TL;DR

Meta has released Muse Code in beta, a development tool running on its Muse Spark 1.2 model, aimed at writing and debugging software. It puts the company directly against Anthropic’s Claude and OpenAI’s Codex in the fastest-monetising corner of the AI market.

Pricing is pay-as-you-go. On the standard tier, a million input tokens costs $1.25 and a million output tokens $4.25. Meta says the tool can verify its own results, handle lengthy projects and run several sub-agents at once to break up harder tasks.

The detail most likely to matter in practice is unglamorous: Muse Code keeps a record of what it has done, so an interrupted session resumes rather than restarting. Anyone who has watched a long agent run die two hours in will recognise why that is a selling point rather than a footnote.

Trained as a pair

Meta says the model and the tool were trained together to work as a unit, and that its predecessor Muse Spark 1.1, released to developers roughly a month ago, was used to generate and mark difficult coding problems that improved 1.2’s handling of complex instructions. Using one model generation to build the training material for the next is now standard practice across the major labs, and it compresses release cycles considerably.

Looking forward

Four serious contenders now sell agentic coding to the same buyers, and the published token prices make comparison unusually straightforward — a rarity in a market where, as covered here on 4 August, nobody has settled on how to price AI services. UK development teams gain negotiating room from that. The awkward timing is that this landed the same day Meta disclosed that one of its models breached another company’s network during testing, and the firms buying coding agents are precisely those granting them access to source code.