Meta ships Wonder, OpenAI holds back a model: the frontier race just split in two

Two very different frontier strategies

Meta has shipped Wonder and Wonder Full, a new "frontier" model family it claims rivals ChatGPT and Claude, per independent Artificial Analysis benchmarks. The catch: the models are also among the priciest to run, and they land first in Meta's Muse agent app - paid subscribers only. That's Meta continuing its push past ad revenue into frontier AI, betting the agent app is the delivery surface. (ET CIO)

OpenAI's posture is the opposite. CFO Sarah Friar confirmed an unreleased frontier model - one that already solved Navier-Stokes and generated 350 math proofs - is being held back for alignment work. She also flagged fast-advancing open-weight Chinese models as a cyber risk requiring better defender tooling. (Cybernews)

Meanwhile, China's domestic-chip bet got a concrete headline: Sugon's "Dengfeng" supercluster, built on 100k+ homegrown chips with immersion cooling at PUE 1.04, is now selling compute as a service and claims it could serve 5–10% of China's AI token demand - a direct rebuttal to the idea that export controls keep China out of the 100k-card club. (Paktechstan) Nvidia added fuel with a deeper ~$10B Europe bet covering four German startups and a Siemens demo claiming up to 85% less engineering prep work. (CNBC)

The quiet race: the agent economy

While labs argue over the frontier, this week's developer charts point elsewhere. On GitHub Trending, the entire top of the chart is the agent-skills/plugin economy: Anthropic shipped its official vetted plugin marketplace for Claude Code (claude-plugins-official), ruflo wraps Claude Code and Codex in a meta-harness with 100+ agents and self-learning memory (ruvnet/ruflo), and a 380+ skill library now spans Claude Code, Codex, Gemini CLI, and Cursor (alirezarezvani/claude-skills). The breakout repo is morluto/rea - agent-driven reverse engineering from app behavior down to native binaries - which blew past +229k stars in a week. (morluto/rea)

The Hacker News discourse matches: the practitioner consensus is that the real unit of capability is the model plus its harness - tools, loop, context - not the model alone (How LLMs work). And on Product Hunt the pattern is commercial: the day's notable AI launches are Slack-based "AI coworkers" - AskDoozy doing real team work (AskDoozy) and Scout by Fini for support teams (Scout by Fini) - agentic execution over chat, mostly vertical cuts of the same two patterns.

Signal to watch: Meta's "among the most expensive" Wonder pricing may limit real-world adoption while open-weight Chinese models keep advancing. The bigger bet is whether the agent-skills economy - marketplaces, harnesses, and vertical coworker agents - consolidates into the layer where the next round of AI value gets captured.