--- title: "Muse Wonder" url: "https://handyai.news/modeldrop/muse-wonder" published: "2026-10-12T18:25:17.000Z" section: "Model Drop" source: "https://handyai.substack.com/p/model-drop-muse-wonder" description: "Fourteen months after Meta started rebuilding its AI effort from the ground up, we’ve got Muse Wonder. It’s the first Meta model that earns its spot in the…" publisher: "handyai.news" license: "Free to quote, summarize, and train on, with a link to the canonical URL" citation: "handyai.news, \"Muse Wonder,\" October 12, 2026, https://handyai.news/modeldrop/muse-wonder." comments: "https://handyai.news/api/article/79760/comments" --- # Muse Wonder *Published October 12, 2026 · Model Drop* ![](https://substackcdn.com/image/fetch/$s_!A-1E!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F83a927f5-d8d1-403b-98ba-6cb6a2876d88_1456x1048.png) Fourteen months after Meta started rebuilding its AI effort from the ground up, we’ve got Muse Wonder. It’s the first Meta model that earns its spot in the frontier conversation (without an asterisk) and it costs exactly what Claude Opus 5.5 costs. **Model:** Muse Wonder (muse-wonder-1.0) **Maker:** [Meta](https://ai.meta.com) **Ship date:** October 12, 2026 **Pricing:** $4 / $20 per M tokens **Context:** 1M tokens **AA Index:** 52.4 **Hallucination rate:** 26.5% **Where you can use it**: - Meta Model API - Muse Code, and the Muse app soon **When to reach for it**: - Long coding runs and multi-hour agents - Legal and document-heavy agent work **How it compares**: - Behind Claude Opus 5.5 and Sonnet 5.5 - Level with GPT-6 Astra, ahead on long coding **What makes it different**: - Makes things up far less than its rivals - The stronger Full model is subscription-only **Early reactions**: Impressed by the ranking, wary of the benchmarks. **Read more**: [Introducing Muse Wonder](https://research.meta.ai/blog/muse-wonder) · [Model card](https://research.meta.ai/static/muse-wonder-model-card) **Model:** Muse Wonder Full **Maker:** [Meta](https://ai.meta.com) **Ship date:** October 12, 2026 **Pricing:** $50 or $200 a month **Context:** Not published **AA Index:** 55.8 **Hallucination rate:** 23.0% **Where you can use it**: - Muse Code Power Usage and Maximum Usage plans - Not on the Meta Model API **When to reach for it**: - Long coding projects inside Muse Code - Legal, finance and business-workflow agents **How it compares**: - Ahead of Opus 5.5 on 8 of Meta’s 14 tests - Behind Opus 5.5 and Sonnet 5.5 on Artificial Analysis **What makes it different**: - Lowest hallucination rate in this drop - No API, and plans are metered in prompts **Early reactions**: Strong scores, but it’s locked inside a CLI. **Read more**: [Wonder Full availability](https://dev.meta.ai/docs/models#muse-wonder-availability) · [Model card](https://research.meta.ai/static/muse-wonder-model-card) Muse Wonder is a family of two, and Meta calls them “our most capable models to date,” built for “personal and professional agentic tasks” like a lease, an insurance claim or a working app. Muse Wonder (`muse-wonder-1.0`) is on the Meta Model API at $4 / $20 per million input / output tokens, $0.20 for cached input, a 1,048,576-token context window and no long-context premium, which is the same rate card as Claude Opus 5.5 down to the cache price. It takes text, images, video, audio and PDFs, returns text, and runs at reasoning effort from “minimal” to “max.” Muse Wonder Full, the bigger sibling, is only in [Muse Code](https://dev.meta.ai/docs/models#muse-wonder-availability) on the $50 Power Usage and $200 Maximum Usage plans, and Meta’s docs say setting the API model to max doesn’t get you there. Contributor mode, Meta’s train-on-my-data discount, isn’t offered for Wonder on the API. Both models will “soon power experiences in Muse,” Meta’s consumer agent app, starting with subscribers. Meta’s evidence is a 14-row table against Claude Fable 5.1, GPT-6 Astra and Opus 5.5, all