Meta released Muse Code, its first artificial intelligence coding agent, on Wednesday, pricing the underlying model at $1.25 per million input tokens to undercut Anthropic and OpenAI.
Key Points:
- Meta launched Muse Code in beta, its first coding agent, running on the new Muse Spark 1.2 model.
- Standard API access costs $1.25 per million input tokens and $4.25 per million output tokens, with a far cheaper tier for users who share their data.
- The company's models still trail OpenAI and Anthropic on the DeepSWE coding leaderboard.
Muse Code Beta Targets Large Repositories
The agent lives in a terminal, installs with a single command, and handles planning, code writing and validation across large repositories rather than suggesting one line at a time. Chief executive Mark Zuckerberg said it takes on complete software engineering tasks. Muse Code is in beta and runs on Muse Spark 1.2, a coding-focused model update that Meta developed and trained alongside the agent itself.
When a job grows too big, Muse Code splits it among sub-agents that work in parallel inside isolated worktrees, so the developer's own working copy is never touched. In one internal test the agent built six features for a game at once, without collisions.
Also Read: iPhone 18 Pro Max Could Be First With Apple's 2nm Chip, 15% Faster
Alexandr Wang Prices Below Claude Code And Codex
Alexandr Wang, who runs Meta Superintelligence Labs, said the company is competing on cost rather than raw capability, and called the tool a strong option for many everyday workflows.
Developers pay $1.25 per million input tokens and $4.25 per million output tokens on the standard tier.
A contributor tier costs more than 10 times less, though anyone taking that discount must let Meta observe how they use the agent and feed the results back into training.
Customers who pay the higher rate can keep their work out of Meta's training pipeline entirely.
Meta shares tumbled last week after the company issued a light revenue forecast and reported dwindling free cash flow for the second quarter, a stretch that left investors asking when the spending pays off. Coding has turned into one of the most profitable uses of generative AI so far, and Zuckerberg needs a way to earn back the hundreds of billions he has committed to data centers.
Muse Spark Still Trails Rivals On DeepSWE
The pitch still has to survive comparison. Muse Spark 1.1 scored 53% on the DeepSWE 1.1 leaderboard, an evaluation that puts coding agents through 113 long-running software engineering tasks under the same harness. GPT-5.6 Sol reached 73% on that test after its mid-July release, while Claude Opus 5 hit 74%.
Meta introduced Muse Spark 1.1 in early July, describing it as a marked improvement on the first version, and researchers inside the company were already using it daily while it trailed the strongest rivals.
Weeks later, a vice president asked engineers across the company to submit at least one weekly code fix through MetaCode, the internal agent, so those corrections could train what comes next.
Read Next: Bitcoin Could Turn Extreme Fear Into Fuel For A $75K Rally