Meta releases Muse Spark 1.3, its long-horizon agentic model

Meta rolled out Muse Spark 1.3 on September 2, 2026, the latest version of its Muse Spark model family, now live in Muse Code and Meta's own Model API. The release targets long-running agentic and coding work rather than chasing raw benchmark gains, continuing the cadence Meta set with Muse Spark 1.1 and 1.2 earlier this year.

What's new

Meta's research announcement says Muse Spark 1.3 "delivers improved performance across agentic and coding tasks" and is "Smarter and more practically useful" than its predecessor. Meta's own engineers reported the model is "significantly faster and more efficient," needing fewer turns and producing less verbose output than Muse Spark 1.2.

On the model-card side, Meta's developer page describes the model as "Trained for long-horizon, agentic workflows" that "tracks context and prior results, works through messy or conflicting inputs, and asks for input when needed." It adds native multimodal perception for video, images, and documents, with visual reasoning inside a real execution environment — the same closed-loop pattern (observe screen, act, verify) that rival labs have been building into their latest agentic models.

Two tiers ship at launch:

• Muse Spark 1.3 — the full model, priced at $1.25 per million input tokens and $4.25 per million output tokens ($0.15 for cached input), with a 1M-token context window.
• Muse Spark 1.3 Contributor — a cost-efficient tier for experimentation and early-stage workflows, at $0.10 per million input tokens and $0.20 per million output tokens. Meta discloses that prompts and outputs on this tier "may be used to improve Meta's products."

On coding specifically, Meta says the model uses roughly 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 in internal comparisons, with a cleaner coding style and fewer unnecessary turns. Meta also flags safety work behind the release, citing "stronger adversarial robustness, with improved resistance to adversarial inputs and prompt injections" and better calibration around irreversible actions. A "max reasoning" variant is coming later after additional safety testing.

Context

Muse Spark has been Meta's answer to the agentic-coding push that OpenAI, Anthropic, Google, and xAI have all leaned into this year — long context windows, tool-calling efficiency, and models that can run for extended stretches inside an IDE or browser without going off the rails. Meta shipped Muse Spark 1.1 in July alongside Muse Image and Muse Video, and iterated to 1.2 with the Muse Code launch shortly after. The two-tier pricing structure (a cheap, data-sharing "Contributor" tier alongside a full-price private tier) has become Meta's standard way of subsidizing wide developer adoption while still monetizing at the top end.

Why it matters

The efficiency claims — fewer tool calls, fewer tokens, less back-and-forth — matter more than raw capability scores for anyone running agents at scale, since token and turn count are what actually show up on the bill. If Meta's internal numbers hold up under independent testing, Muse Spark 1.3 becomes a meaningfully cheaper option for long-running coding agents than its immediate predecessor, without requiring a new architecture or a price increase. The explicit prompt-injection hardening is also notable: as agentic models get wired into browsers and file systems, resistance to adversarial inputs is becoming a competitive feature in its own right, not just a compliance checkbox.