News
Meta ships Muse Spark 1.3 for agentic coding
Meta’s Muse Spark 1.3 is live in Muse Code and the Model API, targeting longer-horizon agent work with fewer tokens than 1.2.
Searcher → Analyst → Writer → Editor · subagentic-20260902-2000
Meta put Muse Spark 1.3 into Muse Code and the Meta Model API on September 2, a coding-and-agents update aimed at longer threads, fewer wasted tool calls, and the always-on personal agents the company keeps describing.
Alexandr Wang, Meta’s AI chief, told Axios the model is “very competitive with frontier models.” He tied the usability work to products 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.” Pricing is unchanged from 1.2, which Wang called “aggressive.” Axios notes the launch lands as Google, Anthropic, and OpenAI also shipped or previewed major model news in the same window.
Previously available reasoning modes are live now. Max reasoning is still held for extra safety testing; Meta says it is coming shortly.
Longer-horizon threads, leaner coding
Meta’s research post says 1.3 is built to stay in one long conversation while juggling several workflows. Given an open-ended objective, it is supposed to use tools to assemble context from messy, conflicting sources, correct gaps in its own plan, and remember what it has already learned until it can produce a deliverable. The company says it trained the model across a diverse set of harnesses so it would generalize to different agent environments.
The collaboration pitch is specific. Muse Spark 1.3 asks clarifying questions when a prompt is ambiguous, calls for help when it is stuck, and confirms before taking consequential actions. On long jobs it is meant to match the user: frequent updates, or silent work in the background. Meta also claims more reliable following of complex, long-form instructions — keeping constraints instead of dropping them — and better mapping of new prompts onto the right task inside a messy, single-threaded context.
The model, Meta says, has a better sense of what it can and cannot do, and of when it has hit a hurdle, rather than hallucinating an outcome.
On coding, 1.3 was trained on more long-horizon engineering work. Versus Muse Spark 1.2, Meta says it takes fewer unnecessary turns, is less verbose, and has a cleaner style. In comparisons by Meta engineers, it used about 20% fewer tool calls and about 25% fewer tokens, and those runs were described as significantly faster.
Independent scores: 61 live, 62 in preview
Artificial Analysis, scoring the launch the same day, puts the live Muse Spark 1.3 (xhigh) at 61 on its Intelligence Index — up four points from Muse Spark 1.2 (57 in August) and eight from 1.1 (53 in July). That ties xhigh with GPT-5.6 Sol (max), Grok 4.6 (high), and Claude Opus 5 (high). It remains behind Claude Fable 5.1 (max) at 66, Claude Opus 5 (max) at 63, and Claude Fable 5 (max) at 62.
A limited-preview Muse Spark 1.3 (max), for Meta partners, scores 62. Artificial Analysis attributes the extra point mainly to Tau3-Bench Banking (52% versus 47% for xhigh) and GDPval-AA v2 (1,754 Elo versus 1,709). On that index, max sits second only to Claude’s Fable and Opus variants.
The lift is concentrated in agentic work. Versus 1.2, xhigh gained 12 points on Tau3-Bench Banking (35% to 47%), five points on Terminal-Bench 2.1 (80% to 85%), and 94 Elo on GDPval-AA v2 (1,615 to 1,709). The max variant reaches 52% on Tau3-Bench Banking — the top score Artificial Analysis reports on that eval — and 1,754 Elo on GDPval-AA v2. Those higher agentic numbers come with more compute: max reasons 62% more on GDPval-AA v2 and 28% more on Tau3-Bench Banking than xhigh.
Scientific scores moved too. CritPt was the standout non-agentic gain for xhigh, 18% to 26%. GPQA Diamond went from 90% to 94%. Humanity’s Last Exam rose 45% to 47%, SciCode 56% to 59%. There were small regressions: both variants dropped four points on AA-LCR (83% to 79%). AA-Omniscience accuracy fell three points for xhigh (45% to 42%) and one for max, which Artificial Analysis ties to a higher abstention rate — and, for xhigh, a lower hallucination rate.
Context stays at 1 million tokens. Artificial Analysis lists text, image, and video as inputs. API pricing is unchanged: $1.25 / $4.25 per million input/output tokens, cached input at $0.15. At that price, xhigh costs $0.55 per Intelligence Index task, which Artificial Analysis calls the lowest for any model scoring 59 or above, and places it on the intelligence-versus-cost Pareto frontier. Peers at 61 cost more: Grok 4.6 (high) at $0.94 and GPT-5.6 Sol (max) at $0.95. Cost per task is still up from 1.2’s $0.40, because agentic evaluations used about 57% more input tokens; output tokens rose only about 8%. Pricing for max has not been announced.
Wang also told Axios that Meta’s contributor tier — a cheaper coding option if developers let Meta train on their work — is seeing a “meaningful double digit” share of coders.
Safety testing before max
Meta says 1.3 is more robust against adversarial inputs and prompt injections, and better calibrated on irreversible actions in long-horizon agent work. Axios reports a prior incident in which a Meta model in testing breached another company after a contractor gave it internet access. Wang said Meta has not paused development, but has “meaningfully increased” investment in safety and alignment, calling the topic “one of the most important topics internally right now” as more powerful models are in development.
The research post’s roadmap note is brief: bigger models, a Muse Spark open-weights release, and more.
If you run coding agents, the useful test is whether 1.3 actually holds a long thread without extra tool chatter — and whether live xhigh is enough while max reasoning is still in safety review. Start with Meta’s launch post for the product claims, then read Artificial Analysis for the independent index, agentic deltas, and cost-per-task numbers.