On August 5, 2026, Meta released Muse Code, its first coding agent, along with a model called Muse Spark 1.2. It’s a terminal beta, it installs with a single command, and it targets large repositories. Two ideas inside deserve attention, the price is aggressive, and the benchmarks can’t be trusted.
Installation on macOS or Linux:
curl -fsS https://dev.meta.ai/install.sh | bash
What actually shipped
Muse Code is a terminal agent powered by Muse Spark 1.2, a new version of the model we wrote about when it premiered. You install it on macOS or Linux with a single curl command. It does what the competition does: it plans changes, writes code, and verifies the result. Meta is building it for the most demanding development work, and it lets multiple agents work on one task at the same time: some implement, others review in the background.
The project is run by Meta Superintelligence Labs under Alexandr Wang, whom Zuckerberg brought in last year. With this, Meta steps onto a field where Anthropic with Claude Code and OpenAI with Codex have been playing for years. It arrives late, and it knows it.
How it differs from Claude Code and Codex
Two things are technically interesting. The rest is catching up.
Persistent subagents. Claude Code spawns subagents for an individual task; each runs in its own context and returns only a result to the main agent. Muse Code does it the other way around: background agents stay alive for the whole session, keep their context, and decide on their own when to report back. According to Meta, this saves repeated information gathering and latency. Whether that holds up in practice will only show in real-world use.
Event-log runtime. This is what caught my attention most. Every model call, every tool run, every edit is written to a local log as the single source of truth. After a crash, the agent picks up exactly where it left off. Claude Code has checkpoints, and a session survives closing the terminal, but “deterministically replay the entire history and resume from the point of failure” is a stronger guarantee. For runs that last hours, it counts.
The other two novelties are more cosmetic. The bundled skills /plan, /grill and /goal (plan it, stress-test the plan, drive the goal to completion) are nice, but conceptually it’s the plan mode and skills that both Claude Code and Codex already have. And the co-training of the model with the harness, which Meta emphasizes, is something Anthropic and OpenAI do too. Meta just admits it out loud.
Where Meta clearly loses is the ecosystem. Muse Code is a beta, terminal only, macOS and Linux only, one model. There’s no IDE extension, no cloud agent, and no mention of open source (Codex CLI is Apache-2.0 and has over 75 thousand stars on GitHub). Claude Code and Codex have years of iteration behind them and mature ecosystems with MCP, hooks and plugins. In developer tools, Meta has nothing behind it.
Pricing: the main weapon
Meta didn’t price the model above the competition. It’s attacking on price. It doesn’t dare to attack on benchmarks.
| Model | Input ($/M tokens) | Output ($/M tokens) |
|---|---|---|
| Muse Spark 1.2 | 1.25 | 4.25 |
| Claude Sonnet 5 | 2 | 10 (until Aug 31, then 3 / 15) |
| Claude Opus 5 | 5 | 25 |
| GPT-5.6 Sol | 5 | 30 |
The standard tier costs $1.25 per million input tokens and $4.25 per million output tokens, and cache comes to $0.15. On output, Muse is roughly six times cheaper than Opus 5 and seven times cheaper than GPT-5.6 Sol. If you want to save even more, there’s the contributor tier, which is about ten times cheaper, but you pay with your data: Meta uses it for training. The model also runs through OpenRouter, and the Meta Model API understands both OpenAI and Anthropic formats, so switching from the competition costs nothing.
The more interesting comparison is with the Claude Max subscription, because these are two different payment models. Muse Code has no subscription; you pay for the tokens you consume. Claude Max is a flat fee, $100 a month for 5x the capacity, $200 for 20x, and Anthropic subsidizes it to shield heavy users from a token bill. One documented case shows this: a developer burned through 10 billion tokens over eight months, would have paid over $15 thousand on the API, and on Max at $100 a month paid around $800 in total. A 93% saving. No pay-per-token price list competes with that, not even Muse with its six-times-cheaper output, because at high consumption a subsidized flat fee beats a metered rate by an order of magnitude. For a lighter user, on the other hand, Muse in pay-per-token mode comes out cheaper than Max 5x. The break-even point lies somewhere around full-day work every day.
Where to be skeptical
Meta picked the benchmarks itself: Terminal-Bench 2.1, DeepSWE 1.1 and Meta Internal Coding Bench. The last one is internal and unverifiable. SWE-bench Verified, which is the standard for these comparisons, is missing. Muse Spark 1.1, meanwhile, did not reach the top in independent tests (Terminal-Bench 2.1: 80.0 versus 85.8 for GPT-5.6 Sol and 84.6 for Opus 5). Whether 1.2 closed that gap on code can’t be credibly determined from an internal bench.
The case study with GPU kernel optimization (over 1,000 tool calls, up to 24 hours of runtime) looks good, but it’s a narrow task on Nvidia Hopper. It doesn’t follow that the model would handle ordinary application development just as well.
And watch out for availability. The Meta Model API launched for the US only. The 1.2 announcement does talk about expanded global access, but before you build production on it, verify that it applies to the Czech Republic as well.
What to take away
Muse Code is a cheap and technically inventive attempt, but it lacks maturity and credible numbers. Behind the persistent subagents and the event-log runtime stands a real architecture. The price is a clear advantage if pay-per-token suits you and you don’t mind entrusting your code to Meta. The rest is so far a promise. The deciding day will be when someone independently measures how many tokens Muse consumes on the same task as Claude Code. Until then, it’s a nice beta with a good price list.