What Is Muse Spark 1.3, and Why Is Meta's Cheap Tier Training on Your Prompts? (September 2026)

Muse Spark 1.3 is Meta's new frontier model, released September 2, 2026, and by independent measurement it now sits within a few points of the best models from Anthropic and OpenAI at a Standard price of $1.25 per million input tokens and $4.25 per million output tokens. The fine print is the story. Meta also sells a Contributor tier of the same model for $0.10 and $0.20, and its own pricing page says the discount is "in exchange for permission to use your prompts and completions to train future Meta models." If you are a working professional, the question is not whether Muse Spark 1.3 is good. It is which tier the tools you use have quietly chosen.

What Muse Spark 1.3 is

Muse Spark 1.3 is the fourth release in Meta's Muse Spark line since April, following 1.1 in July and 1.2 in August. Meta's announcement describes it as trained for "long-horizon, agentic workflows": it holds context across multi-step tasks, "asks clarifying questions when prompts are ambiguous, invokes help from the user when stuck, and confirms before taking consequential actions," and follows long instructions without dropping constraints. Meta says internal comparisons showed roughly 20% fewer tool calls and 25% fewer tokens than 1.2 on coding work.

The specs, from Meta's developer documentation:

• Context window of 1,048,576 tokens.
• Input: text, images, video, PDFs, and audio. Meta warns that audio understanding "is currently not fully supported" in 1.3.
• Output: text only.
• Reasoning settings from minimal up to xhigh, plus a new "max" setting that Meta restricts to the Standard tier of 1.3.
• Hosted only. Meta lists "the Muse Spark open weights release" as a roadmap item with no version or date. Its self-hostable model is a separate, smaller line called Muse Glimmer.

It runs in two places Meta names: Muse Code, its multi-agent terminal coding tool, and the Meta Model API. Muse Code sells subscriptions at $5, $15, and $50 a month.

One wrinkle on "max." Meta's announcement says it "is now available," but VentureBeat reported the same day that Meta said the max configuration was still completing safety testing, and Artificial Analysis described it as a limited preview for Meta's partners. If a vendor advertises max-level results, check whether they can actually call it today.

The two tiers and what the cheap one costs you

Meta's pricing page lists both tiers side by side. The numbers below are per million tokens, and the data terms are Meta's words.

Tier	Input	Output	Cached input	Data terms
Standard (muse-spark-1.3)	$1.25	$4.25	$0.15	"Standard pricing; your prompts and completions are not used to train Meta models."
Contributor (muse-spark-1.3-contributor)	$0.10	$0.20	$0.002	"Heavily discounted token pricing in exchange for permission to use your prompts and completions to train future Meta models."

That is 12.5 times cheaper on input, 21 times cheaper on output, and 75 times cheaper on cached input. The same checkpoint answers both. The differences are the price, the rate limit, and what happens to your data afterward.

Meta's models page frames the Contributor tier as a way to "prototype and scale experiments where that's acceptable," and the developer page for the model puts it bluntly: the Contributor ID is "used to improve our products," the Standard ID is "not used to improve our products." What the terms leave out matters as much. Meta does not publish how long Contributor prompts are retained, whether humans review them, whether attachments and tool results count as "prompts," or whether you can revoke permission for a request already sent. The safe reading is that a Contributor-tier prompt is a permanent donation.

For a solo developer prototyping a side project, that trade can be fine. For anyone whose prompts contain client matters, patient details, financial records, or someone else's source code, it is not a trade you are allowed to make, and the person who made it may not have been you. Our guide on whether AI trains on your data covers the consumer opt-outs for Claude, ChatGPT, and Gemini. Muse Spark is different: the choice is made by whoever wrote the code that calls the API, per request.

How to tell which tier a tool is using

The tier lives in a single string, the model ID, that the tool sends to Meta with every request. You will almost never see it from inside the product. Three practical checks:

1. Ask the vendor the exact question. "Which Muse Spark model ID do you call, and do you ever use the Contributor tier?" A vendor on Standard will answer in one line. A vague answer, or a "we'll get back to you," is itself information. Ask for the answer in writing.
2. Read the data processing agreement, not the privacy page. Look for a clause saying your inputs and outputs are not used to train the vendor's models or any third party's. Many DPAs only cover the vendor's own training. A clause that says sub-processors handle data "in accordance with their terms" is exactly the gap the Contributor tier falls through.
3. Search their pricing and docs for "contributor." Meta's tier name is distinctive. If it appears in a vendor's cost breakdown, architecture notes, or a founder's post about margins, you have your answer. If a tool offers Muse Spark features at a price that only works on the cheap tier, ask how.

