Meta Muse Code launches as a cheaper rival to Codex and Claude Code
Meta entered the agentic coding race on August 5 with Muse Code, a beta AI coding agent built on the new Muse Spark 1.2 model. It is designed to plan, write, test, and coordinate code changes across large software projects, competing directly with OpenAI’s Codex and Anthropic’s Claude Code.
What Muse Code Does
Muse Code runs a main agent alongside persistent background agents that stay active for the entire session. Because those background agents retain what they have already discovered, they do not need to rediscover the project each time a new step arises. A local event log records every model call, tool use, approval, and code change, so if Muse Code crashes mid-assignment it can reconstruct and resume from where it stopped rather than starting over.
Three built-in commands shape how the agent works: /plan turns an assignment into a proposed plan that requires approval before execution, /grill challenges that plan for weaknesses, and /goal keeps the system directed toward a defined outcome. Meta describes the runtime as fully auditable because all actions stay visible and traceable in the event log.
Pricing and the Data Trade-Off
This is where the decision gets real. Meta offers two tiers:
- Standard tier: $1.25 per million input tokens, $0.15 for cached input, $4.25 per million output tokens. Meta says customer usage will not be used to improve its products.
- Contributor tier: $0.10 per million input tokens, $0.002 for cached input, $0.20 per million output tokens. In exchange, Meta may use your activity for product improvement.
For context, GPT-5.6 Terra runs $2.50 and $15 per million tokens, while Claude Opus 5 costs $5 and $25. Muse’s standard tier is meaningfully cheaper. The contributor tier is dramatically cheaper, but that discount comes with a specific policy trade-off your security or data-governance team needs to sign off on.
What the Benchmarks Actually Show
Meta compared Muse Spark 1.2 with Muse Code against GPT-5.6 Terra with Codex and Opus 5 with Claude Code across four evaluations. The results are mixed:
- Terminal-Bench 2.1: Muse scored 82.9%, finishing between Opus 5 at 86.7% and GPT-5.6 Terra at 81.8%.
- DeepSWE 1.1 (long-horizon work inside real codebases): Muse scored 59.3%, trailing GPT-5.6 Terra at 64.8% and Opus 5 at 65%. This is the test most directly tied to the sustained work Meta built Muse to perform.
- Meta Internal Coding Bench: Opus 5 scored 79.4%, Muse scored 70.6%, GPT-5.6 Terra scored 65.4%.
- GDPval-AA v2: Opus 5 scored 1,852, Muse scored 1,631, GPT-5.6 Terra scored 1,577.
Muse beat GPT-5.6 Terra on three of four comparisons, but Opus 5 outscored Muse on all four. The DeepSWE gap is worth noting: Muse trailed both rivals on the test most closely tied to the long-running codebase work it is marketed for.
The Closed Model Question
Muse Spark 1.2 is a closed model. Unlike Meta’s Llama releases, you cannot download the weights and run it on your own infrastructure. Access goes through Muse Code, Meta’s Model API, or OpenRouter. Teams that require on-premise model hosting will need to look elsewhere. Muse Code is available to install directly from the command line on macOS and Linux.
The Operator Takeaway
If your team runs long AI-assisted coding sessions regularly, Muse’s lower token prices could reduce costs meaningfully at scale. The contributor tier’s discount is steep enough to be tempting, but it is a data policy decision, not just a pricing decision. Muse Code remains in beta with no broad production-reliability track record yet. Test it on real work your team does every day before committing.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
