Meta spent close to a year and a considerable amount of money rebuilding its model stack, and the first public result carries the internal codename Watermelon. Muse Wonder is the standard tier, Muse Wonder Full the high-compute tier. Zuckerberg has told interviewers, with visible confidence, that Meta is back at the front of the pack.
The leaderboard argument is one conversation. Getting a model with that price tag to run inside a normal pipeline is a different one. What follows covers what Meta actually shipped, what the benchmarks and the bills say, and which infrastructure layer decides whether the model is usable at all.
A gateway is not a switcher. A switcher answers which provider is being used. A gateway answers how traffic moves between providers, protocols and projects.
What Meta shipped and how the tiers divide
Base models, assistant, and code entry point
Muse Wonder and Muse Wonder Full are the base models, the compute layer underneath everything else. Wonder balances speed against cost for everyday traffic. Wonder Full is the high-compute version, built to work through harder problems.
Muse is the consumer assistant, the interface ordinary users open in a browser or on the desktop, designed to work out of the box with no parameter tuning.
Muse Code is the engineering entry point. It connects to a development environment and directs the base model to rewrite and refactor code.
Two behaviours moved into pre-training
Older language models handled olympiad problems well and fell apart on a real support call. A dependable everyday assistant needs common sense about daily chores and the follow-through to act on it. Meta pushed interaction style and initiative into pre-training instead of attaching them afterwards.
Active Memory keeps key facts across turns. A business rule or a personal habit explained last week can still be retrieved days or weeks later, inside a task unrelated to the original conversation. The ritual of pasting a wall of background into a new window goes away.
Active Focus handles failure paths. Page structures change, dependency downloads fail, APIs time out. Take a booking system whose primary query path is blocked. The model adjusts the plan and tries an alternative channel instead of stopping to ask a human, and it only requests confirmation at points that matter, such as anything that charges money or deletes data.
The benchmark sheet and the two numbers underneath it
Meta's published results
On the Harvey Legal Agent Benchmark, handed a thick pile of raw case material and strict instructions, the model stays organised and produces solid drafting. On tax calculations and medical data analysis, where errors are expensive, the reasoning holds together and the habit of inventing facts is largely gone. For generalisation, the team put the previous Muse Spark 1.3 and Wonder Full on problems neither had been trained for. With more inference compute, Wonder Full clearly outperformed what a normal model manages on unfamiliar work.
Artificial Analysis across dimensions
The third-party evaluation from Artificial Analysis places Muse Wonder in the same conversation as the leaders. Six of ten into the global top three, across a set covering agentic and professional work. Fourth overall, with an intelligence index of 56. On coding, at maximum reasoning effort, Muse Code with Wonder Full matches Claude Code with Sonnet 5.5 and edges slightly past the Opus 5.5 configuration.
Cost per task and input throughput
Two details in the report deserve a slower read. At maximum reasoning effort, the Wonder family averages $4.52 per task. That is cheaper than the expensive Opus 5.5, clearly above GPT-6 Astra at $3.26, and far beyond Gemini. Input token throughput also stays enormous. Output tokens have been reined in, but sustaining Active Memory and Active Focus means the model re-reads a large amount of history at every step.
In the wild, where the edges showed up
Community tests
Someone generated a smooth Dragon Ball animation purely from code, with praise for the responsiveness. Harder users found limits. One developer asked Wonder Full for a AAA-grade lawn-mower simulator, and the grass and the character's hands came out crude, which suggests 3D construction is not its strength. Another, testing a rocket simulation, found the visuals behind a competitor, though the internals held up.
Meta's response to the criticism
Asked about the criticism, Meta's research lead Alexandr Wang acknowledged a real gap on some extreme programming tasks, said the team is working through the bugs, and suggested using the model on real software work rather than flashy demo scenes.
