Meta Muse Spark 1.3 Ships With Two Prices for One Model — The 12.5x Cheaper Tier Trains on Your Prompts

Meta launched Muse Spark 1.3 on 2026-09-02, with paid API access opening the next day. Most coverage led with benchmarks. For anyone actually planning to use it, the thing to understand is that the same model has two prices, 12.5x apart.

1. The two tiers

Input is 12.5x cheaper, output more than 21x. The difference is not performance — it is data terms. The Contributor tier requires permission for Meta to use your prompts and completions to train future models. The Standard tier does not use your data for training.

OpenRouter, which lists the Contributor variant, shows the same numbers and states plainly that prompts and outputs may be used to improve Meta’s products.

2. How to make the call

This is not “cheap versus private” in the abstract. It depends entirely on what you send:

• Batch work on public material (summarising public web pages, translating public documents, drafting marketing copy) → the discount is enormous and the exposure is minimal.
• Customer data, internal source code, unreleased financials, personal data → use Standard. The savings are trivial next to the risk.
• Contract development work → check your client agreement first. Using the cheaper tier puts client material into a training pipeline, which most confidentiality clauses cover.

A simple test: if you would not paste it onto a public web page, do not send it through Contributor.

3. Specs and availability

• Context: 1,048,576 tokens (1M).
• Inputs: text, images, video and files (PDFs, audio); text output.
• Channels: the Meta Model API and Muse Code, Meta’s coding agent for the terminal and CI.
• Weights: closed for this release. Meta’s blog lists an open-weights Muse Spark release as a roadmap item with no timeline.

4. The best configuration is not the one you can buy

This is the detail most easily missed. Meta’s published results come from the max reasoning configuration, but the broadly available version runs at the xhigh setting. The max variant is still completing additional safety testing and remains in limited partner preview, arriving “shortly” per Meta.

In other words, the scorecard and the product are not the same configuration. If you are comparing it against your current model, benchmark the xhigh version yourself rather than copying the official chart.

5. The efficiency claim Meta made itself

Against version 1.2, Meta states coding tasks use about 20% fewer tool calls and about 25% fewer tokens.

That can matter more to a bill than the sticker price: a quarter fewer tokens for the same job is effectively a 25% discount. When evaluating a switch, compare unit price × actual consumption, not unit price alone.

6. Still unverified

• Meta’s official pricing page (dev.meta.ai/docs/pricing-rate-limits) returned a server error during this check. The prices above come from media reporting and the OpenRouter listing; we do not yet have a first-party read of the official page.
• Release date and pricing for the max configuration: Meta says only “shortly”. Unverified.
• Version and timing of the open-weights release. Unverified.

Verified 2026-09-06. Sources: Meta research blog, VentureBeat, OpenRouter listing. Prices can change at any time. For current mainstream model plans, see our Claude tool page.