Meta shipped Muse Code — and Wang says it competes on price, not capability

Meta shipped its first coding agent today. Alexandr Wang, Meta's chief AI officer, announced Muse Code in beta, powered by a new model — Muse Spark 1.2 — that was developed and trained alongside the agent itself.

Wang's own description: "Muse Code is a terminal coding agent, like many of the other coding agents on the market. So you can install it with one command and then use it to take on complete software engineering tasks across a wide variety of use cases, planning changes, writing code, validating the results."

The interesting part is the positioning, and Wang says it plainly rather than leaving it to be inferred: Meta is differentiating on price, not on bleeding-edge capability, against Claude Code and Codex.

The prices back that up. Pay-as-you-go matches the Muse Spark 1.1 rate — $1.25 per million input tokens, $4.25 per million output. Then there is a contributor tier that Wang describes as "more than 10 times cheaper than than even the pay-as-you-go tier", with the trade stated openly: "as part of that, you opt-in to help improve the model in an industry standard way in-line with all the other coding agents." Read that as roughly $0.12 in and $0.42 out if the multiple is literal, in exchange for your code becoming training data. That is a real option for a hobby project and a non-starter for most client work, and it is refreshing that the price of the discount is named rather than buried.

Underneath sits a harness tailored for coding. Models and harnesses can be swapped, but Wang says the two were "co-developed with one another" so you "get the best performance if you use them together." Meta is also now accepting zero-data-retention requests, which Wang calls "a big enterprise feature that is important for folks" — and which is the necessary counterpart to a contributor tier that trains on your work.

One capability claim we are marking as reported rather than confirmed, because it comes from the launch-day summaries and not from Meta directly: that every model call, tool run and edit is written to a local event log before execution, so a crash mid-task resumes from the log instead of starting over. If that holds up it is the most substantive engineering idea in the release — durable execution for an agent, rather than a longer context window — and we will chase it to the docs.

Two things worth noting beyond the launch itself.

Muse Spark 1.2 is in no third-party catalog. We checked OpenRouter and models.dev: 1.1 has been listed by nine providers since April at exactly these prices, and 1.2 appears nowhere. That is what "get it through Meta Model API" means in practice, and it means benchmark comparisons will lag the release by however long resale takes.

And this is the second coding harness announced in two days. DeepSeek is recruiting beta testers for its own Harness framework, also aimed at Codex and Claude Code. Two labs deciding within 48 hours that the model is no longer the product — the harness around it is — is a clearer signal about where this is going than either announcement alone.

A note on our own coverage. This model was invisible to our radar. We tracked Muse Spark 1.1 as live and Muse Spark 2 as armed, with no pattern in between, so the entire 1.x point-release track was unwatched and 1.2 matched nothing at all. That is the same failure shape as Gemma 4 in April: not an entry that was missing, an entry that was structurally unable to match what actually shipped. Muse Spark 1.2 and Muse Code are now on the board, and 1.3 through 1.9 are armed so the next one is caught rather than researched.