M**a Releases Muse Wonder and Wonder Full to Take On GPT-6 Astra and Claude Fable

Ma has released two models at once, Muse Wonder and Muse Wonder Full, and opened access to them on the day of the announcement. The company calls them direct competitors to the best systems from OpenAI and Anthropic. Discussion of the release invokes the names GPT-6 Astra and Claude Fable. There are two access channels: Muse Code subscribers got both models, while developers with access to the Ma Model API can so far connect only Wonder.

What we know about the release

The lineup is structured in a familiar way: a base Wonder and a senior Wonder Full. Judging by the names and the way access is split, Full is meant for the hardest tasks. Wonder looks like the workhorse that can be built into a product through the API. For now that is a logical but still speculative reading: M**a has not disclosed how the models differ in architecture, size or inference cost.

The confirmed facts fit in a short list:

• Number of new models: 2 (Muse Wonder and Muse Wonder Full)
• Muse Code subscribers: access to both models
• M**a Model API: Wonder only
• Positioning: competitors to the OpenAI and Anthropic flagships
• Published benchmarks, pricing and context window size: not provided in the announcement

What the announcement leaves out

The phrase "most powerful models" without numbers proves almost nothing. There are no results on SWE-bench or similar coding benchmarks. There is no data on reasoning, context length or cost per million tokens. There is no comparison with specific versions of competing models under identical conditions. For now, the comparison with GPT-6 Astra and Claude Fable remains a vendor claim, not a measured fact.

Still, the launch structure itself is telling. Muse Code subscribers, users of a coding service rather than a broad audience, were the first to get Wonder and Wonder Full. That signals a priority: M**a is competing for developers and for the everyday workflow in which a model writes, edits and reviews code. This is where OpenAI and Anthropic currently hold their strongest positions and their most loyal paying user base. Releasing Wonder in the API covers the second part of the job: third-party teams can build the model into their products and compare it with the solutions they already use.

The fact that Full is not yet available in the API is also a signal. The company is probably keeping the senior model inside its own product. That makes it easier to control the load on compute and to collect usage data. Or Full may simply be too expensive to offer openly at the prices M**a is ready to set right now. There is no confirmation either way.

How to test the model in practice

Teams that already pay for Muse Code can run Wonder and Wonder Full on their own codebase: refactoring, writing tests, hunting bugs in an unfamiliar module. Track the share of tasks solved on the first attempt and the time spent reviewing the output. That gives you the numbers the announcement lacks.

For those choosing a model for a product, it makes sense to start with Wonder through the API. Assemble your own set of 30–50 typical queries, run it in parallel on your current provider and on the new model, then compare quality, latency and cost. Make the decision to switch only after M**a publishes pricing and independent measurements appear. Until then, the rational stance is this: try it, measure it, but do not rewrite your architecture over a claim of parity.