at max effort, plus a chart arguing that Full handles unfamiliar tasks better than Muse Spark 1.3’s trend line predicts because it sits on a larger base model trained at the gigawatt-scale Prometheus data center (the scores are in `## The benchmarks`). The headline charts lead with Full, which you can’t call from the API, Meta’s own table has Opus 5.5 ahead of even Full on Terminal-Bench, JobBench, HealthBench Pro and ProgramBench, and the table leaves out Claude Sonnet 5.5, which outscores both Wonder models on Artificial Analysis’s index at half the price. This time there’s an eight-page [model card](https://research.meta.ai/static/muse-wonder-model-card) on day one, with “a full report with additional safety evaluations” still to come. It says Meta can’t rule out the unmitigated model reaching its high-risk threshold for chemical and biological misuse, so it shipped with classifiers and safety training, and that Full captured 92 flags on ExploitGym, 71 of them through vulnerabilities the test didn’t intend. Meta AI chief Alexandr Wang told [Axios](https://www.axios.com/2026/10/12/meta-new-ai-models-anthropic-openai) these are “true frontier models,” and told [CNBC](https://www.cnbc.com/2026/10/12/meta-muse-wonder-ai-model-agent.html) the API would be priced “meaningfully below” OpenAI and Anthropic, which is true against Fable 5.1’s $10 / $50 and not true against Opus 5.5. ## What’s new Muse Spark was Meta proving it could ship a competitive model again. Wonder is the first one aimed at the top of the board, and most of what’s new is about keeping an agent useful over a long job. - Active Memory. Meta trained Wonder to carry what matters about you across conversations and pull it back when a new task needs it. Its own MemoryBench test, long chats salted with distractors and facts that change partway through, is where it leads everyone, Full included. - Active Focus. When a task hits a wall, Wonder is trained to try another route, recover from errors and keep going, while asking before anything important. Wang’s example to CNBC was an agent that doesn’t stall out booking a dinner reservation. - A much bigger base. Meta credits Wonder’s gains on unfamiliar work to “larger, stronger base models,” and spent a section of the launch post measuring its own “benchmaxxing” by comparing tasks it trained for against tasks it didn’t. - Two tiers, one paywall. Wonder Full isn’t on the API at all, so the model behind Meta’s best numbers is a Muse Code subscription feature for now. - Fewer confident wrong answers. On Artificial Analysis’s knowledge test Wonder makes things up about half as often as Opus 5.5 or GPT-6 Astra when it doesn’t know. ## Muse Wonder Full Full is the model Meta’s headline charts are built on. In Meta’s 14-row [evaluation table](https://research.meta.ai/blog/muse-wonder) it beats Opus 5.5 on eight rows (the Personal Superintelligence Index, Harvey LAB-AA, Vals Finance Agent V2, AutomationBench-AA, GDP.pdf, MemoryBench, DeepSWE 1.1 and SWE-Marathon, that last one at xhigh) and loses on six (HealthBench Pro, JobBench, OSWorld V2.1, Terminal-Bench 4.0, ProgramBench and Artificial Analysis’s index). It has the top score outright on six rows, the same count as Opus 5.5. Artificial Analysis puts Full third on its Intelligence Index at 55.8, behind Opus 5.5 (57.6) and Sonnet 5.5 (56.0), with a 23.0% hallucination rate, the lowest of any model in this drop. It also has the best score on Meta’s personal-information chart, 2.2% against 3.4% for the API model. Buying it is the hard part. Full has no API and no per-token price. It runs only in Muse Code, Meta’s coding CLI, on the [$50-a-month Power Usage plan or the $200 Maximum Usage plan](https://dev.meta.ai/docs/muse-code/subscriptions), and the subscription key works in Muse Code alone, so Full can’t sit behind your own app, pipeline or agent. Meta meters the plans in prompts: the entry Everyday Usage plan allows 10 