Analysis
Meta's Superintelligence Lab released two new models Monday, Muse Wonder and Muse Wonder Full, which the company is calling its first genuinely frontier-class systems since reorganizing its AI division under Alexandr Wang in August 2025. TechCrunch reported that Wonder is available now through the Meta Model API at $4 per million input tokens and $20 per million output tokens, while Wonder Full is reserved for subscribers to Meta's Muse Code product and isn't exposed through the API at all.
The gap between this release and Meta's last real swing at the frontier matters. Since last year's restructuring, Meta has mostly shipped Muse Spark, a lightweight, low-latency model built for speed rather than capability, while OpenAI and Anthropic kept pushing out new flagship systems on a roughly quarterly cadence. Wonder and Wonder Full are explicitly positioned as the larger, more capable counterweight to Spark -- bigger models aimed at complex tasks like coding rather than fast, cheap inference. The launch was also flagged in a same-day briefing from The Information, which covers Meta's AI division separately from TechCrunch's reporting on the model card itself.
Memory, Not Raw Power, Is The Pitch
Meta's own benchmarking, published in its model card, claims Wonder and Wonder Full match GPT-6 Astra and Fable 5.1 on general benchmark scores. The more specific claim is on MemoryBench, a test of memory and continual learning, where Meta says Wonder scores higher than Fable, Opus, and Astra. The company's blog post called the models "the next step on our scaling ladder and a major leap toward personal superintelligence," and research scientist Kavya Kopparapu framed the design goal in a launch video as making sure "your model can remember what's important to you" over long stretches of use.
That framing lines up with where a meaningful slice of agent-startup funding is currently pointed: systems that hold context across sessions instead of starting cold every time. Meta did not name Google in its comparisons at all, which is its own kind of signal about which rivals it's actually trying to out-benchmark right now.
Where It Sits Against OpenAI And Anthropic
- GPT-6 Astra (OpenAI) -- the model Meta chose to match on general benchmarks rather than claim outright superiority over.
- Fable 5.1 (Anthropic) -- the other named comparison point; Wonder reportedly ties it on overall scores but leads specifically on memory.
- Google's models -- conspicuously absent from Meta's own comparisons, suggesting Meta is picking its fight selectively rather than claiming a clean sweep.
On price, Wonder's $4/$20 per million tokens is, per TechCrunch's reporting, roughly in line with GPT-6 Sol and Opus 5.5 -- not a discount play, a parity play. Wonder Full's API-free distribution, locked behind a Muse Code subscription, is a different go-to-market than the API-first releases OpenAI and Anthropic have favored, and it keeps Meta's most capable model out of head-to-head public leaderboards for now.
What the launch undersells: Meta published no parameter counts, no context-window size, and no independent verification of the MemoryBench results. Every capability claim in this release comes from Meta's own model card -- the same self-graded setup that invites skepticism any time a lab says it has beaten its rivals on a test it designed itself. Price parity with GPT-6 Sol also says nothing about what the models actually cost Meta to serve at scale, a number the company has never disclosed even as its AI infrastructure spending has drawn scrutiny from investors this year.
For founders building products on top of frontier APIs, a fourth well-funded lab shipping at API price parity is one more reason persistent-memory features are becoming table stakes rather than a differentiator. The next real signal is developer adoption of the Muse API over the coming quarter -- not the launch-day benchmark chart Meta chose to publish.
