> ## Documentation Index > Fetch the complete documentation index at: https://docs.tess.im/llms.txt > Use this file to discover all available pages before exploring further. # Muse Wonder > Learn about Muse Wonder, Meta's model for agents that operate the screen, change a repository, and recover when a step breaks. Muse Wonder is Meta's frontier model for **agents that operate the screen, change the repository, and pick the task back up when a step breaks**. Meta launched the series on October 12, 2026, together with Muse Wonder Full. Wonder is on the API, with the identifier `muse-wonder-1.0`. Full is the most advanced model in the series and, for now, is available only through a Muse Code subscription. This page is about Wonder. Meta names two parts of that behavior. Active Memory carries what matters across conversations and brings back the right detail when the task needs it. Active Focus stays on the goal when the task stalls: it tries another path, recovers from an error, asks for clarification, and checks in before an important action. | **Provider model** `muse-wonder-1.0` | **Context** 1M tokens | **Output** Text | **Provider** Meta | | :- | :- | :- | :- | | **Input** Text, image, speech, and video | **Thinking** Reasoning | **Series** Wonder, public API | **Released** October 12, 2026 | Artificial Analysis records that multimodal input on the max-effort variant. Meta has not published a parameter count. ## When to use Muse Wonder and when to use Opus 5.5 The difference is the kind of action. Meta positions Wonder for repository coding, computer use, and tasks where the agent has to remember the context and recover from an error. Opus 5.5 stays ahead when the work needs open-ended judgment, a terminal, or a program written from scratch. | | Muse Wonder | Claude Opus 5.5 | | - | - | - | | Best for | The screen, the repository, and recovery after a failure | Open-ended judgment, the terminal, and from-scratch programming | | Input | Text, image, speech, and video | Text and image | | Context | 1M | 1M | | Where Meta's measurement favors it | DeepSWE, SWE-Marathon, and memory | Terminal-Bench, ProgramBench, and the Artificial Analysis index | > [!TIP] > > Use Wonder when the agent has to see the screen, edit the code, and continue after a failure. For work where judgment shifts at each step, compare it with Opus 5.5 in your own workflow. ## What Meta measured The numbers below come from Meta's table at max effort. Muse Wonder Full scores higher on almost every row, and it is not the model on this page. ### Coding that crosses the repository On DeepSWE 1.1, long-horizon coding, Wonder scored 78.8%. Claude Opus 5.5 scored 74.2% and GPT-6 Astra scored 74.1%. On SWE-Marathon, ultra-long-horizon engineering, the scores were 53.1%, against 46.3% for Opus 5.5 and 42.5% for Astra. On the terminal and on from-scratch programming, the order flips. On Terminal-Bench 4.0, Wonder scored 52.0%, behind Opus 5.5 (59.6%) and Astra (59.1%). On ProgramBench it scored 50.0%, well behind Opus 5.5 (70.5%). ### Memory, tools, and the screen On MemoryBench, Wonder scored 85.6, ahead of Opus 5.5 (81.5) and Astra (81.4). On JobBench, professional tool use, it scored 70.5%, behind Opus 5.5 (75.4%) and ahead of Astra (66.1%). On OSWorld 2.1, computer use on the partial measure, Wonder scored 72.6%. Opus 5.5 scored 81.8% and Astra scored 75.8%. On the binary measure of the same test, Wonder and Astra tied at 34.3%. On domain work, Meta's table is mixed. On Finance Agent v2, 56.6%, close to Opus 5.5 (58.6%) and above Astra (53.5%). On Harvey LAB, legal work from documents, 10.0%, above Opus 5.5 (4.2%) and Astra (8.6%): the whole band is still low. On HealthBench Pro, 60.5%, behind Opus 5.5 (65.6%) and Astra (65.3%). ## What the independent benchmarks show | Benchmark | Muse Wonder | Opus 5.5 | GPT-6 Astra | What it measures | | - | -: | -: | -: | - | | DeepSWE 1.1 | **78.8%** | 74.2% | 74.1% | Long-horizon coding | | SWE-Marathon | **53.1%** | 46.3% | 42.5% | Ultra-long-horizon engineering | | MemoryBench | **85.6** | 81.5 | 81.4 | Context memory | | JobBench | 70.5% | **75.4%** | 66.1% | Professional tool use | | Terminal-Bench 4.0 | 52.0% | **59.6%** | 59.1% | Terminal coding | | ProgramBench | 50.0% | **70.5%** | 52.5% | From-scratch programming | | OSWorld 2.1 (partial) | 72.6% | **81.8%** | 75.8% | Computer use | | Artificial Analysis Index | 52 | **58** | 53 | Composite index, on Meta's table | In Artificial Analysis's own measurement, the max-effort variant scored **52 on the Intelligence Index**, ranking 9th among 216 models in the same class. The median for that price range is 27. In that run the model generated text at 104.1 tokens per second, above a median of 83.6, and took 19.73 seconds to the first token, against a median of 5.87 seconds. That time includes reasoning. Average cost was 4.52 dollars per index task, at a list price of 4 dollars for input and 20 dollars for output per million tokens. On BenchLM, the partial aggregate is **70.44 out of 100**, ranking 16th among 891 models. Coverage is still partial (21 benchmarks), so the score is conservative. > [!NOTE] > > The table was published by Meta