Meta shipped Muse Spark 1.3 on September 2, 2026: a closed multimodal reasoning model built for long-running agentic workflows, multi-agent setups and coding.
Two numbers made the headlines. It scores 75.4 on DeepSWE v1.1, edging past Claude Opus 5, and it hits 98.1% retrieval across a full million tokens of context. Both are real. Both come with a footnote, and the footnotes are what this post is about.
What Is Muse Spark 1.3?
Muse Spark is Meta Superintelligence Labs' closed flagship series. Version 1.3 succeeds 1.2, and the pitch is not producing one good answer but carrying a long task all the way to the end.
- Context window: 1,048,576 tokens (1M)
- Input modalities: text, image, audio and video
- Output: text only
- Weights: closed, not downloadable
- Access: Meta Model API and Muse Code
- Released: September 2, 2026
The headline gain over 1.2 is not a score, it is efficiency: the same work now takes roughly 20% fewer tool calls and 25% fewer tokens. On agent workloads that lands straight on the bill.
Muse Code is Meta's own coding-agent harness. The model spends fewer turns and fewer tokens inside its native environment, which means part of the published performance belongs to the scaffolding rather than the weights.
Two Modes: xhigh and max
This matters. Muse Spark 1.3 has two reasoning modes, and only one was open at launch:
| Measure | max | xhigh | Gap |
|---|---|---|---|
| OSWorld 2.0 | 66.9 | 57.2 | 9.7 points |
| GDPval-AA v2 (Elo) | 1754 | 1709 | 45 points |
| JobBench | 64.9 | 61.2 | 3.7 points |
| Intelligence Index (AA) | 48 | 45 | 3 points |
| Output speed | 226 tokens/s | 176 tokens/s | max is faster |
| Available at launch | No | Yes | max is in safety review |
Most of the numbers that travelled came from max, and max sat behind safety review at launch. Connect to the API today and what you get is xhigh.
Meta is not hiding this, it is in the announcement itself:
The thread continues: "max reasoning coming soon after we finish safety testing."
Reading the scores
Whenever you see a Muse Spark 1.3 figure, check which mode produced it. The two modes are 9.7 points apart on OSWorld 2.0, which is wider than the gap between many separate models.
Benchmark Results
Meta's own scorecard:
| Benchmark | Muse Spark 1.3 | Claude Opus 5 | GPT-5.6 Sol |
|---|---|---|---|
| DeepSWE v1.1 | 75.4 | 74.0 | 72.7 |
| Terminal-Bench 2.1 | 88.8 | 86.7 | 88.8 |
| MRCR v2 (256K-512K) | 98.5 | n/a | 91.5 |
| MRCR v2 (512K-1M) | 98.1 | n/a | 73.8 |
| SWE-Atlas Codebase QnA | 59.4 | n/a | n/a |
Long context is a blowout. MRCR v2 measures whether a model can find and use information buried inside a large body of text. Between 512K and 1M tokens Muse Spark scores 98.1% while GPT-5.6 Sol drops to 73.8%. Nothing else in this table has a 24 point gap.
Coding is a coin flip. DeepSWE puts it 1.4 points ahead of Opus 5, and Terminal-Bench 2.1 is a dead tie with GPT-5.6 Sol. Both of those come from the gated max mode. Independent measurement is more restrained too: on the Artificial Analysis Intelligence Index the model sits at 24th of 644.
So "it beat Opus 5" is true for one row, not as a general claim.
Pricing: Two Endpoints, 12.5x Apart
This is the genuinely interesting part. Meta sells the model through two separate endpoints:
| Item (per 1M tokens) | Contributor | Standard | Gap |
|---|---|---|---|
| Input | $0.10 | $1.25 | 12.5x |
| Output | $0.20 | $4.25 | 21.25x |
| Cache read | $0.002 | $0.15 | 75x |
| Your data | Meta may train on it | Stays private | The real price |
The Contributor tier is not a discount, it is a trade. In Meta's own wording, it offers "heavily discounted token pricing in exchange for permission to use your prompts and completions to train future Meta models".
So your prompts and the model's answers become Meta training data. If you handle personal data, customer records or proprietary source code, this endpoint is closed to you. For an open-source side project, a personal experiment or a non-sensitive batch job, a 21x cut is a serious offer.
Before you pick the Contributor endpoint
If this is work for an employer, do not make the call alone. Once customer data, health data or contractually protected source code is involved, the Contributor endpoint is a data-processing decision, not a pricing one.
