The gap between Big Tech's best and best-adjacent just collapsed to zero — which means the moat everyone thought OpenAI had might've been a mirage.

The Summary

The Signal

Meta spent the past 24 months getting laughed at. Their models were fine. Good enough for some tasks. But nobody was building the next killer app on Llama when GPT-4 and Claude existed. Developers optimized for the best, not the cheapest or most open. That calculation just changed.

Wang's claim that Watermelon matches GPT-5.5 on unnamed benchmarks matters less than the trend line it represents. Meta is throwing 10x more compute at each successive model. They're not trying to be clever about architecture or training efficiency anymore. They're brute-forcing their way to parity through sheer capital expenditure. And when you're Meta, with ad revenue printing $40B+ in annual profit, you can afford to brute force things.

"Watermelon uses an order of magnitude more compute than Avocado."

The timing is telling. OpenAI released GPT-5.5 in April. Meta claims parity by July. That's a three-month lag, down from what used to be 6-12 month gaps between frontier and fast-follower models. The delta is shrinking because the playbook is known: more data, more compute, more RLHF. There's no secret sauce left. Just scale and capital.

But here's the part most people will miss. GPT-5.5 isn't OpenAI's current flagship. GPT-5.6 shipped last month — held back from general release by government request, which tells you everything about where capabilities are actually at. So Meta caught up to OpenAI's April model in July. But OpenAI is already past that. The race hasn't slowed. It's accelerated to the point where "catching up" is a treadmill, not a destination.

This matters for three reasons:

  • Model performance is commoditizing faster than the market expected
  • Open-source-ish approaches (Meta's models leak immediately) can now hit frontier-adjacent performance
  • The value is migrating from model weights to what you do with them — agents, workflows, integrations

The Implication

If Meta can match GPT-5.5 with enough compute, so can Alibaba, DeepSeek, and any other well-funded player. Foundation models are becoming the new cloud infrastructure — necessary but not sufficient. The companies that win Web4 won't be the ones with the best base model. They'll be the ones who figured out how to make agents that actually work, in production, for real tasks.

Watch where Meta deploys Watermelon when it ships. If it goes into Instagram and WhatsApp as embedded agents that billions of people use without thinking about it, that's a different game than another model release that developers ignore. Distribution beats performance when performance is table stakes.

Sources

Business Insider Tech