Meta Ships Muse Code AI Agent and Spark 1.2

Meta entered the AI coding agent market on August 5, 2026 with two simultaneous launches: Muse Code, its first terminal-based coding agent, and Muse Spark 1.2, the third Muse model release in four months. On Meta's own published benchmarks, Muse Code trails Claude Opus 5 across all three published coding benchmarks and beats GPT-5.6 Terra on two of three, trailing only on DeepSWE 1.1 — and the default pricing tier sends developer code straight into Meta's training pipeline.

What happened

Meta shipped two products on the same day:

Muse Code (beta):

• Terminal-based CLI coding agent, installed via a one-line curl command
• Persistent async background agents — keep running after the terminal closes, building repo context over time
• Multi-file editing across large repositories with codebase-aware context
• Fans out to parallel sub-agents in isolated Git worktrees for large tasks
• Crash-safe event log: resumes exactly where it stopped if the agent dies mid-task
• Powered by the co-trained Muse Spark 1.2
• Available via API, OpenRouter, and Meta Model API — macOS and Linux only

Muse Spark 1.2:

• Third Muse model in 4 months (Spark 1.0 → 1.1 → 1.2)
• 82.9% on Terminal-Bench 2.1 and 59.3% on DeepSWE 1.1 — second only to Claude Opus 5 on the former, third on the latter
• Co-trained with the Muse Code harness for best-in-tool agent performance
• Standard pricing unchanged: $1.25/$4.25 per 1M tokens with 1M context
• Contributor tier: $0.10/$0.20 per 1M tokens — 12.5x cheaper on input, but your prompts and completions flow into Meta's training pipeline
• Native subagent orchestration, MCP/tool support

Why it matters

Meta is the last major US AI lab to ship a coding agent — and it arrives to a market that has already moved past "first agent" to "which agent ships the fastest PRs."

The benchmark gap is real. On Meta's own published numbers, Muse Code with Spark 1.2 lands at 82.9% on Terminal-Bench 2.1 (3.8 points behind Claude Opus 5) and 59.3% on DeepSWE 1.1 (behind both Opus 5 and GPT-5.6 Terra). All published results are vendor-run in Meta's own framework — no third-party reproduction exists yet.

This is a market where Claude Code runs dynamic workflows with parallel subagents, OpenAI Codex handles long-running cloud sandbox tasks autonomously, and GitHub Copilot is embedded across every major IDE. Muse Code's differentiator — persistent background agents with crash recovery — is genuinely novel architecture, but both Claude Code (Dynamic Workflows) and OpenAI Codex (cloud sandbox) already offer asynchronous, long-running agent sessions.

The faster story is Meta's model cadence. Three models in four months is a pace no other lab matches, and Spark 1.2 closes some of the gap with frontier models. At $1.25/$4.25 per 1M tokens, it remains the cheapest closed-weight coding model with 1M context.

The training-data trap

The most consequential detail is the pricing structure. The standard tier at $1.25/$4.25 does not train on your data. The contributor tier at $0.10/$0.20 — the default that Meta's docs surface first — sends developer code and prompts into Meta's training pipeline in exchange for a 12.5x input discount.

For enterprise teams, this is a non-starter: proprietary codebases can't flow into a competitor's training data. Meta allows opt-out by selecting the standard tier, but it's a conscious step teams must take, not the default. This mirrors the broader industry divide: Anthropic and OpenAI have committed not to train on API customer data; Meta is betting enough developers won't care enough to switch tiers.

What changes for you

1. If you're on Muse Spark 1.1: Upgrade to 1.2 is a clear win — better benchmarks at the same $1.25/$4.25 pricing with improved tool-use and agentic performance.
2. If you're evaluating Muse Code: Treat it as a beta. The persistent agent model and crash recovery are promising architecture, but the benchmark gap is real. Test against your actual workflows before committing.
3. Enterprise teams: Audit your data flow before adopting Muse Code. The contributor tier's training-data pipeline means your proprietary code is Meta's training fuel unless you explicitly select the standard tier.
4. If you need a production coding agent today: Claude Code, OpenAI Codex, or GitHub Copilot remain the safer bets — they're GA products with proven benchmarks and clear data policies.

FAQ

What is Muse Code?
Muse Code is Meta Superintelligence Labs' terminal-based AI coding agent, launched in beta on August 5, 2026. It handles complete software engineering tasks across large repositories — planning changes, writing code, and validating results — powered by the co-trained Muse Spark 1.2 model.

How does it compare to Claude Code and OpenAI Codex?
On Meta's own benchmarks, Muse Code trails Claude Opus 5 across all three published coding benchmarks and beats GPT-5.6 Terra on two of three, trailing only on DeepSWE 1.1. Its architectural standout — persistent background agents that survive terminal close and crash recovery — is genuinely novel, but both Claude Code and Codex already offer asynchronous, long-running agent sessions. The pricing is aggressive on the contributor tier ($0.10/$0.20), but that tier trades your code for the discount.

Should I use the contributor tier?
Only if you're prototyping, experimenting, or working on non-proprietary code. The contributor tier sends your prompts and completions to Meta's training pipeline. For any production or proprietary work, select the standard tier ($1.25/$4.25), which keeps your data out of training.