Muse Spark 1.3 is Meta’s newest release in its small frontier-efficiency line — and Mark Zuckerberg calls its gains “the biggest jump we’ve made so far on coding and agentic work.” It is available today in Muse Code and through the Meta API, with open weights teased as “coming soon.”
This guide follows the official launch post, the 1.3 methodology document, Meta’s developer documentation, and the launch posts from Zuckerberg and @AIatMeta. It was checked on September 2, 2026, the day of the announcement.
What Meta launched
Muse Spark is Meta’s small-model line positioned for speed and cost efficiency rather than raw scale. Version 1.3 continues the cadence the line has held all summer: 1.1 arrived July 9, 1.2 landed August 5, and 1.3 ships September 2. Each release has targeted the same bottleneck — how much useful agentic work a cheap model can complete per dollar — and 1.3 attacks it on two fronts: raw benchmark capability and per-task efficiency.
The model is live in two surfaces at launch: Muse Code, Meta’s coding assistant, and the Meta API under the model ID muse-spark-1.3, with a muse-spark-1.3-contributor variant for a discounted tier. The official post also confirms a contributor-program data-sharing tier that trades a pricing discount for training-data permissions — the same structure 1.2 introduced.
The announcement
Zuckerberg announced the release with a benchmark collage — the same image reproduced in the next section — and a short claim: frontier performance “almost too cheap to meter,” with the line’s biggest coding and agentic jump so far. The post closes with a two-part teaser: “Next up 🍉 and Muse Spark open weights releases coming soon.”
Muse Spark 1.3 is here — frontier performance almost too cheap to meter. Biggest jump we’ve made so far on coding and agentic work. Try it in Muse Code and our API. Next up 🍉 and Muse Spark open weights releases coming soon.
— @finkd September 2, 2026
The emoji choice is not decoration. Meta has used fruit emoji as release markers before, and 🍉 in Meta’s release cadence points to the Llama family — meaning the open-weights track now covers both Llama and Muse Spark. The replies to the post immediately picked up on the open-weights line, and Meta’s own launch thread confirmed the roadmap language.
Muse Spark 1.3 is out: improved performance on agentic and coding tasks with a focus on real-world usability — and in internal comparisons, ~20% fewer tool calls and ~25% fewer tokens than 1.2 to complete the same work.
— @AIatMeta September 2, 2026
Benchmark evidence
Meta’s launch image compares Muse Spark 1.3 at maximum reasoning against three anchors: its predecessor (1.2, xhigh), OpenAI’s GPT 5.6 Sol (max), and Anthropic’s Claude Opus 5 (max). The full transcription:
| Benchmark | Muse Spark 1.3 (max) | Muse Spark 1.2 (xhigh) | GPT 5.6 Sol (max) | Claude Opus 5 (max) |
|---|---|---|---|---|
| GDPval-AA v2 (knowledge work) | 1754 | 1615 | 1710 | 1824 |
| JobBench (professional tool use) | 64.9 | 61.6 | 45.4 | 65.7 |
| OSWorld 2.0 (agentic computer use) | 66.9 | 47.6 | 62.7 | 68.3 |
| DeepSearchQA (agentic browsing) | 89.4 | 85.9 | 93.0 | 90.4 |
| Agentic IF Index (instruction following, internal) | 57.8 | 46.2 | 60.5 | 59.1 |
| AutomationBench (business workflows, Zapier) | 49.4 | 38.2 | 46.7 | 50.3 |
| MRCR v2 (long-context retrieval, 256K–512K, 8-needle) | 98.5 | 66.3 | 91.5 | — |
| MRCR v2 (long-context retrieval, 512K–1M, 8-needle) | 98.1 | 55.5 | 73.8 | — |
| DeepSWE v1.1 (long-horizon agentic coding) | 75.4 | 55.0 | 73.0 | 74.0 |
| SWEAtlas CodeBase QnA (codebase understanding, Scale AI) | 59.4 | 46.2 | 53.5 | 52.7 |
| Terminal-Bench 2.1 (agentic terminal coding) | 88.8 | 82.9 | 88.8 | 86.7 |
Three readings matter. First, the 1.2 → 1.3 delta is large everywhere: OSWorld 2.0 jumps 47.6 → 66.9 (+19.3), GDPval-AA v2 1615 → 1754, JobBench 61.6 → 64.9, DeepSearchQA 85.9 → 89.4, and MRCR v2 goes 66.3 → 98.5 in the 256K–512K band and 55.5 → 98.1 in the 512K–1M band. Second, 1.3 closes the gap to the much larger frontier anchors: it beats GPT 5.6 Sol on JobBench (64.9 vs 45.4), OSWorld 2.0 (66.9 vs 62.7), DeepSWE v1.1 coding (75.4 vs 73.0), and ties Terminal-Bench 2.1 at 88.8 — while Claude Opus 5 still leads GDPval-AA v2 (1824), JobBench (65.7), OSWorld 2.0 (68.3), DeepSearchQA (90.4), and AutomationBench (50.3). Third, MRCR v2 at 98.5 (256K–512K) and 98.1 (512K–1M) means near-perfect long-context retrieval across the full 1M window — where 1.2 scored 66.3 and 55.5.