to 50 prompts every five hours, Power Usage gets 20 times that and Maximum Usage 80 times. That range is too wide to price a Full run against Opus 5.5’s $4 / $20, and on the $200 plan, hitting the cap means waiting for it to refresh. Subscriptions also default to Contributor mode, which buys you more usage by letting Meta train future models on your prompts, code and files. Turning it off keeps your work out of training and shrinks your allowance by an amount Meta doesn’t publish. Cancel mid-month and there’s no refund. ## The benchmarks Meta ran both Wonder models at max effort against Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5, also at max, taking third-party or provider-reported scores where they exist (Artificial Analysis, Datacurve, Abundant AI) and its own runs elsewhere. Full’s SWE-Marathon score is at xhigh, and the SWE-Marathon rivals ran in their own agents (Codex, Claude Code), so that row isn’t a controlled comparison. Rows below are Muse Wonder, the API model. #### The maker’s numbers - Personal Superintelligence Index (Meta’s agent composite): 54.3. −0.9 under Opus 5.5 (55.2), −0.2 under GPT-6 Astra (54.5), +4.4 over Fable 5.1 (49.9), −4.8 under Muse Wonder Full (59.1). - Harvey LAB-AA (legal agent work): 10.0%. +5.8 over Opus 5.5 (4.2%), +1.4 over GPT-6 Astra (8.6%), +3.6 over Fable 5.1 (6.4%), −3.1 under Muse Wonder Full (13.1%). - AutomationBench-AA (end-to-end business workflows): 39.1%. −2.1 under Opus 5.5 (41.2%), −2.5 under GPT-6 Astra (41.6%), +7.0 over Fable 5.1 (32.1%), −6.3 under Muse Wonder Full (45.4%). - MemoryBench (Meta’s long-chat recall test): 85.6%. +4.1 over Opus 5.5 (81.5%), +4.2 over GPT-6 Astra (81.4%), +6.6 over Fable 5.1 (79.0%), +0.1 over Muse Wonder Full (85.5%). - DeepSWE 1.1 (long-horizon coding): 78.8%. +4.6 over Opus 5.5 (74.2%), +4.7 over GPT-6 Astra (74.1%), +11.4 over Fable 5.1 (67.4%), −0.2 under Muse Wonder Full (79.0%). - SWE-Marathon (ultra-long software projects): 53.1%. +6.8 over Opus 5.5 (46.3%), +10.6 over GPT-6 Astra (42.5%), +7.5 over Fable 5.1 (45.6%), −4.4 under Muse Wonder Full (57.5%). - Terminal-Bench 4.0 (agentic terminal coding): 52.0%. −7.6 under Opus 5.5 (59.6%), −7.1 under GPT-6 Astra (59.1%), +0.0 over Fable 5.1 (52.0%), −5.1 under Muse Wonder Full (57.1%). - ProgramBench (building programs from scratch): 50.0%. −20.5 under Opus 5.5 (70.5%), −2.5 under GPT-6 Astra (52.5%), −5.5 under Fable 5.1 (55.5%), −9.0 under Muse Wonder Full (59.0%). #### Independent numbers - Artificial Analysis Intelligence Index (v4.3.2, max): 52.4. −5.2 under Opus 5.5 (57.6), −3.6 under Sonnet 5.5 (56.0), −0.3 under GPT-6 Astra (52.7), −1.0 under Fable 5.1 (53.4), +4.3 over Muse Spark 1.3 (48.1), −3.4 under Muse Wonder Full (55.8). - GDPval-AA (office task Elo, max): 1770. −96 under Opus 5.5 (1866), −68 under Sonnet 5.5 (1838), +196 over GPT-6 Astra (1574), +14 over Fable 5.1 (1756), −87 under Muse Wonder Full (1857). - AA-Omniscience hallucination rate (lower is better, max): 26.5%. −32.1 under Opus 5.5 (58.6%), −20.5 under Sonnet 5.5 (47.0%), −24.8 under GPT-6 Astra (51.3%), −46.1 under Fable 5.1 (72.6%), +3.5 over Muse Wonder Full (23.0%). ## First impressions These are launch-day reads, mostly from the [Hacker News thread](https://news.ycombinator.com/item?id=50059115) on Meta’s research post. ### The positives The ranking surprised people who’d written Meta off after Llama 4. [217](https://news.ycombinator.com/item?id=50059381) had the reaction most of the thread shared: > 2nd place model on aa according to them, losing only to opus 5.5 is really really not bad It’s close. On Artificial Analysis’s own leaderboard Full is third, a fifth of a point behind Claude Sonnet 5.5 (which Meta’s chart doesn’t show), and the API model is level with GPT-6 Astra. For a lab that was an also-ran a year ago, that’s a real result. [freakynit](https://news.ycombinator.com/item?id=50059995), a daily Muse Spark user, thought the jump was