Blog · October 13, 2026 · Rich Chetwynd Meta announced Muse Wonder with strong reported scores on long-horizon coding and agent workflows. Claude Opus 5.5 still leads some coding benches. After you build with any model, Croft is where your team runs the app.
Quick answer #
Meta announced Muse Wonder today: two models, Muse Wonder and Muse Wonder Full, aimed at personal and professional agents. On Meta’s reported table (max settings), Muse Wonder Full leads several long-horizon coding and agent-workflow benches — DeepSWE 1.1 at 79.0, SWE-Marathon at 57.5 (xhigh), AutomationBench-AA at 45.4. Claude Opus 5.5 still leads Terminal-Bench 4.0-AA (59.6), ProgramBench (** 70.5**), and the overall Artificial Analysis Index (** 58**).
Those figures are Meta’s reported results, not an independent league table. For teams and developers vibe coding an app, the practical split is: pick the model that fits the job; then put the app somewhere private the team can actually run.
What Meta announced #
From Meta’s Muse Wonder research post:
- Two models: Muse Wonder andMuse Wonder Full
- Active Memory (carry relevant context across conversations) andActive Focus (keep working a hard goal, try other approaches, check in before important actions)
- Domain work Meta highlights: law, finance, and health, plus frontier coding
- Muse Wonder is available today through theMeta Model API andMuse Code
- Muse Wonder Full is available today forMuse Code subscribers
- Meta says Muse Wonder will soon power experiences in Muse , rolling out first to subscribers, and will powerartifact generation
Meta did not publish pricing, a general consumer rollout date, or a Vibe Code Bench score in that post.
What Meta’s numbers mean for vibe coding #
Vibe coding is describing an app and steering until it works. Long-horizon coding benches are a useful proxy for that: the model has to keep requirements in mind, change a codebase, test, and recover. Agent and workflow benches matter if the app has to do a job, not just render a screen.
On Meta’s comparison table (Muse Wonder Full max, Muse Wonder max, Claude Fable 5.1 max, GPT-6 Astra max, Claude Opus 5.5 max):
Where Meta reports Muse Wonder Full ahead
- DeepSWE 1.1 (long-horizon coding):79.0 vs Opus 5.5 at 74.2, Astra at 74.1, Fable 5.1 at 67.4. Muse Wonder (the smaller of the two) is78.8 .
- **SWE-Marathon** (ultra-long-horizon engineering):**57.5** (xhigh) vs Opus 5.5 at 46.3. Muse Wonder is**53.1** .
- **AutomationBench-AA** (end-to-end business workflows):**45.4** vs Astra 41.6 and Opus 5.5 at 41.2.
- **Harvey LAB-AA** (legal agent work):**13.1** vs Opus 5.5 at 4.2.
- MemoryBench (remembering context): Muse Wonder85.6 , Muse Wonder Full85.5 , vs Opus 5.5 at 81.5.
Where Meta’s table still has Claude Opus 5.5 ahead
- Terminal-Bench 4.0-AA (agentic terminal coding):59.6 vs Muse Wonder Full at 57.1. GPT-6 Astra is 59.1.
- ProgramBench (from-scratch programming):70.5 vs Muse Wonder Full at 59.0 — a wide gap on this bench.
- **Artificial Analysis Index** (overall):**58** vs Muse Wonder Full at**56** .
- **HealthBench Pro** :**65.6** (Astra 65.3) vs Muse Wonder Full at 62.3.
- **OSWorld V2.1** (partial):**81.8** vs Muse Wonder Full at 78.4.
