Meta is finally catching up to OpenAI, its AI leader says

A Bold Claim From Meta's AI Chief

Meta's superintelligence lab head has made a notable public claim: the company's next flagship model, internally referred to as 'Watermelon,' is now performing on par with OpenAI's GPT-5.5 across key benchmarks. If accurate, this would mark a meaningful shift in a race where Meta has often been perceived as trailing behind OpenAI, Google DeepMind, and Anthropic in raw model capability, even as it has led in openness and distribution.

Why Benchmark Parity Matters

Benchmark comparisons are always somewhat contentious — companies choose which tests to highlight, and real-world performance doesn't always mirror leaderboard scores. Still, closing the gap with a frontier lab like OpenAI would be a significant milestone for Meta's superintelligence efforts, which have been the subject of massive investment, aggressive talent acquisition, and internal reorganization over the past year. Catching up on core capability, rather than just matching on cost or openness, suggests Meta's heavy spending on compute and researchers may be translating into tangible results.

The Open Source Angle

What makes this development particularly interesting for the open-source AI community is Meta's long-standing position as the major lab willing to release large, capable models with permissive licensing through its Llama family. If Watermelon approaches GPT-5.5-level performance and Meta continues its pattern of releasing weights — even partially open or research-focused versions — it could meaningfully raise the ceiling for what's available outside the walled gardens of OpenAI, Google, and Anthropic.

Historically, open models have lagged a generation or two behind the best closed systems. A genuine narrowing of that gap would have ripple effects: startups and researchers building on open weights could suddenly have access to near-frontier capability without relying on API access to closed models, potentially reshaping competitive dynamics across the broader AI ecosystem.

Reasons for Caution

It's worth treating this claim with appropriate skepticism until independent verification arrives. Executives touting their own models' benchmark performance ahead of release is a common industry pattern, and 'matching' a competitor on selected benchmarks doesn't always mean matching it in deployed, real-world use. GPT-5.5 itself is presumably still evolving, and OpenAI has shown it can extend its lead quickly when challenged.

What to Watch Next

The real test will come when Watermelon — or whatever Meta ultimately calls its release — ships and independent evaluators can run their own comparisons. Key questions include whether Meta open-sources the model or keeps it closed, how it performs on reasoning and agentic tasks beyond static benchmarks, and whether this narrows Meta's gap with rivals in actual product adoption, not just leaderboard standings.