For internal tools, the same check is a one-line search of the codebase for the string -contributor. Ask engineering to run it and to add a rule that regulated workloads may only call the Standard ID.

Two Meta-imposed limits also make the Contributor tier visible in behavior. It is capped at 100 requests per minute against 3,000 on Standard, and it cannot use the "max" reasoning setting. A vendor claiming both "enterprise scale" and Contributor pricing is describing something that does not add up.

Where it sits against this week's other launches

Muse Spark 1.3 landed in the middle of the busiest launch week of the year: Claude Fable 5.1 on September 1, Gemini 3.8 Flash on September 2, and GPT-6 Astra on September 3.

By Artificial Analysis's Intelligence Index, Muse Spark 1.3 scores 61 at the xhigh setting and 62 at max. That puts it behind Claude Fable 5.1 at 66 and Claude Opus 5 at 63, level with GPT-5.6 Sol and Grok 4.6, and two points above Gemini 3.8 Flash at 59. Artificial Analysis also calculates the cost of running its full index at $0.55 per task on Standard pricing, the lowest of any model scoring 59 or higher. Its Muse Spark 1.3 analysis did not include a GPT-6 Astra score, so we are not quoting one.

Meta's own claims are stronger, as launch claims usually are. Its announcement shows a scorecard against GPT-5.6 Sol and Claude Opus 5, and VentureBeat reported Meta's numbers for the max setting: a GDPval-AA v2 Elo of 1,754 versus 1,709 at xhigh, and 66.9 versus 57.2 on OSWorld 2.0. Most of that gap comes from the max setting that is not yet broadly available, so treat the xhigh numbers as what you can buy today.

The honest summary: Muse Spark 1.3 is a real frontier-class model at a mid-tier Standard price, and a near-free one on Contributor. Mark Zuckerberg called it "frontier performance almost too cheap to meter," which is accurate for the tier that meters your prompts instead.

Where you can use it

What Meta confirms:

• Muse Code, Meta's terminal coding agent, on its $5, $15, and $50 monthly plans. Meta's subscription documentation says how it uses your inputs and outputs, "including any code you submit, depends on the models you select" and is governed by the Meta Model API terms, so the same Standard-versus-Contributor question applies inside Muse Code.
• The Meta Model API directly, at both tiers, and through OpenRouter.

What Meta does not confirm: any consumer surface. The 1.3 announcement names no consumer app. Meta's April 2026 launch of the original Muse Spark said the family powers Meta AI in the Meta AI app and website and was rolling to WhatsApp, Instagram, Facebook, and Messenger, and some coverage of 1.3 repeats that language. No Meta page found for this review says version 1.3 has reached those apps, so consumer availability of 1.3 is unconfirmed. Meta also publishes no enterprise tier, business agreement, or zero-retention program for the Model API. The Standard tier's "not used to train" line is the whole of the published commitment.

Bottom line

• If you or your vendors use Muse Spark for anything involving client, patient, financial, or regulated data, then the Standard tier is the only acceptable choice, and you need that confirmed in writing before the next request goes out.
• If you are prototyping on your own material with nothing confidential in the prompt, then the Contributor tier is the cheapest frontier-class model on the market, and using it is a reasonable trade as long as you know it is one.
• If a tool you rely on cannot tell you which tier it calls, then treat it as Contributor until proven otherwise, and move sensitive work to a tool that can answer.

---

Sources

1. Introducing Muse Spark 1.3 — Meta AI Research, September 2, 2026
2. Pricing and rate limits — Meta Model API documentation
3. Models — Meta Model API documentation
4. Reasoning — Meta Model API documentation
5. Muse Spark 1.3 — Meta developer model page
6. Muse Code — Meta developer product page
7. Muse Code subscriptions — Meta Model API documentation
8. Introducing Muse Spark: Meta's Most Powerful Model Yet — Meta Newsroom, April 2026
9. Muse Spark 1.3: Meta reaches the frontier — Artificial Analysis, September 2, 2026
10. Meta says Muse Spark 1.3 has frontier performance, but its best results come from a model developers can't broadly use yet — VentureBeat, September 2, 2026
11. Meta Muse Spark Contributor tier hides training consent where security tools cannot find it — Tech Times, September 4, 2026