The benchmark table measures quality at maximum reasoning — the most expensive configuration. The launch post’s efficiency claims cut the other way: ~20% fewer tool calls and ~25% fewer tokens than 1.2 for the same work. A model that scores near the frontier and needs fewer calls is targeting cost-per-completed-task, not leaderboard position.
The Muse Spark lineage
The 1.x line has shipped on a roughly monthly cadence, and each release’s positioning is visible in Meta’s own posts:
| Version | Date | Headline positioning |
|---|---|---|
| Muse Spark 1.1 | July 9, 2026 | First “frontier-efficiency” Spark; 1M context; intro of the contributor variant. |
| Muse Spark 1.2 | August 5, 2026 | Agentic collaboration focus; ~20% fewer tool calls / ~25% fewer tokens vs 1.1; Methodology 1.2 published. |
| Muse Spark 1.3 | September 2, 2026 | Biggest coding/agentic jump in the line; same efficiency deltas claimed vs 1.2; open weights teased. |
Two consistency notes worth knowing. The ~20% / ~25% efficiency deltas appear in both the 1.2 and 1.3 posts — measured against the prior version each time. And the eval-recipe change between 1.1 and 1.3 that some replies flagged is documented: Meta published a methodology document alongside 1.3 describing the updated evaluation harness, so cross-version comparisons should use the per-version numbers from each launch post rather than mixing harnesses.
Agent efficiency deltas
The launch post names the target workload explicitly: long agentic tasks — multi-step coding sessions, computer-use flows, deep research loops. The two efficiency claims are:
- ~20% fewer tool calls than 1.2 to complete the same task — fewer round trips, less latency, less per-call overhead.
- ~25% fewer tokens than 1.2 — less context thrash and shorter reasoning traces, which compounds with the tool-call saving.
Neither claim ships with a published per-benchmark breakdown in the launch post — they are aggregate characterizations. If your workload is tool-call-dominated (browser agents, IDE agents), the tool-call delta matters most; if it is context-dominated (long-document analysis), the token delta compounds with Muse Spark’s 1M window.
API, variants and pricing
Muse Spark 1.3 is served under the model ID muse-spark-1.3, with muse-spark-1.3-contributor as the data-sharing variant. Meta’s developer documentation lists both under the same Standard pricing as 1.2. Context window is 1M tokens.
| Tier | Price per 1M tokens | Rate limits | What you trade |
|---|---|---|---|
| Standard | $0.15 cached input · $1.25 input · $4.25 output | 3,000 RPM / 4,000,000 TPM | Nothing — prompts and completions are not used for training. Same pricing as 1.2 and 1.1. |
| Contributor | $0.002 cached input · $0.10 input · $0.20 output | 100 RPM / 3,000,000 TPM | Permission to use your prompts and completions to train future Meta models. Opt-in only. |
The structural point is that 1.3 does not carry a launch premium over 1.2 — Meta’s pricing page lists 1.3, 1.2 and 1.1 at the same Standard tier, so the upgrade path is a one-line model-ID change. There is no long-context premium (1M context costs the same whether the window is nearly empty or nearly full), the built-in web_search grounding tool is billed separately at $2.50 per 1,000 search queries on top of token cost, and Muse Image is billed flat at $0.01 per generated image.
Muse Code integration
Muse Code is the first-party surface where 1.3’s efficiency claims show up directly: fewer tool calls per task means faster completions and lower session costs inside the assistant. Zuckerberg’s post names Muse Code first among the launch surfaces, and the launch post describes the same agentic-collaboration positioning 1.2 introduced — the model plans, calls tools, and reports back in a single session rather than requiring external orchestration.
