Trying Meta's most capable new model on its $50 monthly plan means checking a setting before you hand it your code. Muse Code subscriptions default to Contributor mode, which lets Meta use submitted content to train future models. You can turn it off and keep access to Muse Wonder and Muse Wonder Full on the plans that include them, according to Meta's subscription documentation.
Meta released Muse Wonder and Muse Wonder Full on October 12, describing them as its strongest models so far, trained for prolonged work involving tools and personal context. Wonder is available through the Meta Model API and Muse Code. Full launches in Muse Code subscriptions. Meta's research announcement also says Wonder will reach its Muse assistant, with subscribers first.
Full's entry price is $50 a month on the Power Usage plan, with a $200 Maximum Usage plan also available, according to Meta's model page. A team evaluating the API version gets a different model, with different benchmark results, from the one Meta calls its most advanced.
Turning off contribution keeps the models available
Contributor mode covers the prompts, code, files and context you send, along with the model's responses. Subscribers receive more usage in exchange for allowing training on that material. Switching the setting off means future content isn't used to train Meta models unless you turn it back on.
The same documentation explicitly says Wonder and Wonder Full are available in either mode on eligible plans. That gives you a concrete decision before a trial: leave contribution enabled for the higher allowance, or disable it before submitting work you don't want used for training. Meta describes the change in terms of future content. The documentation doesn't describe switching the setting off as deleting earlier submissions.
The subscription also limits where its included usage applies. The subscription applies to the credential connected during Muse Code's CLI onboarding, and that credential is for Muse Code use only. Additional API keys created under the associated Meta Model API account are billed separately on a pay-as-you-go basis. Meta's instructions separate the subscription from API spending by an application you build.
Those monthly plans have usage limits. Meta advertises Power Usage at 20 times the Everyday Usage allowance and Maximum Usage at 80 times it, while its subscription guide says subscribers who exhaust a limit must upgrade or wait for it to refresh. Maximum Usage customers must wait. Benefits and availability can vary by region, with the applicable terms shown during onboarding.
The API model has its own price and scorecard
Muse Wonder's API identifier is muse-wonder-1.0. Meta lists prices of $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens. Both Wonder and Full have a one-million-token context window, but the pricing table gives Full no per-token API price and labels it subscription-only in Muse Code.
At those published rates, a call containing 100,000 uncached input tokens and producing 10,000 output tokens would cost $0.60 in model-token charges. That's a calculation from Meta's rates, rather than a measured cost for completing a coding task. An agent that makes repeated calls while testing and revising work accumulates charges across those calls.
Full's advantage depends on the test. On DeepSWE 1.1, Meta's table reports 79.0% for Full and 78.8% for Wonder, a gap of 0.2 percentage points. On SWE-Marathon, the scores are 57.5% and 53.1%. ProgramBench shows a wider difference: 59.0% against 50.0%. Paying for Full buys access to the stronger variant, but those figures don't support one uniform improvement across coding work.
The tasks differ considerably. DeepSWE has 113 tasks across 91 repositories and five languages, graded with functional and regression tests. SWE-Marathon evaluates substantial software projects. ProgramBench asks agents to reconstruct complete programs in a clean-room environment, then checks them against hidden behavioral tests. The model card defines ProgramBench success as passing at least 95% of the tests.
A coding agent is part of the comparison
Meta's results mix controlled runs with results from outside evaluators and rival providers. Its methodology says controlled internal coding comparisons use the same benchmark-specific agent harness, meaning the software that supplies tools and manages the model's work. External results retain the harness or agent product used by the reporting organization.
The card also explains how Meta selects scores: it uses a comparable third-party or provider-reported result where available. Otherwise, it takes the highest of its internally validated run, a self-reported score or an official leaderboard result. It cautions that competitor runs performed internally are best-effort because prompts and tools may not be tuned optimally. Those selection rules matter when a small difference between two numbers is part of a buying decision.
SWE-Marathon needs particular care. Its GPT-6 Astra and Claude Fable 5.1 entries use Codex and Claude Code respectively, so Meta explicitly says that row isn't a controlled comparison using one common harness. Full's 57.5% result also uses xhigh reasoning effort, an exception to the max-effort settings described for the main comparisons. Those details appear in the card's benchmark notes.
The broader scorecard has limits, too. Meta lists Artificial Analysis Intelligence Index scores of 56 for Full and 52 for Wonder, against 58 for Claude Opus 5.5. Full leads some rows and trails others. For an application using Wonder through the API, Wonder's column is the relevant starting point. Full's subscription results describe a separate option.
Memory testing covers long conversations
Meta puts Active Memory and Active Focus at the center of its launch. Memory means retaining important context across conversations and retrieving relevant details. Focus describes continued work through obstacles, including trying another approach and asking for clarification. The company connects both capabilities to agents handling work that extends beyond a single response.
Its published memory test is more specific than that product description. The model card defines MemoryBench as an internal evaluation using complex conversations of 64,000 to 100,000 tokens. They include distracting conversations and turns that contradict or update earlier facts. An LLM judge grades answers against weighted criteria for each question.
Wonder scores 85.6 on that test and Full scores 85.5, according to Meta's table. Full produces no improvement in this particular memory result. The exercise tests retrieval and handling changed information in lengthy conversations. It doesn't establish how a deployed assistant manages a user's changing circumstances over months.
The card also warns that models can make mistakes during multi-step reasoning, particularly in unfamiliar scenarios, and that performance varies with tools and serving configuration. For anyone building around Active Memory, those stated limits make corrected facts a useful trial case: after you replace an old requirement, does the agent use the new one when it resumes work?
Meta recommends confirmation before irreversible actions
Meta says it trained Wonder to handle private information carefully and resist adversarial instructions. Its model card recommends additional protections suited to the application, including human confirmation before irreversible actions. Those responsibilities remain relevant whether a developer chooses API calls or a coding subscription.
The card contains risk assessments and safety results, while the launch announcement promises a further report with additional safety evaluations. Reading the existing card gives developers more evidence than the launch demonstrations alone. Its caveats also help define a trial: check the agent's completed work and decisions, alongside the benchmark score that prompted you to try it.
The next disclosures to watch are that additional safety report and any change to Full's distribution. Meta's current model page puts Full inside Muse Code and Wonder on the API. Until that changes, evaluate the version you can deploy, and check Contributor mode before sending work from a repository into a subscription session.