Model guide
Meta Muse Wonder: pricing, benchmarks and Hermes Agent setup
Use Meta’s new API model through OpenRouter or a direct custom endpoint. Compare Wonder with Wonder Full, understand the subscription boundary, and budget for long agent tasks.
Published October 12, 2026 · Availability and prices checked on the publication date · 中文版
What Meta released
Meta announced Muse Wonder and Muse Wonder Full on October 12. Its release post describes better long-running work, Active Memory and Active Focus. For a Hermes Agent user, the first decision is access: ordinary Wonder has an API; Full is currently a Muse Code subscription model.
| Muse Wonder | Muse Wonder Full | |
|---|---|---|
| Meta API | Available: muse-wonder-1.0 | Not available at launch |
| OpenRouter | meta/muse-wonder | No Full API entry at launch |
| Muse Code | Available | Power Usage / Maximum Usage |
| Hermes Agent | API route described below | Cannot select Full through these API routes |
Meta’s model documentation lists a 1,048,576-token context window and text, image, video, audio and PDF inputs for Wonder. A provider accepting a modality does not mean every Hermes tool sends that format; check the route and tool you actually use. Setting Wonder’s reasoning effort to max does not turn it into Full.
Use Wonder on a managed Hermes instance
- Open the dashboard and find your Hermes instance.
- Click its current text model and choose the OpenRouter provider. Search for Muse Wonder.
- If the model is missing from the shortlist, search the full OpenRouter catalogue using
meta/muse-wonder. For manual entry, choose Custom platform…, enter platformmetaand model IDmuse-wonder. The resulting selection ismeta/muse-wonder, routed through OpenRouter on Hermes;metahere is its catalogue namespace, not a direct Meta API connection. - Confirm the switch, wait for the instance to be ready, and begin a fresh conversation. Verify the selected model on the instance card.
Paid model access depends on your account; Wonder is not a free-model or trial entitlement. OpenRouter requests use your saved OpenRouter key when enabled, otherwise your eligible managed credits. To pay OpenRouter directly, add your own key on the API keys page. A Meta API key belongs to the direct Meta endpoint, not the OpenRouter key field.
Self-hosted Hermes: OpenRouter configuration
Run hermes model, choose OpenRouter, supply your own OpenRouter key when prompted and select or enter meta/muse-wonder. The corresponding model section in ~/.hermes/config.yaml is:
model:
provider: openrouter
default: meta/muse-wonder
Keep OPENROUTER_API_KEY in Hermes’s local credential setup or ~/.hermes/.env, separate from this model block. Merge these fields into your existing configuration rather than replacing the whole file. See the Hermes configuration reference and upstream model selector.
Restart your CLI session after changing the configuration. For a running messaging gateway, use hermes gateway restart and start a new conversation; the gateway documentation explains its graceful restart behavior.
Optional: use Meta’s API directly
With a funded Meta API account, run hermes model and choose a custom OpenAI-compatible endpoint. Set the base URL to https://api.meta.ai/v1, provide your Meta API key, and use the bare model ID muse-wonder-1.0. These differ from the OpenRouter ID and key above. Meta documents this wire format in its Chat Completions reference; Hermes’s custom-endpoint implementation supports this setup. This guide verifies the configuration against documentation; it does not report a paid inference test.
API pricing and a worked budget
| Ordinary Wonder, per 1 million tokens | USD |
|---|---|
| Uncached input | $4.00 |
| Cached input read | $0.20 |
| Output | $20.00 |
These are Meta’s published API rates and OpenRouter’s listed token rates on October 12. Meta says long context has no higher token-price tier. Check Meta pricing and the OpenRouter model page before a large run.
For 100,000 uncached input tokens + 10,000 billed output tokens, the token estimate is $0.60: $0.40 input plus $0.20 output. If all that input qualifies for the listed cache-read rate, it becomes $0.22. Cache hits are conditional, not a guaranteed discount. Reasoning, repeated tool turns, search fees and hosting can increase the final bill. A million-token window is a capacity limit, not a flat-price allowance.
What the coding results show
Meta’s published benchmark table reports:
| Benchmark | Wonder | Wonder Full | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| SWE-Marathon | 53.1% | 57.5% | 42.5% | 46.3% |
| DeepSWE v1.1 | 78.8% | 79.0% | 74.1% | 74.2% |
| ProgramBench | 50.0 | 59.0 | 52.5 | 70.5 |
| Terminal-Bench 4.0 (AA) | 52.0% | 57.1% | 59.1% | 59.6% |
Both Wonder versions lead the two comparison models on SWE-Marathon and DeepSWE in this table. Full trails Opus on ProgramBench and both comparison models on Terminal-Bench. This supports trying Wonder for sustained repository work; it does not establish a win on every coding task.
The model card describes the evaluation setups, including different agent harnesses and reasoning settings. Treat these as reported system results, not a controlled Hermes speed comparison. Separately, Artificial Analysis lists Full at 56 on its Intelligence Index. A live ranking can change; that number is not a promise about your project.
For your own comparison, use the same repository snapshot, requirements, tool permissions and time budget. Judge the final diff, passing tests and actual cost rather than completion time alone.
Active Memory and Active Focus in Hermes
Meta’s release describes remembering relevant information across conversations and staying on track through obstacles and changing requirements. Those capabilities do not give a third-party Hermes session access to your Muse app memories. Meta’s Chat Completions API is stateless: the caller supplies conversation history.
In Hermes, configure memory, tools and saved task state in Hermes itself. The model can reason over what Hermes supplies. A larger context window does not replace durable storage, and a model switch does not import another product’s private conversations.
Wonder Full subscriptions and privacy
Full currently requires Muse Code Power Usage ($50/month) or Maximum Usage ($200/month). Meta advertises 20× and 80× the base usage respectively; these are usage tiers, not unlimited inference. The subscription documentation says subscription benefits attach to the Muse Code CLI key. Separate API keys remain usage-billed, so buying the subscription does not fund Hermes API calls.
Meta’s model documentation distinguishes Standard API data treatment from Contributor subscriptions. Subscriptions default to Contributor, which permits use of contributed data for training; ordinary Wonder’s API is Standard-only at launch. Review the tier and regional conditions shown during onboarding before using work repositories.
Frequently asked questions
Can I use Muse Wonder with Hermes Agent?
Yes. Use OpenRouter with meta/muse-wonder, or a custom Meta endpoint with muse-wonder-1.0 and a Meta API key. Full is not available through these API routes at launch.
Does Muse Code’s $50 or $200 subscription cover Hermes usage?
No. Subscription benefits apply to the Muse Code CLI key. Hermes API requests are billed through the API account or eligible managed credits used by the selected route.
Does Active Memory automatically remember my Hermes conversations?
Do not assume it does. Meta’s Chat Completions API is stateless. Hermes must supply history and manage its own durable memory; Muse app memories are not automatically imported.