> ## Content Index
> Fetch the complete content index at: https://bytevyte.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Muse Spark 1.2 Contributor tier: $0.10 tokens, paid in code
- URL: https://bytevyte.com/muse-spark-1-2-contributor-tier-0-10-tokens-paid-in-code/
- Published: 2026-08-25T15:35:02.000Z
- Updated: 2026-08-25T15:35:02.000Z
- Description: Meta's Muse Spark 1.2 Contributor tier cuts token prices to $0.10 per million if you let it train on your code. We weigh the trade-offs.
- Author: Bytevyte Editorial
- Tags: ai-beats

Meta has introduced the **Muse Spark 1.2 Contributor tier**, a pricing plan that trades token discounts for training data: input drops to $0.10 per million tokens and output to $0.20, against $1.25 and $4.25 on the standard tier, when developers agree to let Meta train future models on their prompts and completions. The tier went live on Meta's API on August 5 alongside Muse Code, the company's first terminal-based coding agent, and reached OpenRouter on August 23\. It is the first mainstream "pay with your data" pricing for a frontier-scale coding model, and it undercuts OpenAI, Anthropic, and Google on price by an order of magnitude.

The two SKUs share the same checkpoint, so the tier is a data license attached to identical weights rather than a watered-down model. Meta's API lists muse-spark-1.2 and muse-spark-1.2-contributor with the same 1,048,576-token context window, a 131,072-token output limit, multimodal input across text, image, video, and audio, text-only output, and configurable reasoning effort. What differs is the rate card and the data policy. OpenRouter routes both model IDs, which puts the Contributor rate one click away from any tool that already sends traffic through the aggregator.

## Inside the Muse Spark 1.2 Contributor tier

Standard pricing is unchanged from Muse Spark 1.1, so teams already paying for the older model get the upgrade at no cost. The Contributor rate is what Meta's model documentation describes as heavily discounted token pricing in exchange for permission to use prompts and completions to improve its products, a phrase broad enough to cover any future model family, not just the next Spark.

The rate card published on Meta's model page:

|                      | Standard       | Contributor                |
| -------------------- | -------------- | -------------------------- |
| Model ID             | muse-spark-1.2 | muse-spark-1.2-contributor |
| Input per 1M tokens  | $1.25          | $0.10                      |
| Cached input per 1M  | $0.15          | $0.002                     |
| Output per 1M tokens | $4.25          | $0.20                      |
| Training use         | No             | Yes                        |
| Rate limit           | 3,000 req/min  | 60 req/min                 |

The gaps are wide: 12.5x on input, 21.25x on output, and roughly 75x on cached input. The delta of about $1.15 per million input tokens is what Meta effectively pays for the right to train on your traffic, a price the company never states in those terms. Read strictly, the tier is a purchase offer: the currency is compute, and the asset is the right to train on your repository. On a working agent setup, the trade stops looking like a coupon and starts looking like a licensing agreement in reverse, where you receive compute and Meta receives the traces of real software work, including bug fixes, test runs, and the reasoning steps a terminal agent takes. With caching, the blended cost of a contributor session works out to roughly $0.04 per million tokens.

## A data flywheel dressed as a discount

Muse Code makes the arrangement structurally tighter. The terminal agent was co-trained with Muse Spark 1.2, with harness trajectories folded into the model's training data through rejection sampling, and it logs every call to a replay-safe event log, which lets Meta reconstruct agent runs during training instead of discarding failures. A contributor-tier session is cheaper compute, and it is also a training example produced inside a harness Meta already controls. Every prompt that fixes a build error or refactors a function is a labeled sample of real engineering, captured at the point of use rather than scraped after the fact.

The price position matters as much as the data. At $0.10 in and $0.20 out, the Contributor tier prices below DeepSeek Flash and near OpenAI's GPT-5.6 Luna at $0.20/$1.20 and Google's Gemini 3.1 Flash-Lite at $0.25/$1.50, while the standard tier sits close to Gemini 3.6 Flash at $1.50/$7.50\. Cached input at $0.002, a 75x cut from the standard $0.15, makes long agent sessions with repeated context the clearest win for heavy users. Meta has chosen to compete at the bottom of the coding-model market, where the value of each request's data outweighs the cost of serving it. The comparison set matters for buyers: the Contributor tier lands in the same price band as lightweight models like GPT-5.6 Luna, while running the full 1M-token context window and the 131,072-token output limit of a premium coding model.

