Meta shipped Muse Spark 1.3 this week through Muse Code and the Meta Model API — the fourth release in a five-month sprint that started in April. The xhigh tier is live now. The heavier max variant exists only as a limited partner preview while safety testing finishes. That cadence alone says something about how aggressively Meta is iterating, but the real story is the price-performance math.
On Artificial Analysis’s Intelligence Index, max scores 62 and xhigh lands at 61 — up from 57 in August and 53 in July. Those gains aren’t evenly distributed. The index weights GDPval-AA v2 at 20 percent, Terminal-Bench 2.1 at 16 percent, and τ³-Bench Banking at 14 percent, and Meta’s biggest jumps land exactly there. On τ³-Banking, where agents operate tools in a simulated banking environment, max hits 52 percent — the only outright number-one finish the model claims. The available xhigh tier manages 47 percent, tying Claude Fable 5.1 max and GLM-5.3-Flash rather than beating them. Terminal-Bench 2.1 climbs from 80 to 85 percent on xhigh, but Claude Fable 5.1 still dominates at 91.4 percent in its max tier.
What sells Muse Spark 1.3 is economics. At $1.25 per million input tokens and $4.25 per million output, one index task costs $0.55. No model scoring 59 or higher is cheaper — rivals at the same index level run between $0.94 and $1.23. It’s slightly pricier than the 1.2 release at $0.40, but still undercuts everything in its performance class. There’s a catch, though: the max variant burns 62 percent more reasoning tokens than xhigh to buy its edge. Meta is paying for performance in compute, not cleverness.
Outside agentic tests, the picture is mid-pack. GPQA Diamond rises from 90 to 94 percent, respectable but behind Gemini 3.8 Flash high at 95.3 and Grok 4.6 high at 94.9. CritPt in research physics jumps from 18 to 26 percent — still far behind GPT-5.6 Sol max at 32.3. Two scores actually dropped against 1.2: AA-LCR fell from 83 to 79 percent, and factual accuracy in AA-Omniscience slipped by up to three points because the model now declines to answer more often when unsure. That’s a trade-off worth watching. Refusing to answer preserves credibility, but it also means the model is doing less work. Neither Meta nor Artificial Analysis has priced the max tier yet, and an open-weights release is apparently in the pipeline. If the open version lands anywhere near this performance, it resets the floor for what budget AI can do.
