Meta Launches Muse Spark 1.3: New AI Model Takes Aim at OpenAI and Anthropic

Meta has released Muse Spark 1.3 on Sept. 2, 2026, introducing an update focused on coding, long-horizon agentic work and practical usability. The model is available through Muse Code and the Meta Model API, while Meta said its max reasoning mode will follow after additional safety testing.

The launch comes as Meta competes with Anthropic, OpenAI and Google in an increasingly crowded market for advanced AI models. Meta AI chief Alexandr Wang told Axios that the update is intended to support the company’s work toward personal AI agents.

Coding And Agentic Work Drive The Update

According to Meta, Muse Spark 1.3 was trained on longer-horizon coding tasks and is designed to handle extended agentic workflows. The company said the model can maintain requirements across multi-step tasks, manage multiple workflows in a single conversation and ask users for clarification when instructions are ambiguous.

Meta’s engineers also found that Muse Spark 1.3 used about 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 in their comparisons. It described the change as an improvement in efficiency as well as coding usability.

The company’s published evaluation methodology says Muse Spark 1.3 was evaluated using the Meta Model API. For comparable tests, Meta compared it with Muse Spark 1.2, Anthropic’s Claude Opus 5 and OpenAI’s GPT-5.6 Sol, using max reasoning for the newer model and the competing models and xhigh reasoning for Muse Spark 1.2.

Meta reported a 75.4% score for Muse Spark 1.3 on DeepSWE 1.1, a benchmark measuring long-horizon software engineering. It also reported an 88.8% score on Terminal-Bench 2.1, which evaluates coding agents operating in terminal environments.

For long-context retrieval, Meta reported scores of 98.5% in the 256,000-to-512,000-token range and 98.1% in the 512,000-to-1-million-token range. The evaluation methodology describes these tests as retrieval tasks using eight target items distributed across the context.

Alexandr Wang Links Model To Personal AI Agents

Meta AI chief Alexandr Wang said Muse Spark 1.3 is intended to support the company’s longer-term development of consumer AI products.

“It’s very competitive with frontier models,” Wang said. “A lot of the usability improvements that we’ve made will be really helpful for things that Mark [Zuckerberg] has talked about on earnings calls, like personal agents that can work 24/7 on your behalf and help you achieve your goals.”

Wang’s comments to Axios position the model as part of Meta’s broader effort to develop AI systems capable of handling tasks on behalf of users rather than simply responding to individual prompts.

Meta Continues Shift Toward Paid Model Access

Muse Spark 1.3 is available through Muse Code and the Meta Model API, according to Meta. The company has also said that an open-weights version of Muse Spark is part of its roadmap.

It creates a distinction between Meta’s current proprietary Muse Spark offering and its open-weight strategy. Meta released Muse Glimmer in August as an open agentic model designed to run on consumer devices, while Muse Spark 1.3 remains accessible through Meta’s products and API.

Meta’s latest release also follows a rapid sequence of Muse Spark updates. The company introduced the original Muse Spark in April, followed by Muse Spark 1.1 in July and Muse Spark 1.2 in August. Meta’s Sept. 2 announcement describes 1.3 as the latest step in that development cycle.

AI Infrastructure Spending Remains A Major Focus

The model rollout comes as Meta continues to increase spending on infrastructure to support its AI ambitions.

Meta’s latest financial guidance calls for 2026 capital expenditures of $125 billion to $145 billion, including principal payments on finance leases. The company said the increase from its previous guidance reflected higher component prices and additional data center costs to support future capacity.

Meta has separately expanded its data center and custom-chip investments as it builds capacity for AI workloads. In March, the company said it was developing four new generations of its Meta Training and Inference Accelerator chips over two years.

The scale of those investments provides broader context for Meta’s model strategy, such as advances in AI software are being accompanied by significant spending on computing infrastructure, data centers and specialized hardware.

Open-Weights Release And Max Reasoning Mode Ahead

Meta said previously available reasoning modes for Muse Spark 1.3 were available at launch, while max reasoning would arrive after additional safety testing.

The company also reported improvements in adversarial robustness and resistance to prompt injections. It said the model has improved calibration around irreversible actions in complex agentic tasks.

Meta’s roadmap also includes an open-weights release of Muse Spark, although the company did not provide a specific release date in its announcement.

Muse Spark 1.3 represents Meta’s latest push to improve coding and agentic performance while reducing the computational work required for some tasks. Its longer-term position against competing models will depend on independent evaluations, wider deployment and how quickly Meta incorporates the technology into products used by consumers and developers.