Meta’s frontier model codenamed Watermelon is claimed to have parity with OpenAI’s GPT-5.5 on unnamed, internal benchmarks, while consuming roughly ten times more compute than Meta’s Muse Spark during training.

Meta has been ramping up infrastructure spending, guiding investors to expect about $125-$145 billion in 2026 for chips, data centers, and related infrastructure, up from a prior forecast.

Wang signaled an upcoming Muse Spark update with stronger coding and more agentic abilities as Meta accelerates toward competitive capabilities in AI benchmarks.

Independent benchmarks and public evaluations matter; internal benchmarks alone are not reliable, and the absence of a public evaluation table leaves Watermelon’s frontier parity unverified.

The overall takeaway is that any parity claim is provisional and contingent on future public, reproducible results.

OpenAI has released GPT-5.5 and later GPT-5.6 with limited public access, reportedly restricted due to government considerations; Meta has not independently verified Wang’s claims and declined to comment.

The industry context shows a trend toward balancing accuracy, speed, deployment costs, safety, and context length, not just chasing benchmark parity.

Industry dynamics suggest a broader race where competitors push for deployable products and developer tools, not solely higher benchmark scores.

Key points to watch: Watermelon allegedly matches GPT-5.5 on internal tests, uses substantially more compute, and the claim remains unverified and not a basis for procurement decisions.

Meta’s CEO signaled concerns about the speed of adapting to AI changes and noted layoffs were not as clean as desired, reflecting pressure to deliver results.

A shift toward a few cash-rich giants dominating AI scale creates a winner-takes-all dynamic and raises barriers for smaller players.

Watermelon is described as bringing stronger reasoning, coding, language understanding, and multimodal capabilities, though public benchmark details have not been disclosed.