Meta ships Muse Wonder and Wonder Full and the benchmarks say frontier — but five hours of real engineering work tells some users a very different story. Underneath the divided reviews sits a deeper divide: Zuckerberg isn't building a machine god, he isn't signing pacing letters, and he says things are about to get faster from here.
Meta is back at the frontier — and it's playing a different game than everyone else.
Muse Wonder's benchmarks say frontier; real-work tests from serious engineers say not quite — and part of the mess is that Wonder and Wonder Full share a name but not a quality bar. The deeper division isn't about the model at all. OpenAI and Anthropic are either trying to or feel they inevitably are building machine gods that inspire awe; Meta is trying to build what it views as just a normal technology. Thirty-four days after Muse launched, the frontier duopoly is back to five-ish labs — and the newest member doesn't share the others' theology.
In 2025 the Meta AI story was expensive researcher poaching and internal strife; most observers wrote the company out of the race. Now, in the season of the personal agent, Muse Spark powers a hit consumer app and the new Wonder and Wonder Full models are meant to compete head-on with, as Chief AI Officer Alexander Wang put it, the big boys.
Wonderful scored 79% on DeepSuite 1.1 — almost five points ahead of GPT-6 Astra and Opus 5.5 — led on agentic tests, and hit state-of-the-art on Harvey Lab legal work, Val's finance agent, and Automation Bench. Artificial Analysis put Wonderful tied for second overall, two points behind Opus 5.5, calling it a material gain just six weeks after Spark 1.3.
Muse Wonder runs $4.52 per task — nearly a 3× jump from Muse Spark 1.3, about a dollar more per task than GPT-6 Astra and roughly $1.50 cheaper than Opus 5.5. Meta is pricing like a frontier lab now, because on the benchmarks it is one.
Wonder flopped on the lawnmower-simulator-style 3D tests that dominate X, and the dunks flowed. But NLW argues the real problem is upstream: because 3D games and renderings look good on social media, frontier labs feel forced to optimize for a use case that doesn't represent real code work at all — and that's extremely net negative for the entire industry.
— Alexander Wang, Meta Chief AI Officer, responding to the 3D demo backlash. Wang's ask — "please try Muse Wonder on your actual coding work" — plus his hands-on replies to individual critics all week signaled an unusually engaged launch team owning its misses in public.
Every's senior-engineer benchmark — rewriting a vibe-coded slop database from scratch — scored Muse Wonder at 52%, far behind even GPT-6-1 Sol at 71%. Researchers running real programming tasks called the quality itself bad, and at least one prominent builder went back to Codex and Claude Code after five hours.
A recurring pattern emerged: good experiences with the full model, bad experiences with the slimmed-down Wonder — and because they share a name, the lesser model's results dragged down the better one's reputation. Some argued the two should have launched at different times; Wang's reply: "Fair feedback, DMing you to understand more."
Users on the $50 plan reported 13% of a five-hour limit consumed in under ten minutes, and some had to upgrade to the $200 plan mid-task. Wang said a bug may have been miscounting usage — but whatever the cause, it didn't help first impressions.
Theo put Wonder Full in his S tier alongside Opus 5.5 after it found an existing PR and recommended merging that instead of duplicating work; Charlie Holtz said it one-shot a real security vulnerability other agents missed. Clare Vo's category breakdown: bad at front-end design and most writing, but on agentic tasks it crushed — first of eight frontier models she tested.
Power users today run Instinct, Muse, GrokBot, and Dots side by side, but NLW argues that won't scale: coding agents deliver value without knowing anything about you, while personal assistants get more useful in proportion to what they know about you and the systems they can access. Most people will struggle to grant that access to even one assistant — let alone half a dozen.
Muse has been great at immediate tasks like booking flights but terrible at long-term ones — false-alarm meeting reminders, doom loops on Hilton's booking site. Wonder ships two features aimed squarely at that gap: active memory, which tracks what matters to you across conversations, and active focus, which keeps a task going when it gets hard, tries other approaches, and recovers from errors.
Zuckerberg, Wang, and the AI at Meta account all explicitly named professional agents alongside personal ones — a deliberate messaging campaign. That may not mean Meta wants to fight Claude Code for enterprise deals; more likely it knows a personal agent consumers truly invest in will inevitably do professional work — and small business, where Citrini Research sees Meta's real shot, is already deep in its apps.
For many, the story isn't the endpoint but the delta: Muse launched 34 days ago, and now Meta has joined the frontier. Ethan Mollick found the pace more impressive than the model, and even skeptics of the billion-dollar talent spree conceded the crack-team thesis worked. Markets agreed — Meta rose 4% on the news.
— Mark Zuckerberg, to the Financial Times. Zuckerberg declined to sign the Pacing the Frontier letter and counter-signaled with his essay 'The Future Is for Everyone.' Now he tells the FT that Meta has reached the frontier and is 'already working on that next step up' — and he declined to give a P(doom) number, saying AI can be engineered with normal guardrails.
OpenAI and Anthropic are either trying to or feel they inevitably are building machine gods that inspire awe; Meta is trying to build what it views as just a normal technology. Who's right remains the debate — but that is the fundamental dividing line in frontier AI right now.
2026 began with three US frontier labs; coding agents' dominance narrowed the field to two. Now, with Wonder — and if Gemini 4 Argon can do what Google says — it's back to five-ish. Though with more powerful models reportedly just behind closed doors, the duopoly could reassert itself soon.
Anthropic is serving Opus 5.5 at a verified 285 tokens per second on its upgraded Fast Mode, while OpenAI's docs suggest a 280 theoretical max for GPT-6-1 Sol — at a 50% price premium. Both are expensive and niche today, but as ultra-fast modes become standard, big latency shifts will change how we interact with models in ways we won't know until we get there.
Users report queries being routed away from Opus 5.5 to a new model with very impressive results, with both Opus- and Fable-style prompts affected — raising the possibility of a double-headed release timed ahead of Anthropic's IPO. The consensus from one New York AI dinner: Fable 5.5 will be stunning, but OpenAI beats it in December with Belle.
Both OpenAI and Anthropic have restricted testing of their most advanced internal models and could choose to pace the frontier by shipping mid-size releases like Opus and Sol — or we could see another round of acceleration as new, larger frontier models are unleashed.
The Information reports OpenAI is in talks to license the mothballed Sora model to a stealth startup founded by former Sora lead Bill Peebles, with ex-Disney chairman Jeffrey Katzenberg and former Dropbox CFO Sujay Jaswa as co-founders. Sora was shut down in April when enterprise coding won the compute-allocation fight; it's unclear whether the startup would get full access to the weights.
Decart's hardware-agnostic DOS stack claims 8× faster inference at one-eighth the cost, developed with early access to Amazon's Trainium3 chips. Amazon walked away at $6B last month and is now back at a 75% premium over Decart's last mark — preparing, it seems, for a world where hardware efficiency is the only benchmark that matters.
NVIDIA — already an $800M investor in Reflection AI — is exploring options up to a full acqui-hire after Reflection's Beam release; at the $25B March valuation, a deal would top the $20B Groq purchase as its largest ever. After buying Hugging Face and acqui-hiring Poolside and Essential AI, the open-source intent is unmistakable — though observers split on whether it means competing with the frontier labs or just fixing Nemotron.
Nadella called it Microsoft's biggest custom-silicon bet to date, with multi-gigawatt deployments starting next year to make Maia the primary inference system for Microsoft's own products. It lands squarely in the everyone-competes-with-everyone era: chip companies moving up the stack, application companies moving down, and the open question of what constraint — distribution, data, or power — margins get built around.