Meta Muse is now on every EBITDAI plan: the cheapest model we have ever run

What shipped, by plan

Meta Muse Spark 1.3 is now in the model picker on all three plans. Meta sells the model in two tiers with the same weights and different terms, so the picker shows them separately: "Meta Muse Spark 1.3 (trains on your data)" is Meta's Contributor tier, and "Meta Muse Spark 1.3 (not used for training)" is Meta's standard tier.

• Free plan. The Contributor tier, and nothing else. 150 requests a week, no card.
• Lite, $5/month or $48/year. The Contributor tier, and nothing else, plus 100x more usage a month than free, on it.
• Pro, $15/month or $144/year. Both Meta tiers, plus Kimi k3, DeepSeek V4.1 Flash, DeepSeek V4 Pro and Gemini 3.8 Flash, 7.5x more usage a month than Lite across every model, your own key for every provider, and the QuickBooks, Campfire and Stripe connectors.

Neither Meta option is ever the default on a paid plan. You have to pick it.

How cheap, exactly

List prices per 1M tokens, from each vendor's published rate card:

Meta rates: Meta, Meta Model API pricing and rate limits, September 2026. DeepSeek rates are off-peak and double at peak; V4 Pro is 3x Flash across the board. Meta's reasoning tokens are hidden from you and billed as output.

Against the standard tier of the same model, the Contributor tier is 12.5x cheaper on input, 21.25x cheaper on output and 75x cheaper on cached input. That last ratio is the one that decides what an agent session costs.

Why cached input is the number that matters

An Excel agent turn is mostly re-reading. The system prompt, the modeling rules, the tool definitions and the conversation so far get resent on every step, and only the tail of each request is new. In our agent loop 97% of input tokens come back as cache hits. So the rate that dominates a session is not the headline input price, it is the cached one, and that is where the Contributor gap is widest.

A full model build in our harness runs 716K input tokens, 97% of them cached, and 10K output tokens. Price that profile against the table above: on the Contributor tier the build lands at a fraction of a cent, and on DeepSeek V4.1 Flash at a few cents.

We measured rather than guessed, on the same conventions as our published Excel agent benchmark: seeded workbooks, identical assertions, cost per task from real invoiced token counts, raw transcripts kept. On the hard suite, ten tasks against existing workbooks rather than blank sheets, Meta Muse Spark 1.3 on the Contributor tier passed 9 to 10 of 10 at $0.003 to $0.005 a task. DeepSeek V4 Flash averaged $0.024 a task on the same suite, and Gemini 3.8 Flash averaged $0.19. Building a SaaS model from scratch cost $0.003.

Those figures are not comparable to the numbers on the benchmark page, which come from a 30-task suite that mostly builds on blank sheets. Compare models within one suite, not across the two.

What that buys you on a plan

The practical effect is on how much you get to run. Priced at that build, Lite's included monthly AI usage on Meta Muse, 100x more than free, goes a long way. Pro carries 7.5x the usage of Lite across every model, and it can spend all of it on Meta Muse at that same price per build.

Set those against a $20 Claude Pro plan, which buys 85 builds a month with its entire allowance spent on Claude for Excel. Lite is $5/month with 100x more usage than free and Pro is $15/month with 7.5x the usage of Lite. And the Claude allowance is shared with chat and coding, while yours is dedicated to Excel. The free plan is capped by request count instead: 150 a week on the Contributor tier, cheap enough to run that a free plan with a real agent behind it exists at all.

One quality note: we run Meta Muse at minimal reasoning effort, and at that setting it hardcoded closing balances instead of writing formulas on one bank reconciliation task, so check formulas on reconciliation work or switch models for it.

Now the training part

Meta's Contributor tier is discounted in exchange for your data. Meta's own words on the price list: "heavily discounted token pricing in exchange for permission to use your prompts and completions to train future Meta models." The tier is labelled "Used to improve our products."

Concretely, what leaves your machine on that tier is your prompt, the model's reply, and every cell value, header, formula and sheet name the agent reads out of your workbook to answer you. Meta trains on all of it. If you ask the Contributor model to fix the debt schedule in a live client model, the contents of that model become training data. That is not a risk or an edge case, it is the deal.

Meta's standard tier carries the opposite label: content submitted there is "not used to improve our products." It is offered on Pro. It is also the priciest included model per turn we run, because you are paying for the model and for Meta not learning from your numbers. A trivial two-operation turn on it cost $0.019.

Separately from Meta, there is what EBITDAI keeps. Free-plan conversations are logged and reviewed to improve the product, and we say so on the security page. On paid plans we store no prompts and no workbook content at all: they pass through the relay and are not kept. The two policies are independent. A paid plan stops our logging; it does not stop Meta's training if you pick the Contributor model.

So we made the choice loud. The picker labels say which is which. Whenever the Contributor model is selected, a dismissible amber bar sits at the bottom of the pane repeating what goes to Meta. Close it and it comes back the next time you open the pane or switch back to that model. And if you would rather Meta had nothing at all, pick another model. Switching is one click mid-session and you keep your place in the build.

Who should use which

The Contributor tier is the right default for work that is throwaway, public, or yours to give away. Learning the add-in. Scratch scenarios on invented numbers. Public filings and index data. Template building. Cleaning a downloaded CSV. It is fast, it passes almost everything we throw at it, and it costs close to nothing.

Client numbers are the other case. A live deal model, a portfolio company's actuals, a cap table, salary detail, anything under an NDA: run those on Meta's standard tier, or on DeepSeek, Kimi or Gemini, and read each provider's terms yourself rather than taking our word for it. Our security page lists where every model is served from and what each vendor says about training.

Try it

Start on the free plan: 150 requests a week on Meta Muse Spark 1.3, no card, no trial clock. That is the Contributor tier, so run it on data you do not mind sharing. When it earns a place in your work, Lite is $5/month with 100x more usage than free, on Meta Muse. Both Meta tiers, 7.5x the usage of Lite across every model and your own keys for every provider are on Pro at $15/month. See pricing for what each plan includes, and the benchmark for how we measure.