Meta's leaner agentic-coding model — the fourth Muse Spark in five months, keeping the 1M-token context window but using ~20% fewer tool calls and ~25% fewer tokens than 1.2.
Last verified · 2026-09-02 · by Moe Ameen
Muse Spark 1.3 is Meta Superintelligence Labs' agentic-coding model, launched September 2, 2026 and announced by CEO Mark Zuckerberg. It is the fourth model in a family that did not exist five months earlier — the original Muse Spark shipped in April 2026, version 1.1 in July, and 1.2 in August — and unlike those earlier jumps, 1.3 is an efficiency-and-reliability refinement rather than a ground-up overhaul.
The design center is long-running, multi-step agentic work, not one-shot chat. Muse Spark 1.3 keeps the 1 million-token context window and is tuned to sustain longer-horizon tasks — collaborating with a user and juggling several workflows in a single long thread, following complex instructions more reliably, and orchestrating tools with less waste. The headline improvement over version 1.2 is doing more with less: Meta reports roughly 20% fewer tool calls and about 25% fewer tokens on comparable tasks, plus a cleaner, less verbose coding style. On safety it claims stronger resistance to adversarial inputs and prompt injection and better calibration on what counts as an irreversible action during agentic tasks. A separate "max reasoning" tier is planned but held back pending additional safety testing.
The multimodal side is where it touches creative work. Muse Spark 1.3 perceives video, images, and documents and reasons over them in real execution environments — useful for summarizing a long recording, pulling the structure out of a PDF report, or describing a clip in detail. You reach it through Muse Code (Meta's beta coding agent) and the Meta Model API, with OpenRouter integration for existing workflows. Standard API pricing held at $1.25 per million input tokens and $4.25 per million output (cached input $0.15 per million), with a lower-cost "contributor" variant whose traffic Meta uses to improve its products. On Artificial Analysis's Intelligence Index, a partner-only "max" preview of 1.3 was reported to score 62 — third overall — though that configuration is not one most users can run yet.
The honest framing: Muse Spark 1.3 is a reasoning-and-coding model, not a content-production tool. It can draft copy, write a script, summarize a video, or reason over a document, but it hands back raw text or code in an agent session or an API response. It does not keep a brand voice across a week of posts, render a carousel or an avatar video, size anything per platform, or schedule and publish to your social channels. Treat the specific benchmarks, the preview status of max reasoning, and the exact pricing as a launch-window snapshot — verify current numbers on Meta's own pages.
The most creator-relevant thing about Muse Spark 1.3 is its multimodal reading: hand it a 40-minute webinar recording, a podcast transcript, or a dense PDF report and it will pull out the structure — the key moments, the quotable lines, the section outline. That is genuinely useful raw material. But it stops at a text description of your video; it cannot touch the video itself. It will tell you "the strong segment starts around minute 12," but it will not cut that segment into a vertical clip, reframe it to keep the speaker centered, burn in animated captions, or push it to TikTok. That gap — between knowing what to make and actually rendering and shipping it — is exactly the span Kompozy covers.
Here is the concrete handoff. Take the outline or key-moments list Muse Spark 1.3 produced from your long-form source and bring it into Kompozy. Kompozy's Clipped Shorts turn the actual long-form video into vertical cuts; the same brief fans into a brand-exact Carousel, Quote Graphics built from the lines the model surfaced, a Blog Article and an Email Newsletter, and native Text Posts — all held to one voice by your Persona Brief and banned-word filters, and each rendered as finished media, not a description of media. Then Kompozy schedules and publishes the whole set across eight social platforms plus blog and email from one queue with Autopilot and a per-post review pipeline. Muse Spark 1.3 reads your source and tells you what is in it; Kompozy is what turns that reading into the finished, on-brand posts and gets them live.
Muse Spark 1.3 is Meta Superintelligence Labs' agentic-coding model, launched September 2, 2026 as the fourth Muse Spark release in about five months. It keeps the 1 million-token context window and refines version 1.2 for efficiency — roughly 20% fewer tool calls and 25% fewer tokens on comparable tasks — and is available in Muse Code and on the Meta Model API.
Meta reports about 20% fewer tool calls and 25% fewer tokens on comparable tasks, plus a cleaner coding style, better instruction-following on long tasks, stronger prompt-injection resistance, and better calibration on irreversible actions. A "max reasoning" tier is planned but held back pending additional safety testing.
On the Meta Model API, the standard model is $1.25 per million input tokens and $4.25 per million output (cached input $0.15 per million), unchanged from 1.1 — but 1.3 uses about a quarter fewer tokens per task, so the effective cost is lower. A cheaper "contributor" variant is also offered in exchange for Meta using your traffic to improve its products. Confirm current numbers on Meta's pricing page.
No. Muse Spark 1.3 drafts text and code and can summarize or describe media you feed it, but it returns raw output in an agent session or API response. It does not render video or carousels, keep a brand voice across a week, size posts per platform, or schedule and publish anywhere. A content engine like Kompozy turns its drafts into finished, on-brand formats and publishes them across platforms.