Muse Spark 1.3

Muse Spark 1.3 launched September 2, 2026, per Meta's own developer page and AI Research blog post, independently corroborated by Bloomberg, Axios, and Neowin the same day. It follows the same roughly-monthly Muse Spark cadence tracked here before: base Muse Spark (2026-04-08), Muse Spark 1.1 (2026-07-09), Muse Spark 1.2 (2026-08-05), now 1.3. Meta's own materials describe it as trained on "months of broad adoption of Muse Code and Meta Model API" — the page does not state whether this is a new pretraining run or a post-training update to 1.2's checkpoint, unlike Google's explicit disclosure for Gemini 3.8 Flash.

Pricing is unchanged from Muse Spark 1.2: the Standard tier runs $1.25 in / $4.25 out per 1M tokens ($0.15 cache-hit) and carries a no-training-use guarantee; the Contributor tier is $0.10 in / $0.20 out ($0.002 cache-hit) but opts prompts and completions into Meta's training data — the same roughly 12x/21x discount-for-privacy trade this site has already flagged on 1.2's own page (confirmed side by side on developer.meta.com/ai/models/muse-spark/, checked 2026-09-03).

Artificial Analysis tracks two reasoning-effort configurations for 1.3: "max" (its new top tier, and the canonical/default page for this model) and "xhigh" (the same ceiling tier 1.2 was tracked at here). The two don't move in the same direction across benchmarks — max scores higher on Humanity's Last Exam (49.1% vs xhigh's 47.5%) and Terminal-Bench 2.1 (85.8% vs 85.4%), but xhigh scores higher on GPQA Diamond (94.1% vs max's 93.8%). This site tracks max as the default going forward, consistent with how Artificial Analysis presents the model's own canonical page — but that means the GPQA Diamond comparison against 1.2's own tracked 90.4% (itself an xhigh-tier figure) mixes tiers; on a matched xhigh-to-xhigh basis the improvement is larger (94.1% vs 90.4%). Of the four benchmarks 1.2 and 1.3 share here, only one — LiveBench — supports a clean, tier-matched, un-saturated real-vs-noise verdict: both models were scored at xHigh Effort, and the 3.6-point gain clears LiveBench's own noise band. Humanity's Last Exam's tracked 3.6-point gain (49.1% vs 45.46%) mixes tiers the same way GPQA Diamond's does — 1.3's tracked figure is AA's max-effort run, 1.2's is xhigh — and on a matched xhigh-to-xhigh basis (47.5% vs 45.46%) the gain shrinks to 2.04 points, barely above HLE's own 2.0-point noise floor rather than the comfortable margin the mixed-tier figure implies. Terminal-Bench 2.1's 5.6-point gain reads as a tie against that benchmark's own ±10.6 margin regardless of tier. GPQA Diamond's 3.4-point gain is moot either way — the benchmark is saturated and every comparison on it is tainted regardless of gap size.

Four of the eleven benchmarks this site tracks carry an independently-run Muse Spark 1.3 score, one day after launch: Humanity's Last Exam, Terminal-Bench 2.1, and GPQA Diamond (all Artificial Analysis), and LiveBench (81.6 overall on the 2026-06-25 release round, xHigh Effort tier — Reasoning 89.7, Coding 81.1, Agentic Coding 64.1, Mathematics 95.9, Data Analysis 79.6, Language 82.8, Instruction Following 78.0). Terminal-Bench 2.1's own canonical board, tbench.ai, still lists only Muse Spark 1.1, not 1.2 or 1.3.

Coverage gaps mirror 1.2's own, checked live on each as of 2026-09-03: LiveCodeBench (vals.ai has no Muse Spark row past 1.1), SWE-bench Verified (vals.ai's newest Meta row is still 1.2, at 86.60%), Toolathlon-Verified (toolathlon.xyz's newest Meta row is still 1.2, at 75.9%), DeepSWE (deepswe.datacurve.ai's newest Meta row is still 1.2's 55%, both on "Best" and "all effort levels" views), ARC-AGI-2, Agents' Last Exam, and HMMT Feb 2026 (all three have never carried any Muse Spark version, 1.3 included). Meta's own launch materials self-report 75.4% on DeepSWE v1.1 — a real gap from 1.2's own independently-verified 55% if the boards ever converge on 1.3, but that comparison cannot be made yet because no independent DeepSWE run of 1.3 exists.

Licensing is a harder story than 1.2's own "coming soon." Meta shipped 1.2 closed, with an August 10 promise from Chief AI Officer Alexandr Wang that open weights were imminent — still unfulfilled as of this writing. For 1.3, Meta has gone further: reporting describes the weights decision as explicitly undecided, citing the EU AI Act's Article 53 — the exemption for genuinely open-source models does not apply once a model is classified as carrying systemic risk, which frontier-scale models are liable to be. Wang separately claimed 1.3 is "competitive with Anthropic's Claude Fable 5.1, better than OpenAI's GPT-5.6 Sol at coding, and ahead of any current Chinese model" — a qualitative, self-reported cross-vendor comparison with no benchmark number attached in the reporting that carries it; this site records it as Meta's own claim, not a finding.