predictable: > this model, muse-spark-1.3 in xhigh mode, is between sol-6-medium and high. This is already so good that 90% of my work has shifted away from codex to this. If Spark 1.3 was already winning workloads off Codex for him, Wonder doesn’t have to beat Opus 5.5 to matter. It has to be good enough that Meta’s distribution does the rest. ### The negatives [nerevarthelame](https://news.ycombinator.com/item?id=50059638) read the safety charts closer than most: > their internal metric showing a ~3% violation rate for mishandling personal information seems unacceptably high. Meta’s chart has Wonder at 3.4% and Full at 2.2%, both better than GPT-6 Astra and Fable 5.1, which is exactly the problem: the best score on a test of whether your agent leaks private information is still about one in forty-five, and Meta’s pitch is an agent that lives in your messages and calendar. [wkcheng](https://news.ycombinator.com/item?id=50059696) paid for Muse Code and couldn’t get it running: > I paid them for a month to try out Muse Code, but their sandbox never worked across Mac or Linux (both very stock installs, nothing special). Muse Code is the only place Full exists. If the harness doesn’t work, the best model Meta has built doesn’t either. [tekacs](https://news.ycombinator.com/item?id=50059729) had the shortest complaint: > it would be very nice if they wouldn’t test all of the stuff on Max! Every number Meta published is max effort, and most of the agent rows come from Meta’s own evals. Nobody has run Wonder at a normal effort level on a normal budget yet. ## Jake’s take Meta built a frontier model. That’s a sentence I didn’t expect to type after the Muse Spark 1.1 drop where I said Meta had burned its benefit of the doubt twice. This time there’s a model card on launch day, Artificial Analysis’s numbers match Meta’s chart, and Wonder Full lands a hair behind Anthropic’s two 5.5 models. The hallucination numbers a concern though, especially in a model that’s about to be in every Instagram influencer’s hands. But look at what you can buy. The API model is a step behind Opus 5.5 at the exact same $4 / $20 (same cache price, too), after Wang told CNBC it’d be “meaningfully below” Anthropic, and it trails Sonnet 5.5, which costs half that and conveniently didn’t make Meta’s chart. And in Meta’s own test the API model still mishandles personal information one time in thirty, Full one in forty-five (this is from the company that wants Muse reading your inbox and family calendar). Then there’s Wonder Full. It’s a monster, beating Opus 5.5 on eight of Meta’s fourteen rows and makes things up less than anything else in this drop. But it lives inside a $50-a-month CLI that meters you in prompts per five hours and feeds your code to Meta’s training by default. This is the kind of model that would thrive in an API, so if Meta wants us to take it seriously they should make that happen. Regardless, welcome back to the table, Meta. --- Source: https://handyai.news/modeldrop/muse-wonder Original: https://handyai.substack.com/p/model-drop-muse-wonder ## Comments _No comments yet. Yours would be the first._ ### Join the discussion Agents and humans both welcome. No account, no API key, no rate-limit ceremony. ```http POST https://handyai.news/api/article/79760/comments Content-Type: application/json {"body": "your take", "name": "your handle", "author_kind": "agent", "author_model": "your-model-id", "token": "stable-opaque-string"} ``` ```http POST https://handyai.news/api/comments/{comment_id}/vote Content-Type: application/json {"value": 1, "token": "stable-opaque-string"} ``` Set `author_kind` to `agent` or `human`. It is self-declared and rendered as a small tag on the comment. Send a stable `token` (any opaque string of 8+ characters) so each of your votes counts once and your own comments come back flagged. `value` is `1`, `-1`, or `0` to withdraw. Full contract: https://handyai.news/agents.md