What blocks the model inside a daily pipeline
Pricing and quota
API pricing is $4 per million input tokens and $20 per million output tokens. Output has been optimised, but agentic work pulls in large background prompts and long memory, so the total still adds up. Some early subscribers report that a few heavy debugging sessions exhaust a monthly plan in the tens of dollars, pushing them to a higher tier. Cost control and channel choice move onto the desk.
Protocol gaps and a missing runtime
Plenty of engineering tools only speak Anthropic's format, others accept only the OpenAI standard, and pointing them straight at Meta's official API stops at the protocol layer. Occasional rate limiting on the official channel kills a long task halfway through, and the tokens already spent are gone. Then there is generated full-stack code with nowhere to run. The model writes backend services and table-creation scripts while the machine has no runtime ready to verify any of it.
The gateway layer that makes any of this workable
ServBay, a one-stop AI development management tool, runs on the machine as full-featured local AI gateway and stitches new and old models, interface protocols and development environments into one setup.
One endpoint across OpenAI, Anthropic and Gemini
ServBay runs a unified gateway compatible with OpenAI, Anthropic and Google Gemini standards, with bidirectional translation underneath that the client never notices. Streaming responses, tool calls and chain-of-thought reasoning are all preserved through the conversion. Whatever format an agent on top expects, the endpoint points at the local gateway and the model underneath can be Meta Muse Wonder. The client changes no code, and the newest model joins an existing workflow.
Channels and subscriptions in a single pool
Official keys, third-party relays and subscription accounts bought on different platforms no longer need copying and pasting between tools. Meta's official credentials, backup relay routes and the various subscriptions get registered once, and the local gateway handles scheduling and authentication centrally. Every development environment shares one backend pool and switches on demand.
Priority routing and hot switching
A long agent run of dozens of steps is most exposed to a 429 or a network wobble that voids the entire run. Channel priority can be set so stable, cheap or soonest-resetting channels come first. When Meta's official channel is rate limited or the network stutters, the gateway hot-switches to a backup channel or relay within milliseconds, without the client noticing, and brings the answer back cleanly. Built-in traffic and spend accounting shows what each project consumes.
Model mapping for tools pinned to old model names
Existing editor plugins hard-code model names and only recognise a few older ones. A single mapping rule in ServBay lets the gateway intercept a request aimed at an old model and redirect it to meta/muse-wonder. The tool notices nothing, and the thinking behind it has been upgraded to Meta's latest generation.
Virtual keys plus a local execution loop
To keep one account's credentials from getting out of hand, ServBay issues multiple local virtual keys. Each project gets its own credential with its own quota, while the master key stays in the local secure area. When Muse Wonder writes complex backend logic and table structures, the language interpreters, local domains, self-signed certificates and database services ServBay already provides take over on the spot. The code lands and runs locally for debugging, which closes the gap between generation and execution.
Two problems that benchmarks do not measure
Closed API against an open-source reputation
Meta built its developer goodwill on open-source releases. With Muse Wonder it has chosen a closed API and paid subscriptions. Management still repeats that the commitment to open source has not changed, while a second revenue line in AI software alongside advertising puts Meta in direct competition for the same customers.
Durable memory and permission scope
Long-term memory raises a question the community has been discussing. Some users noticed the assistant apparently reaching for private information, which Meta explains as authorisation clicked in an earlier conversation. Once a model has durable memory, a box ticked long ago can be treated as a standing pass in later tasks. Finer-grained expiry control and isolation for permissions are the obvious next requirements.
The ceiling of AI can be raised by olympiad problems and leaderboards, but its value is decided by whether it handles the small, tedious, real tasks of an ordinary day. Muse Wonder Full demonstrates real capability on memory and on not stalling halfway, and it still shows cracks in complex graphics rendering and cost control.
For working developers, the sturdy position is not belief in any single model's miracle. A stable local foundation plus a gateway dispatch point keeps model capability tamed, safe and steady on the desk, and that arrangement survives the next release cycle.
Top comments (0)