Muse Code costs $1.25/$4.25 per million tokens - or $0.10/$0.20 if you let Meta train on your code. That gap is not a rounding error. It is a 12-to-21× price difference, and it is the most consequential engineering decision buried inside an otherwise routine beta launch.
Muse Code is Meta's terminal coding agent, launched in beta on August 5, 2026 and powered by the new Muse Spark 1.2 model.
It targets complex software engineering across large repositories - planning changes, writing code, validating results - and runs persistent async background agents for the whole session instead of spawning per task. On raw capability, the reviews are solid. The pricing model is where it gets strange.
What the two tiers actually mean for your team
The standard tier is straightforward: pay $1.25 per million input tokens and $4.25 per million output tokens. That is roughly in line with Claude Code on Sonnet 5 at $3/$15 and cheaper than Codex on GPT-5.6-Sol.
The contributor tier - $0.10 input, $0.20 output - is something different. It offers significantly lower costs in exchange for Meta using your prompts and completions to train its models. The real-world coding data it collects is scarce and superior to scraped or synthetic data because it is checkable against actual test suites. Every bug your agent finds, every PR it writes, every failing test it reasons through: that becomes a training signal Meta would otherwise have to generate itself.
Meta's advertising business rests on data users produced while doing something else. Muse Code applies the same structure to the most valuable workforce on the internet, and this time Meta is paying developers to take part
- by absorbing the cost gap instead of charging for it outright.
The non-obvious consequence: the cheaper your token bill, the more valuable your codebase probably is to Meta's next training run. Teams doing the most interesting, hardest-to-generate coding work - internal tooling, novel architectures, complex domain logic - are the ones subsidizing their access with the most signal.
How Muse Code compares to Claude Code and Codex on price
Before deciding which tier to use, it helps to see where Muse Code actually sits.
| Agent | Model | Input ($/1M) | Output ($/1M) | Open source harness? |
|---|---|---|---|---|
| Muse Code (standard) | Muse Spark 1.2 | $1.25 | $4.25 | No |
| Muse Code (contributor) | Muse Spark 1.2 | $0.10 | $0.20 | No |
| Claude Code | Claude Sonnet 5 | $3.00 | $15.00 | No |
| Codex (Terra default) | GPT-5.6 Terra | $2.00 | $12.00 | No |
| DeepSeek Harness | V4-Pro-0813 (API) | varies | varies | Yes (MIT) |
| OpenCode | any BYOK | varies | varies | Yes (MIT) |
Claude Opus 5 shipped July 24 at $5/$25 per million tokens with a 1M-token context window and became the default Opus in Claude Code.
Claude Sonnet 5 arrived June 30 at $3/$15, with introductory $2/$10 pricing through August 31. At introductory Sonnet 5 rates the gap between Claude Code and Muse Code standard narrows considerably. The contributor tier still wins on cost alone - until you account for IP.
The practical ceiling for contributor-tier use is clear: anything under an NDA, anything touching customer data, anything your employment contract defines as proprietary, and anything with a security classification. That covers the majority of what engineering teams at mid-size and larger companies actually do.
What DeepSeek Harness changes about this calculus
Six days after Muse Code launched, a different release landed that most coverage underweighted.
DeepSeek released DeepSeek Harness v0.1 in developer preview on August 13, 2026, giving agent builders a runnable open-source stack instead of another model announcement.
DeepSeek Harness adopts an "everything is a plugin" design philosophy: all agent capabilities - models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and UI - are composed of plugins that can be freely replaced and flexibly reconfigured.
DeepSeek Harness passed 135,000 GitHub stars in four days
- faster than anything in this category since Claude Code's public launch. On August 13, 2026, the source code was simultaneously released under the MIT License.
The practical implication for teams worried about the Muse Code data deal: every component of DeepSeek Harness - models, tools, session state, filesystems, sandboxes, the agent loop, orchestration, and the web experience - all sit behind composable Cordis services. You can wire it to any inference provider, including a self-hosted endpoint. No training data agreement, no proprietary model you cannot inspect.