Nothing comparable exists at OpenAI or Anthropic, which have kept their coding models at premium rates and avoided training on customer traffic as a pricing lever. The Muse Spark 1.2 Contributor tier collapses both decisions at once: it is the first time a major lab has sold a frontier-scale coding model at a fraction of its standard rate in exchange for data, turning the developer ecosystem into the training pipeline. Each session Meta serves at $0.10 buys a trace of real engineering work that rivals would have to generate synthetically or license separately. The bet is that the training value of one real coding session exceeds the cents it costs to serve it.

## Who should take the deal

The deciding factor for most teams is what the code is. For open-source and non-proprietary work, the Contributor tier is close to free, roughly 92 percent cheaper on input and 95 percent cheaper on output, and the 60-request-per-minute cap still covers interactive development. The heaviest users pay the most in data: teams streaming million-token contexts all day generate exactly the long-horizon traces that are hardest to synthesize, and they are the ones the cap most limits. The same cap rules out high-volume CI pipelines and batch agent fleets, which get 3,000 requests per minute on the standard tier, fifty times more headroom. It also sets a ceiling on bulk extraction: at 60 requests per minute, the cheap compute is not practical for scraping the model or feeding a rival training run.

For proprietary repositories, the choice is contractual. On a closed codebase, granting training rights is a data-licensing decision, not a billing option, and the rate cap removes the argument that the discount is free money. Meta's published policy statement for the tier is short: prompts and outputs are used to improve its products, with no public wording on retention, deletion, or how trained-in code is kept separate from the company's other services. The Contributor discount is permanent rather than promotional, so the terms are not a launch special that expires. Teams that need the standard tier's no-training guarantee pay $1.25 and $4.25 and receive none of the discount.

The ownership question is the one the rate card leaves open. Users on the standard tier can point to Meta's published policy that their data is not used for training; Contributor users hold only the wording that their traffic improves Meta's products, with no commitment about what those products become. The stakes sharpen once the promised open-weight release of Muse Spark 1.2 lands, because contributor traces that enter a training run would be baked into weights anyone can download.

The tier is one arm of a broader push at Meta Superintelligence Labs, the unit built around former Scale AI chief Alexandr Wang. On August 10, Meta released Muse Glimmer 30B open weights under the Apache 2.0 license, its first open-weight release in more than a year, and Zuckerberg has said Muse Spark 1.2 weights will follow. Open weights give developers an out from the data trade entirely: run the model on your own hardware and the Contributor contract never applies. The cheap tier is one path into the ecosystem; the open release is the other. On performance, Muse Spark 1.2 posts 82.9 on Terminal-Bench, the agent benchmark where Anthropic's Claude Code has been the reference, and the free Muse Code CLI, which carries no subscription, puts Meta directly into the same terminal workflows. Muse Spark 1.2 is also the third Muse model release in four months, a cadence that pairs each checkpoint with new tooling, and the terminal-agent focus is deliberate, since harnesses produce longer, more structured traces than chat interfaces.

## Why this matters

The Muse Spark 1.2 Contributor tier reframes what developers pay for model access: the currency is the work product as much as cash. For a CTO, the question is whether roughly $1.15 per million input tokens is fair compensation for the traces of enterprise coding, and whether the 60-request-per-minute cap lets the deal survive production scale. Whatever the answer, Meta has set a new baseline, and OpenAI, Anthropic, and Google now have to respond on price, on data terms, or both.

## Related Articles

- [Meta's Muse Spark 1.1 API Undercuts Rivals at $1.25 per Token](https://bytevyte.com/metas-muse-spark-1-1-api-undercuts-rivals-at-1-25-per-token/)
- [Meta Muse Code brings open-weights autonomy to the coding agent race](https://bytevyte.com/meta-muse-code-brings-open-weights-autonomy-to-the-coding-agent-race/)
- [Grok 4.5 Price War: SpaceXAI Undercuts Claude Opus by 75% to Dominate Coding Agents](https://bytevyte.com/grok-4-5-price-war-spacexai-undercuts-claude-opus-by-75-to-dominate-coding-agents/)

✔Human Verified

---

*Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.*