Meta releases Muse Wonder and claims parity with ChatGPT and Claude in coding

Meta has released Muse Wonder, a model internally named Watermelon, and claims it matches ChatGPT and Claude in programming and professional tasks. This is the company's own claim, and there are no independent measurements yet. Meta also hasn't said whether it will publish the weights. The answer determines which obligations the company faces under the European AI Act.

What is claimed and what has not been shown

According to TNW's description, Muse Wonder is the most capable model Meta has released to date. The key claim is that the model has caught up with the best closed systems from OpenAI and Anthropic exactly where people pay money for them: generating and editing code, working with documents, analytics, and other office scenarios.

For Meta, this is a serious claim. For years the company built its reputation on open models that lagged noticeably behind the leaders on complex tasks. A claim of parity changes the conversation, but for now it rests on words alone. The available materials name no benchmarks and give no figures, test conditions, or comparisons against specific versions of competing models. Without a table of results, the claim of having "caught up" cannot be verified.

• Model: Muse Wonder, internal name Watermelon.
• Claimed level: on par with ChatGPT and Claude in coding and professional work.
• Independent benchmarks: not provided in the primary materials.
• Weights status: Meta has not clarified whether it will publish them.
• Pricing and API access terms: not disclosed.

Why the weights question comes down to Brussels

The AI Act puts general-purpose models in a separate category. Requirements for their providers have applied since August 2025. The European Commission's powers to impose fines took effect in August 2026. A model released now gets no transition period: it is new and falls under the rules immediately.

Weight openness matters here. Providers of models under a free license with public weights are exempt from some obligations, including preparing technical documentation for regulators and downstream developers. But the exemption has a limit. It does not apply to models with systemic risk, and a presumption of such risk arises when training exceeds a threshold of 10^25 floating-point operations. Meta has not disclosed Muse Wonder's compute budget. For a model aiming at the frontier, exceeding this threshold looks likely, but that is an assumption, not a confirmed fact.

If the threshold is crossed, Meta will have to notify the European Commission, evaluate the model for dangerous capabilities, run adversarial testing, report incidents, and ensure cybersecurity, regardless of whether the weights are closed or open. If the model falls below the threshold and ships with open weights, the burden is noticeably lighter. That is why the silence on weights looks like a deliberate pause rather than an oversight.

What teams choosing a model should do

Don't carry Meta's claim into product decisions without checking it. Run Muse Wonder on your own tasks: your codebase, your documents, your tests. Public benchmarks often diverge from reality, and for coding the gap is usually the most noticeable.

If you work in the EU or serve European customers, request from Meta before integrating the documentation it is required to provide to downstream providers: a model description, its limitations, its copyright policy, and a summary of the training data. If the weights stay closed, integration goes only through Meta's API, and legal responsibility for compliance is allocated differently than with self-hosted deployment.

Watch for three signals: publication of a technical report with figures, a decision on the license and weights, and the first independent comparisons on tasks close to yours. Until those appear, Muse Wonder remains a candidate for testing, not a ready replacement for your current stack.