DeepSeek-V4-Pro-0813, the production release of DeepSeek's flagship model, scores 87.9 on Terminal-Bench 2.1 on its own harness
- a meaningful jump from 72.1 for the April preview. That is a competitive number. The catch is that V4-Pro weights are not public; you are still sending prompts to DeepSeek's API, which carries its own set of data-residency questions for teams in regulated industries or under EU jurisdiction.
OpenCode (formerly sst/opencode) has 193,678 GitHub stars under MIT and shipped v1.18.13 on August 4, 2026
- a day before Muse Code launched. It remains the most-starred model-agnostic open harness, and it will pick up DeepSeek Harness as a backend option once the plugin ecosystem stabilizes.
Beagle in action#engineering, 11:02amMuse Code's open-source posture is more complicated than it looks
OpenAI released its Codex CLI as open source under the Apache 2.0 license and followed with its gpt-oss open-weight models; Google's Gemini CLI harness is likewise Apache-licensed. With Muse Code, Meta lands closest to the posture of Anthropic - whose Claude Code remains proprietary - while the company that once argued open source was the path forward now asks developers to pay per token for a model they cannot inspect, or to subsidize that access with their own data.
On August 10, Zuckerberg announced Meta will open-source Muse Spark 1.2's weights, alongside an essay arguing for American open-source AI. The weights are not out yet and no date has been given, so treat it as announced rather than shipped.
Muse Glimmer - Meta's 30B open-weight agentic model - arrived under Apache 2.0, the company's first fully open release since Muse Spark replaced the Llama family in April. In fact, Muse Glimmer launched with a more permissive license than Llama ever carried. That model fits a 24GB GPU, handles code and images, and is free to self-host with no contributor tier complication. It is not the frontier model Muse Spark is, but for teams with clear IP constraints it is the cleaner option today.
The decision tree is actually short. If your code would be uncomfortable in a competitor's training set, use the standard tier or an open harness. If it is open-source or genuinely throwaway, the contributor tier is fine - the economics are real. The mistake is treating it as a default just because the number is small.
Muse Code pricing: common questions
What is the difference between Muse Code's standard and contributor tiers?
Standard pricing is $1.25 per million input tokens and $4.25 per million output tokens - you pay the full rate and Meta does not use your data for training. The contributor tier costs $0.10/$0.20 per million tokens, roughly 12-21× cheaper, but grants Meta permission to use your prompts and completions to train future models.
Is Muse Code open source?
No. The Muse Code harness is proprietary, and Muse Spark 1.2 weights are closed at launch. Zuckerberg announced the weights will be open-sourced, but no date has been confirmed. Muse Glimmer - a separate 30B model - is available today under Apache 2.0 and runs on a single 24GB GPU.
How does Muse Code compare to Claude Code on cost?
At standard rates, Muse Code ($1.25/$4.25) is cheaper than Claude Code on the current Sonnet 5 introductory price ($2/$10 through August 31). Claude Code's default Opus 5 tier runs $5/$25. Neither Anthropic nor Meta offers a data-for-discount trade on their coding agents; Muse Code's contributor tier is currently unique to this market.
What is DeepSeek Harness and why does it matter here?
DeepSeek Harness is an MIT-licensed open-source agent runtime released August 13, 2026 with a plugin-first architecture where the model, loop, sandbox, and UI are all swappable. It reached 135,000 GitHub stars in four days. It matters as an alternative for teams that want to avoid both Muse Code's data terms and Claude Code's proprietary harness - bring your own model, your own keys, no training-data deal required.
Should regulated teams use the contributor tier?
No. Teams handling customer data, proprietary IP, or anything under an NDA, employment contract, or data-residency obligation should treat the contributor tier as off-limits by default. The standard tier or a self-hosted open harness like OpenCode or DeepSeek Harness are safer defaults when the data governance stakes are high.