Even the standard tier sits on the cheap side of the market:
| Model | Input | Output |
|---|---|---|
| Muse Spark 1.3 (Contributor) | $0.10 | $0.20 |
| Gemini 3.8 Flash | $0.75 | $3.75 |
| Muse Spark 1.3 (standard) | $1.25 | $4.25 |
| Qwen3.8-Max | $2.00 | $6.00 |
| GPT-6 Astra | $10.00 | $50.00 |
| Claude Fable 5.1 | $10.00 | $50.00 |
What Does That Mean on a Monthly Bill?
Make it concrete. Take an agent-heavy workload: 500M input tokens and 20M output tokens per month. Input dominates because the agent rereads the same context on every turn.
| Model / tier | Input (500M) | Output (20M) | Monthly total |
|---|---|---|---|
| Muse Spark 1.3 (Contributor) | $50 | $4 | $54 |
| Gemini 3.8 Flash | $375 | $75 | $450 |
| Muse Spark 1.3 (standard) | $625 | $85 | $710 |
| GPT-6 Astra | $5,000 | $1,000 | $6,000 |
Two conclusions.
First, the Contributor tier is in a different league entirely: the same workload costs 111 times more on GPT-6 Astra. That is no longer a discount, it is a different business model. Meta is not selling cheap tokens, it is buying training data and paying for it in rebate.
Second, and less discussed: Muse Spark's standard tier is more expensive than Gemini 3.8 Flash. Picking Muse Spark as "the cheap option" on price alone is a mistake. What justifies the standard tier is not cost, it is the million-token retrieval score and four-modality input.
What Changed From 1.2 to 1.3?
Meta is not promising a score jump in this release, it is promising efficiency. To finish the same task:
- roughly 20% fewer tool calls
- roughly 25% fewer tokens
On agent workloads that can be worth more than a score. If a task runs 40 turns and each turn makes a tool call, 20% fewer calls pulls both the bill and the wall-clock time down. The $710 monthly figure above would have been around $900 doing the same work on 1.2.
Meta's other listed changes are harder to measure: the model now asks clarifying questions, keeps better track of the task map across long threads, and is better calibrated about its own limits, meaning it is less prone to pretending it can do something it cannot.
That last one matters in agent setups. A model that keeps attempting work it cannot do burns tokens and produces wrong output.
Are There Open Weights?
No. Muse Spark 1.3 is closed, with no downloadable weights and no Hugging Face repository. Meta has said an open-weights Muse Spark release is coming "soon", but no date, variant or licence has been confirmed.
The Muse you can download and run today is Muse Glimmer: 30 billion parameters, Apache 2.0, distilled from Muse Spark. Architectural relatives, but Glimmer is small and local while Spark is large and API-bound.
How to Use It
There are two channels:
- Meta Model API: direct API access. The standard endpoint serves xhigh mode; the Contributor endpoint serves the same model at the discounted rate.
- Muse Code: Meta's own coding-agent interface, where the model spends fewer turns and fewer tokens.
Meta has not announced a free tier. The Contributor rate is cheap but not free, and you pay in data. The choice between endpoints is made at the API key level, so one account can route sensitive work to the standard endpoint and non-sensitive batch jobs to Contributor.
One warning: the max mode you see in benchmark tables is not on the API yet. Plan against max scores and the performance you actually get will be lower.
There is a modality detail too: the model accepts audio and video as input but produces text only. Video summarisation and audio transcription are on the table; video or audio generation is not.
Who Is This Right For?
- Anyone working with very long context: the strongest case by far. A 24 point MRCR v2 lead above 512K tokens is not something a competitor closes with price.
- High-volume agent workloads: if you can accept the data trade, the Contributor tier's 21x cut is the most aggressive offer on the market.
- Video and audio input: few models take all four modalities.
- Enterprises on confidential data: stay on the standard endpoint. $1.25/$4.25 is still an eighth of the frontier shelf.
- Everyday coding: no rush. The DeepSWE lead is 1.4 points and it came from the gated max mode.
- Running locally: not Spark, Muse Glimmer.
Conclusion
The real story of Muse Spark 1.3 is not the benchmark table, it is two footnotes: most of the headline scores come from a mode you cannot use yet, and the eye-catching price comes with your data as the payment.
Set those aside and something solid remains: a retrieval score at a million tokens that nobody comes close to, at an eighth of frontier pricing.
Which raises the question I keep going back and forth on: would you send your prompts and completions to a vendor's training set for a 21x discount? Where is your line, and does it move when it is a side project instead of work?
Originally published on Proje Defteri, where this post is kept up to date.
Also on the site: more English posts on AI models, Arduino and IoT, and free browser tools for makers and developers - token counter, LLM cost calculator, LCD and OLED bitmap converters.
Your support means a lot! ✨ Comment 💬, like 👍, and follow 🚀 for future posts!
Top comments (0)