A compact slice of the coding and overall scores Meta reported:
| Benchmark | Muse Wonder Full | Claude Opus 5.5 | Lead on Meta’s table |
|---|---|---|---|
| DeepSWE 1.1 | 79.0 | 74.2 | Muse Wonder Full |
| SWE-Marathon | 57.5 | 46.3 | Muse Wonder Full |
| AutomationBench-AA | 45.4 | 41.2 | Muse Wonder Full |
| Terminal-Bench 4.0-AA | 57.1 | 59.6 | Claude Opus 5.5 |
| ProgramBench | 59.0 | 70.5 | Claude Opus 5.5 |
| Artificial Analysis Index | 56 | 58 | Claude Opus 5.5 |
Honest read for a team building apps with AI: Muse Wonder looks strong when the work runs long — multi-step coding, business workflows, keeping context. Opus 5.5 still wins some of the coding suites that look more like from-scratch programs and terminal agents, and it leads Meta’s overall index. GPT-6 Astra and Claude Fable 5.1 sit in the mix on the same table.
Treat this as Meta’s snapshot on launch day. Public app-building leaderboards (see best models for vibe coding) measure something different, and they move.
After you build, you still need a host #
Whichever model writes the app — Muse Wonder, Claude, GPT, Cursor, Grok — a chat preview is not a place your team can work. You still need:
- A private server that stays up
- Company login , so only your people get in
- A database everyone shares
- Schedules for jobs that should run without someone clicking
- Monitoring , so a bad deploy is visible
- Backups and rollback when a prompt breaks something
That is hosting. It is also where most vibe-coded prototypes die.
Croft is the place the team runs it #
Croft does not sell models and does not pick a vendor. Croft is an MCP server, so it works with any MCP-capable assistant. Meta’s post does not describe a Muse-specific Croft integration; use the client you already have.
The loop is the same:
- Describe the app in your assistant (or in Ask Croft in the browser).
- Deploy to your croft.
- Invite the team. They sign in once.
You get a live URL on your own private server, behind one company login, with a database, schedules, and monitoring included. Connect setup: Claude, ChatGPT, Cursor, Grok, or any MCP client.
Related #
- [Best models for vibe coding](https://oncroft.net/best-models-for-vibe-coding/) — public benches, assistants, and how to choose
- [Vibe coding hosting](https://oncroft.net/vibe-coding/) — where the app lives after the chat window
- [AI app hosting](https://oncroft.net/ai-app-hosting/) — private hosting for apps you build with AI
- [Connect your AI assistant](https://oncroft.net/connect/) — MCP setup
- [Google Gemini 4 Argon for vibe coding](https://oncroft.net/blog/google-gemini-4-argon-vibe-coding/) — the last frontier-model drop, same hosting gap
FAQ #
What is Muse Wonder?
Meta’s new frontier series, launched 13 October 2026 (NZ): Muse Wonder and Muse Wonder Full. Meta says they target personal and professional agent work, with Active Memory and Active Focus. Availability in Meta’s post: Meta Model API and Muse Code today; Full for Muse Code subscribers; Muse product rollout coming, subscribers first.
Is Muse Wonder the best model for vibe coding?
On Meta’s table it is the one to watch for long-horizon coding and agent/business workflows. Opus 5.5 still leads Terminal-Bench, ProgramBench, and the Artificial Analysis Index. Use the model that matches the job; rankings will move.
Can I connect Muse Wonder to Croft?
Croft works with any MCP-capable assistant. Meta did not publish a Croft (or MCP) setup in the announcement. Point your MCP client at Croft, build, deploy.
Where should a team host an app built with Muse Wonder?
On a private host with company login and a database. That is Croft. Public preview URLs are for demos, not internal tools.
How much does Croft cost?
Solo from $24/month on annual billing, Team $99/month on annual. 7-day free trial, no card to start. Pricing →
Where to learn more #
- **Meta research post:**[Muse Wonder](https://research.meta.ai/blog/muse-wonder)
- **Croft:**[Get your croft](https://app.oncroft.net/signup) — 7-day free trial, Solo from $24/mo annual
Stake out your croft. #
Your team's first app could be live before lunch.
Start free trial 7 days free, no card to start. From $24/month - cancel anytime and take everything with you.