Roadmap: watermelon and open weights
The closing teaser — “Next up 🍉 and Muse Spark open weights releases coming soon” — establishes two parallel tracks:
- 🍉 (Llama): the next Llama-family release, in Meta’s fruit-marker convention.
- Muse Spark open weights: downloadable 1.3-class weights, extending the open-weights track that Llama started to the Spark line.
No dates were given for either. The open-weights tease matters most for two audiences: teams that fine-tune on private data, and teams with data-residency or cost constraints that make hosted-API-only lines a non-starter.
Safety notes
The launch post includes standard model-safety language: red-team evaluation coverage for agentic computer-use and browsing capabilities, and a note that the contributor tier’s data use is disclosed at opt-in. Meta’s methodology document covers the eval harness details. As with any agentic model, sandbox untrusted content and apply least-privilege tool permissions in production.
Caveats
All performance figures are vendor-reported. The comparison table, the efficiency deltas, and the competitor scores come from Meta’s launch materials with Meta-chosen configurations (“max” for 1.3, Sol, and Opus; “xhigh” for 1.2). Competitor numbers on a vendor chart are directional, not adversarially tuned — if you are migrating a pipeline, benchmark on your own workloads.
Two claims have no public breakdown. The ~20% tool-call and ~25% token savings are aggregate figures without per-benchmark methodology in the launch post. Treat them as marketing-grade characterizations until Meta publishes the underlying data.
Open weights have no date or license. “Coming soon” is the entire roadmap statement as of publication. The license terms — which determine commercial use — were not announced.
Use it if
Use Muse Spark 1.3 if your workload is high-volume agentic work — coding sessions, computer-use automation, browsing research — where cost-per-completed-task beats peak benchmark scores. The +19.3-point OSWorld jump over 1.2 with no price increase is the strongest small-model efficiency move in the current market.
Look elsewhere if you need the absolute top of a specific leaderboard — Claude Opus 5 still leads GDPval-AA v2, JobBench, and OSWorld 2.0 on Meta’s own table — or if you require open weights today; wait for the teased release and read its license first.
FAQ
What is Muse Spark 1.3?
Meta’s newest small frontier-efficiency model, released September 2, 2026, with the line’s biggest coding and agentic improvements. It is available in Muse Code and via the Meta API under the model ID muse-spark-1.3.
How much better is 1.3 than 1.2?
On Meta’s launch table: OSWorld 2.0 47.6 → 66.9 (+19.3), GDPval-AA v2 1615 → 1754, JobBench 61.6 → 64.9, DeepSearchQA 85.9 → 89.4, MRCR v2 98.5 and 98.1 across the two long-context bands (1.2 managed 66.3 and 55.5). Meta also claims ~20% fewer tool calls and ~25% fewer tokens than 1.2 for the same work.
How does it compare to GPT 5.6 Sol and Claude Opus 5?
It beats GPT 5.6 Sol on JobBench (64.9 vs 45.4), OSWorld 2.0 (66.9 vs 62.7), coding (75.4 vs 73.7), and ties it on Terminal-Bench 2.1 (88.8). Claude Opus 5 leads GDPval-AA v2 (1824), JobBench (65.7), OSWorld 2.0 (68.3), and DeepSearchQA (90.4).
What does it cost?
The same Standard pricing as Muse Spark 1.2, per Meta’s developer docs — no launch premium. A discounted contributor tier trades a pricing reduction for permission to train on your prompts and completions.
Are open weights available?
Not yet. Zuckerberg’s announcement says “Muse Spark open weights releases coming soon” with no date or license specified. The 🍉 teaser points to the next Llama-family release on the same roadmap.
What is the context window?
1M tokens. Meta evaluates MRCR v2 in two bands: 98.5 at 256K–512K and 98.1 at 512K–1M — vs 66.3 and 55.5 for 1.2.
Sources
Sources checked September 2, 2026:
- Introducing Muse Spark 1.3 — Meta research blog
- Muse Spark 1.3 Methodology — Meta research resources
- Muse Spark 1.3 model docs — Meta developer platform
- Mark Zuckerberg launch post (benchmark table and roadmap teaser)
- @AIatMeta launch thread (efficiency deltas)
Related reading: Gemini 3.8 Flash guide and Gemini agentic video understanding guide.