On September 2, 2026, Meta released Muse Spark 1.3 in Muse Code and the Meta Model API. The update is aimed at long-running agent and coding work: staying on task across a messy thread, gathering context with tools, asking for help when blocked, and preserving detailed instructions without drifting.
Meta's most useful launch claims are not that Muse is broadly "smarter." They are operational. In comparisons run by Meta engineers, version 1.3 used about 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2, while producing less verbose output. That is a cost and latency claim about completing work, not just a benchmark score.
MetaWhat changed in Muse Spark 1.3
Meta says the model was trained across several agent harnesses so it could generalize beyond one fixed tool environment. It is designed to map new prompts to the right task inside a long, interrupted conversation, revise its plan when sources conflict, and keep track of what it has learned before producing a deliverable.
The behavioral changes matter for real agents. Muse Spark 1.3 is trained to ask clarifying questions when a request is ambiguous, request help when it is stuck, and confirm before consequential actions. Meta also says the model has better awareness of what it does not know and is less likely to claim a task succeeded when it did not.
Meta reports lower resource use on common engineering workflows
Muse Spark 1.2 normalized to an index of 100. Muse Spark 1.3 values apply Meta's reported approximate reductions of 20% in tool calls and 25% in tokens. Lower is better.
Source and period: Meta engineer comparisons published September 2, 2026. This is vendor-reported relative usage, not an independent benchmark or an absolute token count.
The chart normalizes Muse Spark 1.2's resource use to 100 because Meta did not publish absolute tool-call or token totals. It visualizes the stated relative reduction, not an independent benchmark run. Actual savings will vary with the coding harness, task mix, reasoning setting, and how often an agent needs to recover from errors.
How to read the benchmark evidence
Meta's evaluation spans professional work, computer use, web research, business automation, software engineering, long-context retrieval, and instruction following. The methodology document is unusually useful about limitations. It says third-party model runs are best-effort and may not use provider-optimized prompts, tools, or runtimes. It also reports the highest comparable result available from Meta's run, an official leaderboard, or another provider's self-reported result, so the table is not one uniform independent experiment.
Several benchmarks are still concrete enough to explain what Meta is targeting. GDPVal-AA v2 covers 220 professional tasks across 44 occupations. OSWorld 2.0 runs 108 stateful workflows on a full Ubuntu desktop. AutomationBench checks 600 simulated business workflows with deterministic end-state assertions. DeepSWE v1.1 tests 113 software tasks across 91 repositories and five programming languages.
Those suites measure different things, and a lead on one does not establish a general winner. The release evidence supports a narrower conclusion: Meta trained 1.3 specifically for sustained, tool-using work and reports better efficiency than its own prior model. Independent evaluations will be needed to validate cross-company rankings.
Pricing and the contributor tradeoff
Meta kept standard Muse Spark 1.3 pricing unchanged from its predecessor. It also continues a lower-cost contributor tier for developers who permit their interactions to be used to improve Meta's models. Axios reported that a meaningful double-digit share of coders choose that option.
That is not merely a discount. It is a data-use decision, and organizations should evaluate it against source-code confidentiality, customer obligations, and internal AI policy. Teams handling sensitive repositories may reasonably prefer the standard tier even when the contributor price is lower.
Availability, safety, and what comes next
Muse Spark 1.3 is rolling out now in Muse Code and the Meta Model API. Meta says the previously available reasoning levels are live, while max reasoning will arrive after additional safety testing. It also says open weights are on the roadmap, without publishing a release date.
The release lands one month after Muse Code and Spark 1.2, covered in our Muse Code guide, and alongside new workhorse models from Google and Anthropic. See our comparisons of Gemini 3.8 Flash and Claude Fable 5.1.
That pace is a reason to keep model choice portable. For teams using multiple providers through metir, the practical test is not a vendor's aggregate chart but cost, reliability, and completion quality on the team's own tasks. Muse Spark 1.3 has a credible efficiency claim. Production traces and independent benchmarks will show how broadly it holds.
Sources:
- Introducing Muse Spark 1.3 | Meta AI Research
- Muse Spark 1.3 Evaluation Methodology | Meta AI Research
- Meta debuts Muse Spark 1.3 as personal agent work continues | Axios
- Muse Spark models and developer access | Meta AI
Image credits
Header and in-body portrait: Alexandr Wang, chief AI officer at Meta, provided by Meta Platforms via Wikimedia Commons, licensed under CC BY-SA 4.0. The portrait does not depict the Muse Spark 1.3 launch.