at max effort. The Intelligence Index, speed, latency, and cost per task are from Artificial Analysis on that same variant. The BenchLM aggregate uses partial coverage. Results depend on the harness, the tools, and the effort. Validate in your own workflow before changing a production setup. ### Where it doesn't lead * On from-scratch programming, **Opus 5.5 opens a wide gap** (70.5% versus 50.0% on ProgramBench). * On terminal coding and computer use, Opus 5.5 and GPT-6 Astra stay ahead. * Time to first token rises while the model is reasoning. For a short classification, another model answers sooner. * In Artificial Analysis, the max variant writes more than the median for the range: 130 million output tokens on the index, against 85 million. Ask for short answers when the format matters. * Muse Wonder Full scores higher on Meta's table. It is not the API model. ## Where it adds the most value ### A change that crosses several files The fix does not fit in one file. Wonder walks the repository, edits in parallel, and checks the result before it stops. **Expected result:** the change you asked for, verified, with no extra file outside the scope. ### A task that stalls midway The plan breaks on a tool, a source, or a step that will not close. The model tries another path, says where it stopped, and continues an important action only after you confirm. **Expected result:** the task finishes, or it stops with the blocker written down, with no silent action. ### A process that lives on the screen The figure is in a system, a PDF, and a spreadsheet, and someone has to work across all three. Wonder sees the screen, joins the sources, and returns what it checked. **Expected result:** every number points to the screen or the file it came from. ## How to apply it in Tess **1. Pick a task with a screen, code, or several tools** Prefer a repository change, a process with several tools, and a check against the screen. For short classification at volume, another model answers sooner. **2. Lock what counts as done** State the outcome, the files that may change, and the test that proves it is finished. State what is out of scope. **3. Set tools and approvals** Enable only the tools you need. Require confirmation before sending, publishing, paying, or changing data. **4. Ask for an explicit recovery** Tell the agent to try another path when a step fails, and to stop, with the reason, if the second attempt fails too. **5. Sample before you let it run** Run one full case, check what changed, and only then repeat the flow. ## Three playbooks to get started ### 1. A repository change with a check **Input:** the repository, the expected behavior, and the tests that already exist.\ **Tools:** reading and editing files, running the project's tests. ```text theme={null} Implement only the behavior described. Edit the files you need and run the tests named here. If a test fails, fix it and run it again. If it fails a second time, stop and say what broke. Do not refactor anything you were not asked to touch. ``` **Done when:** the requested tests pass and the diff stays inside the scope. ### 2. A process that can fail midway **Input:** the goal, the sources, and the action that needs confirmation.\ **Tools:** the process integrations, read-only until confirmation. ```text theme={null} Carry out the goal in steps. If a tool or a source fails, try one alternative path. If that fails too, stop and describe the blocker. Before you send, publish, or change data, show what you are about to do and wait for confirmation. ``` **Done when:** the result is done, or the blocker is written down, and no important action went out without confirmation. ### 3. A check between the screen and a file **Input:** the PDF or spreadsheet and the system where the same figure appears.\ **Tools:** file reading and the system screen. ```text theme={null} Check each number in the file against the screen. Return the value, the unit, and the place it came from, in the file and on the screen. If the two disagree, list both. Do not pick one, and do not change the system. ``` **Done when:** every number has both sources, or the disagreement is explicit. ## Limitations and best practices * Muse Wonder Full is not this model. It leads Meta's table, and access today is through a Muse Code subscription. * A 1M-token context does not replace selecting information. Send only relevant content. * List price on Meta's API is 4 dollars for input and 20 dollars for output per million tokens. A cache read costs 0.20 dollars. * Meta describes strong safeguards against chemical, biological, radiological, nuclear, and cyber misuse, and says defensive security work remains possible. A full safety report was still to come at launch. * Review a sample before publishing, sending, or automating decisions. * Keep human approval for financial, legal, clinical, or high-impact decisions. See also: [Models and Costs](/en/models-and-cost) · [Claude Opus 5.5](/en/claude-opus-5-5) · [Muse Wonder — Meta](https://dev.meta.ai/models/muse-wonder) · [Muse Wonder — Meta research](https://research.meta.ai/blog/muse-wonder). This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.