Meta on August 5, 2026 launched its first coding agent, Muse Code, alongside the new Muse Spark 1.2 model that powers it. The move is a direct bid to enter the developer-tools market dominated by Anthropic and OpenAI. According to the head of Meta Superintelligence Labs, the company plans to compete mainly on price rather than on capabilities alone.
Key takeaways
- Muse Code is Meta's first coding agent, available in preview and installed with a single terminal command.
- Muse Spark 1.2 is the model powering the agent, trained alongside it for software-engineering tasks.
- Background agents stay active for the whole session, cutting latency on multi-step problems.
- The model optimized GPU kernels over 1,000 tool calls in 24 hours for NVIDIA Hopper chips.
- The cheapest tier is more than 10 times cheaper than pay-as-you-go but requires opting in to data use for training.
An agent in the terminal, a model underneath
Muse Code is a command-line tool that takes on complete software-engineering tasks — planning changes, writing code and validating results. Alexandr Wang, head of Meta Superintelligence Labs, described it in an interview as ready to work on large repositories right after a one-command install. Like the Claude assistant from Anthropic and Codex from OpenAI, it combines app building and agent management in one interface.
Beneath the agent runs a so-called harness — a model-management layer tuned for coding projects. Muse Code relies on background agents that are not spawned separately for each task but stay active throughout the session. According to Meta AI, this reduces latency on multi-step problems.
The runtime is built on a local event log, which — as Meta's documentation states — makes it „replay-exact” and restart-safe after a crash, allowing long-running tasks without losing state. The agent also ships bundled commands: „/plan” creates an approval-gated plan, „/grill” stress-tests it for weak points, and „/goal” drives the work to completion.
Muse Spark 1.2 — a model trained on code
The new model extends the Muse Spark line that Meta released in its API a month earlier. The company scaled the compute for training on coding tasks and broadened training-environment diversity. The model handles long-horizon work — whole repositories and large projects — and part of its training data came from a self-improvement loop in which the older Muse Spark 1.1 generated material for the newer version.
Meta showed results on the Terminal-Bench 2.1, DeepSWE 1.1 and its own Meta Internal Coding Bench benchmarks, but gave no direct numerical comparisons with Anthropic or OpenAI models. In one case study, the model tuned KDA and MLA kernels for NVIDIA Hopper chips over more than 1,000 tool calls across 24 hours, beating the baseline substantially.
Competing on price, not capabilities
Meta is differentiating Muse Code and the Muse Spark family against rivals mainly on price, not capabilities, Wang said. Developers can pay pay-as-you-go, but a „contributor” tier comes in „more than 10 times cheaper”, per Wang. In return, users on the cheapest plan must opt in to letting their data improve the model.
Meta's new model will also arrive on the OpenRouter platform, which hosts popular models, including open-weight offerings from Chinese labs such as DeepSeek AI and Z.ai. The company is also starting to accept requests for zero-data retention, under which it would not keep developer data — Wang called it an important feature for enterprise customers. That matters, because Meta draws roughly 98 percent of its revenue from ads built on user data.
Why it matters
Meta's pricing strategy is a bet on commoditizing coding agents. By placing its model next to Chinese open-weight models on one platform, it treats coding models as interchangeable infrastructure where unit cost counts. The „data-for-discount” tier leans on the group's advertising DNA but creates tension with enterprise trust — hence the parallel zero-data-retention offer. If Muse Code matches the leaders on quality, price pressure could hit the margins of paid assistants sold on capabilities rather than cost.
What's next
- Muse Code stays in preview and Wang declined to share user numbers, so adoption will only become clear from later data.
- Meta is accepting zero-data-retention?Zero-data retention: A mode in which the provider does not store data sent to the model or use it for training. requests, opening a path to enterprise deployments.
- Availability on OpenRouter will extend the model's reach beyond Meta's own developer page.
