Meta Muse Spark 1.3

Muse Spark 1.3, released on 2 September 2026, is Meta's newest agentic and coding model and the third release of the Spark line in two months. Meta's pitch is unusual for a frontier launch: the headline claim is not a benchmark record but efficiency. In comparisons run by Meta engineers against Muse Spark 1.2 the new model is described as significantly faster while using around 20% fewer tool calls and around 25% fewer tokens, taking fewer turns where they are not needed and writing less verbose code. Meta published a full scorecard with numbers, which is worth noting because the previous Spark releases came with charts only. Read honestly, it is a split result. In coding the model leads its comparison set: DeepSWE v1.1 75.4 (Muse Spark 1.2: 55.0; Opus 5: 74.0), SWEAtlas CodeBase QnA 59.4 (46.2; 52.7) and Terminal-Bench 2.1 88.8, tied with GPT 5.6 Sol. In long-context retrieval the gap is enormous: MRCR 256K-512K 98.5 against 66.3 for its own predecessor and 91.5 for GPT 5.6 Sol, and MRCR 512K-1M 98.1 against 55.5 and 73.8. In the agent rows, however, it usually finishes behind Anthropic's Opus 5 - GDPVal-AA v2 1754 against 1824, JobBench 64.9 against 65.7, AutomationBench 49.6 against 50.3 - and behind GPT 5.6 Sol on agentic browsing (DeepSearchQA 90.3 against 93.1) and instruction following (Agentic IF Index 57.8 against 60.5). The one agent row it does win outright is the binary score on OSWorld 2.0 computer use, 32.0 against 31.4. One caveat comes from Meta's own table headers: version 1.3 is measured at max reasoning while 1.2 is measured at xhigh, so the two columns are not the same effort setting. Beyond the numbers, Meta describes behavioural training rather than raw capability: the model asks clarifying questions on ambiguous prompts, calls for user help when stuck, confirms before consequential actions, and was trained to recognise its own limits instead of hallucinating an outcome. Safety work focused on the same ground - resistance to prompt injection and better calibration of what counts as an irreversible action. The model is available through Muse Code and the Meta Model API; Meta has announced that a Muse Spark open-weights release and larger models are on the roadmap, but has not published parameter counts or architecture for this version. Meta did not put a price in the launch post, but the developer documentation does, and it splits in two. On the Standard tier, where Meta states your prompts and completions are not used for training, the model costs USD 1.25 per million input tokens and USD 4.25 per million output, with cached input at USD 0.15. The much-quoted USD 0.10 and USD 0.20 figures belong to a second, separate model id — muse-spark-1.3-contributor — which is discounted precisely in exchange for permission to train on your prompts and completions. Two more details from the same documentation: the context window is exactly 1 048 576 tokens, and the max reasoning level used in Metas own benchmark table is available on the Standard tier only. Meta also warns that audio understanding in 1.3 is not fully supported and quality may be degraded, pointing audio workloads back to Muse Spark 1.2.

▸Videos

No videos yet.

▸Reviews