Meta Launches Muse Wonder While Restricting Wonder Full to Muse Code
Meta launched its Muse frontier model family, but reserved its highest-performing model, Muse Wonder Full, exclusively for paid Muse Code agent subscribers starting at $50 per month. Developers using the standard Model API only receive Muse Wonder at $4 per million input and $20 per million output tokens, though Claude Code CLI can be pointed directly at Meta's endpoint via environment variables.
Why it matters
You can wire Claude Code CLI directly to Meta's $4/$20 Muse Wonder API today using seven standard environment variables without paying the $50 subscription fee.
TL;DR
- 01Meta's top-tier Wonder Full model is gated inside Muse Code subscriptions ($50 to $200 per month) and unavailable via API.
- 02The standard Muse Wonder API costs $4 per 1M input tokens and $20 per 1M output tokens, beating Opus 5.5 on DeepSWE 1.1.
- 03Claude Code CLI can run against Meta's API by setting the Anthropic base URL and model environment variables.
Key facts
- Muse Wonder API Input Price
- $4 per 1M tokens
- Muse Wonder API Output Price
- $20 per 1M tokens
- Muse Code Power Usage Plan
- $50 / month
- Muse Code Maximum Usage Plan
- $200 / month
- Wonder Full Composite Score
- 59.1 (self-reported)
- Muse Wonder API Composite Score
- 54.3 (self-reported)
Model Fragmentation and Agent Lock-in
Meta has introduced a sharp division between model inference and agent environments. The standalone Model API only exposes muse-wonder-1.0 at $4 per million input tokens and $20 per million output tokens. Meanwhile, the flagship Muse Wonder Full model is exclusively packaged into Meta's Muse Code agent tool under Power Usage ($50 monthly) and Maximum Usage ($200 monthly) plans. On Meta's internal composite scorecard, Wonder Full leads at 59.1, whereas the standalone API version scores 54.3, sitting behind Opus 5.5 (55.2).
Benchmark Discrepancies and Features
While Wonder Full achieves 79.0 on DeepSWE 1.1 and 57.5 on SWE-Marathon under extra-high reasoning, the accessible API variant Muse Wonder remains competitive on engineering benchmarks. It records 78.8 on DeepSWE 1.1 (versus 74.2 for Opus 5.5) and 10.0 on Harvey's LAB-AA benchmark (versus 4.2). The Muse agent environment introduces Active Memory for persistence across conversations and Active Focus for sustained autonomous iterations across multi-agent worktrees.
Routing Claude Code CLI to Meta API
Developers do not need to install the proprietary muse CLI or purchase a subscription to benchmark Meta's coding capabilities. Because Meta serves an API compatible with standard SDK structures, Claude Code CLI can be pointed directly at api.meta.ai by overriding provider environment endpoints in shell configurations.
Try it in 2 minutes
export ANTHROPIC_BASE_URL="https://api.meta.ai"
export ANTHROPIC_AUTH_TOKEN="$MODEL_API_KEY"
export ANTHROPIC_MODEL="muse-wonder-1.0"
export ANTHROPIC_DEFAULT_OPUS_MODEL="muse-wonder-1.0"
export ANTHROPIC_DEFAULT_SONNET_MODEL="muse-wonder-1.0"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="muse-wonder-1.0"
export CLAUDE_CODE_SUBAGENT_MODEL="muse-wonder-1.0"bash
✓ When to use
- Use when you want to benchmark frontier DeepSWE agent performance without abandoning Claude Code's terminal UX.
- Use when setting up multi-agent evaluations across git worktrees to avoid per-seat platform lock-in.
✕ When NOT to use
- Do not subscribe to Muse Code if your workflow relies strictly on standard IDE extensions without standalone CLI terminal agents.
- Avoid Muse Wonder API for codebases requiring broad multilingual reasoning where Claude Opus 5.5 retains higher composite benchmark marks.
What to do today
- Export ANTHROPIC_BASE_URL and ANTHROPIC_MODEL variables in your shell profile to test Muse Wonder inside Claude Code CLI.
- Isolate agent tasks into dedicated Git worktrees before running multi-agent tools to prevent file collision.
- Evaluate whether DeepSWE performance justifies adding an external provider before purchasing proprietary agent seats.