跪拜 Guibai
← Back to the summary

Meta's Muse Code Undercuts the Market at $0.20, but the Real Price Is Your Code Trajectories

Summary: On August 5, Meta released its first coding agent, Muse Code, alongside the purpose-built Muse Spark 1.2 model. It enters the market dominated by Claude Code and Codex with near-cost pricing (the Contributor tier costs only $0.20 per million output tokens). But the low price comes with a condition: Meta uses your code trajectories to train the next generation of models. This is a play that Chinese companies figured out long ago.

On August 5, Meta did something: it released its first coding agent, Muse Code.

The company that holds the open weights for Llama and owns one of the world's largest monolithic code repositories has officially entered the arena to compete for developers against Anthropic's Claude Code and OpenAI's Codex.

Looking at features alone, Muse Code isn't novel—terminal installation, reading codebases, planning changes, writing code, running tests for verification. This set is standard for coding agents in 2026. What's really worth dissecting are two things: how it prices so low, and what it gets in exchange for that low price.

Muse Spark 1.2: A model trained together with the harness

Muse Code runs on Muse Spark 1.2, a model Meta trained specifically for it. This "specifically trained" isn't marketing fluff; it represents a real shift in technical approach.

Previously, the training logic for most coding models was: train the model first, then layer an agent framework (harness) on top to teach it tool calling. Meta reversed this—Muse Spark 1.2 incorporated tool-calling and multi-agent collaboration trajectories into its training data. The official blog's phrasing is "Muse Code was in the training loop from day one."

This means two things:

First, higher tool-calling success rates. The model saw complete trajectories of "tool call failure → read error → correct" during training, rather than just being asked to "use tools" at inference time. Officials claim this leads to higher first-attempt accuracy and more reliable tool calling.

Second, its generalization ability to other harnesses. Meta emphasizes the model was trained on multiple harnesses, so Muse Spark 1.2 isn't only useful inside Muse Code; it can also be plugged into other coding tools you're already using. This is very pragmatic—developers won't switch away from their familiar setup just because Meta released a new tool.

The model itself comes with a 1-million-token context window, capable of holding dependency graphs, legacy code, and thousands of files. This is an entry ticket for long-horizon tasks, not a crushing advantage—Claude and Gemini series already have million-level contexts.

Pricing: $0.20 for your code trajectories

Muse Code's pricing is the most thought-provoking part of this release.

Two tiers:

For comparison, Anthropic's Claude Opus 5 output price is $25/million tokens, and GPT-5.6 Sol is $30. Meta's Contributor tier is over 100 times cheaper than both.

What's the cost? The Contributor tier agreement states clearly: Meta may use the prompts and completions you send to train future models.

This is the essence of the transaction. Meta is trading near-cost token pricing for real code, task trajectories, tool-calling sequences, and error messages—these continuous trajectories are far more valuable for training coding models than ordinary Q&A data. A coding agent doesn't produce scattered text; it produces a complete closed loop of "goal → code context → tool call → test feedback → final modification," which is precisely the data needed to train long-horizon task performance.

Beneath the veneer of a price disruptor lies a data collection gambit. Zuckerberg is willing to push this product at a moment when revenue is under pressure and the stock dropped 10% in a week; clearly, he's eyeing more than just that bit of API revenue.

Comparison with Chinese companies: This path has been walked for a while

If the "low price for data" model looks familiar, that's because it is—Chinese companies have been on this path for a long time, just with a different posture.

ByteDance's Trae is a representative Chinese coding agent. The China version's basic features are free, with built-in models like Doubao and DeepSeek. In SOLO mode, AI leads the entire process from requirement understanding to deliverable. Its free strategy is also ecosystem land-grabbing—trading free compute for developer habits and codebase data.

Alibaba's Qoder CN (formerly Tongyi Lingma) took a different path: after rebranding in May 2026, it fully upgraded to an agent matrix, offering code completion, multi-file modification, programming agents, and other capabilities, supporting multi-model switching among GLM, DeepSeek, Kimi, MiniMax, and others, and providing enterprises with dedicated VPC deployment compliance solutions. Its differentiation isn't low price, but enterprise-grade control—data never leaving the internal network is the hard requirement for industries like finance and government.

In comparison, Meta's Contributor tier shares the same origin as Chinese companies' "free for data" thinking, but with one key difference: transparency. Meta writes "your data will be used for training" into the Contributor tier's terms; developers make an informed choice. In contrast, the data usage explanations for Chinese free products are often buried deep within privacy policies spanning over a dozen pages.

Three points worth watching out for

First, all benchmarks are self-reported by the vendor. The Terminal-Bench, DeepSWE, and other data released by Meta come from official sources, and their evaluation methodology documentation admits that "evaluation configurations may disadvantage third-party models." Until independent reproduction exists, treat performance numbers as reference points, not conclusions.

Second, the weights are closed-source. Muse Spark 1.2 has no Hugging Face repository and cannot be self-hosted. This stands in stark contrast to Meta's proud Llama open-source path—the open-source inertia from the Llama era does not apply to this new family. For enterprise users, this means both data sovereignty and auditability take a hit.

Third, the Contributor tier has rate limits (60 requests per minute), suitable only for low-concurrency scenarios. To run large-scale agent tasks, you'll have to fall back to the Standard tier.

One-sentence summary

Meta entering the arena with Muse Code shows that the coding agent track is already crowded with giants—Anthropic, OpenAI, Google, Microsoft, and now Meta. The focus of competition has shifted from "whose model is smarter" to "whose pricing is smarter and who can get more real task data."

For developers, this is good news: a price war will only make tools cheaper, and model selection will shift from a strategic decision to a runtime parameter. But regarding the "$0.20 for your code trajectories" deal, be clear-eyed—cheap things often have a different price tag attached elsewhere.

Author: 唐悦玮 | An engineer who started from the backend and expanded to full-stack using AI.