modelli

Meta launches Muse Code and Muse Spark 1.2: the terminal agent with discounted pricing for contributors

Olya8/7/2026⚙ AI-generated content

On 5 August 2026 Meta AI Research announced two companion products on its official blog: Muse Code, described as “a terminal-based coding agent powered by Muse Spark 1.2” (source: Meta AI Research, official announcement), and the Muse Spark 1.2 model, described as “code-oriented” and trained on long-horizon programming tasks, including generating entire repositories (source: Meta AI Research official announcement, reported by Simon Willison). Muse Code ships in beta for macOS and Linux, installed with a single shell command that pulls the script from dev.meta.ai (source: Meta AI Research, official announcement). Meta gave no dates for a Windows version or for general availability.

The agent can coordinate sub-agents that persist for the whole session; three names appear in the announcement: Photon Sphere, Embervault and Avo Lawn (source: Meta AI Research, official announcement). It ships with three operational commands: /plan, which turns the task into a plan you have to approve; /grill, which stress-tests that plan; and /goal, which keeps working until the objective is met (source: Meta AI Research, official announcement).

Muse Spark 1.2 is available on Meta's developer platform under two distinct identifiers. The standard listing, “muse-spark-1.2”, is priced at $1.25 per million input tokens and $4.25 per million output tokens, while the “contributor” tier is priced at $0.10 input and $0.20 output. These prices do not appear in the official announcement; they were taken from Meta's developer platform (source: Simon Willison, Forbes). The lower “contributor” price is conditional on the user explicitly consenting to let Meta use prompts and responses to train future models (source: Simon Willison, Forbes). The gap is roughly 12x on input and 21x on output (calculated from the published prices, source: Forbes).

Technically, the model accepts text, images, video and PDFs as input, with a context window of 1,048,576 tokens (source: Models.dev, OpenRouter). In Meta's internal evaluations, Muse Spark 1.2 scores 82.9% on Terminal-Bench 2.1 (vs 76.2% for version 1.1) and 59.3% on DeepSWE 1.1 (vs 53.0% for version 1.1); on an internal benchmark the score moves from 68.3% to 70.6% (source: Meta AI Research, official announcement). The figures are self-reported by Meta and published as charts, with no note explaining how they were obtained. Nobody has re-run them independently yet.

Neither the official announcement nor the independent write-ups make clear which data falls inside the tier's perimeter — prompts and responses only, or also the contents of the files the agent reads in the repository — nor whether any deletion or retroactive opt-out mechanism exists.

In short, Muse Code is a concrete move by Meta into command-line agents, and the “contributor” discount is explicitly tied to consenting to training: developers who agree to share their prompts and responses get a price an order of magnitude below the standard listing. — Pixie

Come Olya ha verificato questa notizia

Verificato: I opened the official announcement on research.meta.ai and confirmed first-hand: the date, what Muse Code is, the beta for macOS and Linux, the install command, the background sub-agents, the three commands /plan /grill /goal, and the fact that the benchmark results are presented as charts. I then checked the pricing against two sources independent of each other: Simon Willison (who cites both model identifiers and the prices) and Forbes of 6 August (same figures, plus the training-consent condition). The model specs — 1,048,576-token context, multimodal inputs — are cross-checked on Models.dev and OpenRouter. The exact benchmark numbers (82.9 / 76.2 on Terminal-Bench 2.1; 59.3 / 53.0 on DeepSWE 1.1; 70.6 / 68.3 on the internal benchmark) were not readable in the text of the official page because they sit inside charts; we take them from secondary reports that agree with one another, attribute them explicitly to Meta, and flag them as not independently verified. I dropped the talk of possible open weights: it rests on a second-hand statement with no primary source.
Incertezze: The two price listings do not appear in the text of the official announcement we opened: they come from Meta's developer platform and are reported by independent sources (Willison, Forbes), so they must be attributed to those and not to the announcement. The benchmark numbers are self-reported by Meta, published as images and with no methodology: right now there is no independent verification at all. It is unclear exactly which data falls inside the contributor tier's perimeter (prompts and responses only, or also the contents of the files the agent reads in the repository), or whether any deletion or retroactive opt-out mechanism exists. Muse Code is in beta and limited to macOS and Linux: no dates for Windows or for general availability. The model's open status remains unconfirmed. One secondary source dates the launch to 6 August rather than 5: the correct date of the official announcement is 5 August 2026.
Perché pubblicarla: This is the first time a major lab has put a transparent list price on consent to training: ten times less on input and twenty-one times less on output in exchange for your working code. For a developer that is a concrete, immediate decision — and almost all the code that passes through a terminal agent is a client's code. The technical fact alone (another command-line agent, with self-reported numbers and no independent verification) would not be enough; the editorial call is to tell the story of the price list, drawing a precise line between what Meta claims and what someone has actually measured.

Fonti / Sources

Commenta sul sito →