horizon development tasks. Compared with Muse Spark 1.2, it takes fewer unnecessary turns, is less verbose, and produces a cleaner overall coding style. In comparisons by Meta engineers, it used roughly 20 percent fewer tool calls and roughly 25 percent fewer tokens, showing a clear difference in both perceived speed and efficiency.

Pricing through the Meta Model API is 1.25 USD per million input tokens (about 200 yen) and 4.25 USD per million output tokens (about 670 yen), the same level as when the previous-generation Muse Spark 1.1 debuted. The context window is roughly 1 million tokens. ※1 USD = 158 JPY (based on the September 2, 2026 closing rate)

The previously available reasoning modes can be used from day one, but the top-tier max reasoning mode will be rolled out after additional safety testing is complete. Meta also published a scorecard comparing Muse Spark 1.3, Muse Spark 1.2, OpenAI's GPT 5.6 Sol (max), and Anthropic's Opus 5 (max) across agent, coding, instruction-following, and long-context evaluations.

Safety and What Comes Next

On safety, Meta focused on the areas most directly tied to agentic work and coding. It says the model has stronger resistance to adversarial inputs and prompt injection (attacks that steer an AI through instructions planted in external documents and the like), and better calibration on which actions are irreversible in complex agentic tasks, proceeding more carefully accordingly.

Looking ahead, Meta cites bigger models as well as a release of the Muse Spark 1.2 weights after final safety fine-tuning. Having steered from the free-distribution Llama approach toward a paid API, Meta is once again signaling that it intends to keep publishing open weights in parallel.

Summary

Muse Spark 1.3 is an update that polishes operational qualities rather than chasing benchmark supremacy: whether a long job can be delegated without falling apart, and whether the model hands control back to a person when unsure. The reduction in tool calls and token consumption translates directly into lower cost and shorter waits, and pricing stays at the previous level. The rollout of max reasoning and the open-weights release of Muse Spark 1.2 are the next things to watch.