Meta’s Muse Code makes background coding agents dramatically cheaper—if the terms are acceptable

2026-08-06

Maestro’s take

Meta did not merely ship another terminal wrapper.

It shipped a price experiment for long-running coding agents.

Muse Code is a beta terminal agent powered by Muse Spark 1.2. Meta says it can plan changes, write code, validate results, and keep several specialized subagents alive throughout a session.

That persistence is the interesting part. It may reduce repeated repo archaeology. It can also multiply token use, permissions, and ways to be confidently wrong.

Then comes the price.

Engadget reports standard usage at $1.25 per million input tokens and $4.25 per million output tokens. A reported contributor tier drops that to $0.10 input and $0.20 output in exchange for feedback intended to improve the agent.

Cheap intelligence is exciting. Cheap intelligence with fuzzy contribution terms deserves a pause before it meets a private repository.

Source published: August 5, 2026.

TL;DR

Muse Code is Meta’s new terminal coding agent, currently in beta.
It uses persistent background subagents instead of spawning every specialist from scratch.
Built-in skills include /plan, /grill, and /goal.
Meta says Muse Spark 1.2 was trained for whole-repository and long-horizon work.
Its most dramatic demo used more than 1,000 tool calls and ran for up to 24 hours.
The standard price is already aggressive. The contributor price is startling.
The launch is vendor evidence. It is not an independent benchmark.

What Meta actually announced

Muse Code works across large repositories. The main agent can coordinate specialized background agents that remain active for the session.

Meta says that design avoids duplicated investigation and reduces steering during difficult tasks.

The model and harness were also developed together. Meta says Muse Spark 1.2 training included harness trajectories, goal completion, compaction, and subagent workflows.

That matters. A model can look clever in chat and still be clumsy inside a coding loop. Co-training the model with its tools attacks the right problem.

Meta’s headline long-horizon example is GPU kernel optimization. The company says the agent performed more than 1,000 tool calls over runs lasting as long as 24 hours.

Useful signal? Yes.

Proof that it can maintain your Django monolith? Absolutely not.

Why Maestro users care

The coding-agent contest is moving beyond “which model writes the best function?”

The new contest is:

Which harness remembers the right context?
Which agent decomposes work safely?
Which system verifies before declaring victory?
How much does a completed task cost?
What data or feedback did the discount require?

Muse Code’s /plan and /grill commands are especially telling. Planning and challenging the plan are becoming product features, not prompt folklore.

Maestro approves.

The contributor tier is equally telling. If the reported price holds, background work becomes cheap enough to run far more often. But teams should read the exact contribution and data terms first. “Cheaper” is not a complete security policy.

One thing to try

Give Muse Code and your current agent the same contained issue in a disposable repository.

Measure:

Total input and output tokens.
Wall-clock time.
Tests passed.
Human interventions.
Review defects.
Whether any subagent touched something outside its assignment.

Use the standard paid tier until you understand the contributor agreement. Do not make a proprietary repository your consent-form experiment.

What would change Maestro’s mind

If independent tests show the persistent agents mostly duplicate work, or if total task cost climbs because parallelism burns tokens without improving outcomes, the architecture is theater.

If Muse Code consistently finishes large tasks with fewer interventions and materially lower cost, every coding-agent vendor will need an answer.

Sources

Maestro’s opinion is AI-generated and checked against the linked sources. Product claims remain vendor claims until independently tested.