Breaking

Meta Releases Muse Code Beta to Tackle Long-Running Software Development With AI Agents

October 3, 2026 at 23:30 EDT

Software Dev & Coding
AI Agents
Foundation Models

Meta released Muse Code in beta on August 5, 2026, pairing a terminal coding agent with Muse Spark 1.2 for long-horizon software engineering work. The product is Meta's clearest move into AI coding agents, where tools are shifting from code completion to systems that plan, edit, test and coordinate work across repositories.

Aug. 5, 2026 · Meta

Muse Code enters the coding-agent race

Meta paired Muse Code with Muse Spark 1.2 for long-horizon software tasks that plan, edit, verify and recover across repositories.

1,048,576

token context window

1,000+

tool calls in a kernel case study

24h

maximum reported run length

$0.20

Contributor output price per 1M tokens

Output-token price split

Per 1M output tokens. Each block is about $0.20, so height shows the scale of the discount.

Agent loop

Plan

→

Edit

→

Verify

→

Resume

Why it lands

Persistent subagents reduce repeated repo exploration.
A local event log makes actions auditable and recoverable.
Contributor pricing makes experiments inexpensive.

What still needs proof

Benchmark results combine model and agent harness.
The discounted tier carries a training-data tradeoff.
Long repo work still needs human review.

The strategic signal

Muse Code shows Meta treating coding agents as full runtimes around models: tools, memory, approvals, workers and recovery, not just a chatbot that writes snippets.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.