Meta announced the Muse Spark 1.3 artificial intelligence model, whose main improvements focus on coding, long-horizon agentic tasks, and complex multi-step workflows. The model became available as of September 3, 2026, through the Muse Code programming tool and the Meta Model API, about a month after the release of Muse Spark 1.2.
Meta describes the new release as its most capable model to date, but this wording remains a characterization issued by the company and does not mean that it outperforms all models in every test or use case.
Improved capabilities for managing extended tasks
Muse Spark 1.3 was designed to handle multiple tasks within a single long conversation. When given an open-ended goal, it can use tools to build the required context, identify and correct gaps in the work plan, and then transfer the information it learns during execution to the final result.
Meta also says that the model has become more capable of handling ambiguous instructions without immediately resorting to assumptions. It can ask clarifying questions, request the user's help when it cannot complete a task, and obtain prior approval before irreversible operations or operations that could have significant consequences. The company also highlights improved preservation of constraints within lengthy instructions and the ability to connect older requests with newer ones within the same conversation.
Meta indicates that the model is more accurate in estimating its capabilities, with the aim of reducing cases in which it implies that it completed a task it cannot actually perform.
Gains in coding and resource use
According to comparisons conducted by Meta engineers, Muse Spark 1.3 performs similar tasks using about 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2. These gains are also associated with reducing unnecessary dialogue rounds and producing less detailed but cleaner code, according to the company's description.
In the DeepSWE v1.1 benchmark for long-horizon software development tasks, the model scored 75.4 points. The benchmark covers 113 tasks across 91 software repositories written in TypeScript, Go, Python, JavaScript, and Rust. According to Meta's table, this result surpassed GPT-5.6 Sol and Claude Opus 5, while advanced competing models remained ahead of it in other evaluations.
What do the numbers actually prove?
The DeepSWE results show a measurable improvement on the stated benchmark, but they are not sufficient on their own to prove that Meta has closed the gap with OpenAI and Anthropic in all usage scenarios. The company itself cautions that tests on models from other organizations may not be optimized for each model and therefore may not reflect its highest possible performance.
Artificial Analysis also gave the max version of the model 62 points on the Intelligence Index and stated that Muse Spark 1.3 has a context window of one million tokens and supports text, image, and video inputs. Practical comparisons remain tied to the benchmark used, the evaluation method, and the agentic architecture surrounding the model.
Availability, pricing, and open limitations
Meta did not change its API prices despite announcing performance improvements. Standard prices per million tokens are $1.25 for inputs, $0.15 for cached inputs, and $4.25 for outputs.
The company says it has strengthened the model's resistance to prompt-injection attacks and malicious inputs, and trained it to exercise greater caution before performing irreversible operations. The existing reasoning modes are currently available, while the company plans to release the max reasoning option after completing additional safety tests.
The roadmap includes larger Muse Spark models and open-weight releases, but Meta has not clarified whether it will publish the weights of Muse Spark 1.3 itself. Therefore, the release's practical value will depend on the results of independent use, when the additional reasoning capabilities become available, and whether the openness plans will specifically include this model.