Meta Ships Muse Spark 1.3: A Leaner, Smarter Coding-Agent Model

Meta just pushed another update in its fast-moving push into AI coding: Muse Spark 1.3, the latest version of the model behind its terminal coding agent, Muse Code. Meta Superintelligence Labs announced it on September 2, 2026, and it began rolling out in Muse Code and the Meta Model API this week, with paid developer API access opening September 3.

If you've lost track of Meta's cadence here, it's been rapid: Muse Spark 1.1 in July, Muse Code plus Muse Spark 1.2 in early August, and now 1.3 — all in about two months, as Meta chases Anthropic's Claude Code and OpenAI's coding tools. Meta's own benchmark scorecard put 1.3 up against frontier rivals like GPT-5.6 and Opus 5.

Here's what actually changed, and whether it matters.

What's new in Muse Spark 1.3

More reliable on long, messy work. The headline theme is real-world usability on longer-horizon tasks. Meta says 1.3 sustains multi-step work across a single long thread better, generates its own context from messy and conflicting sources, proactively fixes gaps in its own plan, and keeps track of what it has learned toward a final deliverable.

It collaborates more, and guesses less. 1.3 is trained to ask clarifying questions when a prompt is ambiguous, ask for help when it's stuck, and confirm before taking consequential actions — and to adapt to how you like to work, whether that's frequent check-ins or quiet background execution. Meta also says it has a better sense of its own limits, so it's less likely to hallucinate an outcome when it hits a wall.

Better instruction-following. Across multi-step tasks, Meta reports 1.3 holds onto detailed requirements more reliably — dropping fewer constraints and drifting less from the requested workflow. That's the failure mode that makes agents frustrating on complex tasks, so it's a meaningful thing to target.

Leaner, faster coding. For developers specifically, 1.3 was trained on more long-horizon coding tasks and, in Meta engineers' own comparisons against 1.2, used roughly 20% fewer tool calls and 25% fewer tokens while producing a cleaner, less verbose coding style. Fewer tokens and turns means lower cost and faster runs — a practical win if the quality holds up.

Stronger safety on agentic tasks. Meta highlights improved adversarial robustness, including better resistance to prompt injection (a genuinely hard, unsolved problem for coding agents that read untrusted files and web content), plus better judgment about what counts as an irreversible action.

Availability and what's coming

Muse Spark 1.3 is live now in Muse Code and the Meta Model API. Muse Code installs on macOS or Linux with a single shell command, and pricing is unchanged from 1.2 — keeping its two token-based tiers: a Standard pay-as-you-go tier and a much cheaper Contributor tier that trades a discount for letting Meta train on your prompts and completions.

Meta also flagged the roadmap: a higher "max reasoning" mode is coming after additional safety testing, along with bigger models and — notably — a Muse Spark open-weights release. If that ships, it would put a capable coding model in the same open-weights lane Meta has used with Llama.

Should you care?

If you already work with a terminal coding agent, 1.3 is worth a look for two reasons: the efficiency gains (fewer tokens and turns means lower bills) and the better long-instruction following, which is exactly where agents tend to fall down. The honest caveats haven't changed — Muse Code is still a Mac-and-Linux beta, there's no Windows client, and the cheapest tier asks you to trade your code and prompts for the discount, which is fine for side projects but a real consideration for proprietary work.

Meta is clearly serious about the coding-agent race, and 1.3 reads as a steady, practical step rather than a flashy leap. In a field where the frontier moves monthly, "steadier and cheaper" is a perfectly reasonable place to compete.

Trying it out? See our Muse Code breakdown for the full feature, pricing, and comparison detail against Claude Code, Cursor, and Kiro.