Meta’s chief AI officer Alexandr Wang told employees during an internal town hall on July 2 that the company’s next frontier model, codenamed Watermelon, matches OpenAI’s GPT-5.5 on key benchmarks. If the claim holds under independent evaluation, it would mark the closest Meta has come to matching the industry’s leading closed-source system.
The disclosure, first reported by Business Insider and confirmed by Benzinga, represents a significant escalation in Meta’s AI ambitions and a direct challenge to the pricing power that OpenAI and Anthropic command in the enterprise market.
An Order of Magnitude More Compute
Watermelon is the successor to Avocado, Meta’s internal codename for Muse Spark, which launched in April as the company’s first entry in its current model family. Wang told staff that Watermelon uses “an order of magnitude more” compute during training than its predecessor, a statement that only makes sense in the context of Meta’s $200 billion Hyperion data center investment in Louisiana and its fleet of roughly 600,000 H100-equivalent GPUs already in production.
The compute claim matters because it positions Meta squarely in the scaling-laws camp at a moment when some researchers argue that simply throwing more hardware at training runs yields diminishing returns. Meta is betting the opposite: that with sufficient compute and data, an open-weight model can match proprietary systems that charge $15 to $60 per million output tokens.
The GPT-5.5 Benchmark, With Caveats
Wang’s comparison target is notable. OpenAI released GPT-5.5 in April and has since shipped GPT-5.6, though the newer model’s rollout was paused at the Trump administration’s request over national security concerns. By benchmarking against 5.5 rather than 5.6, Meta is claiming parity with a system that, while powerful, is already one generation behind OpenAI’s internal frontier.
The specific benchmarks Wang referenced have not been publicly disclosed, which makes independent verification impossible for now. AI benchmark selection is notoriously susceptible to cherry-picking, and internal claims made during employee town halls carry less weight than peer-reviewed evaluations or third-party leaderboards.
Still, even a rough equivalence to GPT-5.5 would represent a dramatic capability jump for Meta’s AI efforts and a credible commercial threat. If Watermelon ships as an open-weight release following Meta’s established pattern, enterprise users could deploy GPT-5.5-class intelligence without per-token API fees.
Why This Matters for the AI Market
The competitive dynamics here are straightforward. OpenAI generated over $11 billion in annualized revenue as of early 2026, largely on the strength of API pricing that reflects its models’ capability lead. Anthropic charges similar premiums. If Meta can deliver equivalent performance through open-weight releases, it collapses the pricing umbrella that funds both companies’ research budgets.
Meta’s business model makes this play rational. The company monetizes AI through advertising efficiency and user engagement on Instagram, Facebook, and WhatsApp, not through API access fees. Every dollar that GPT-5.5-class capability costs in the open market is a dollar Meta can eliminate for developers building on its platform.
For enterprise buyers currently paying premium rates for frontier intelligence, a credible open-weight alternative at GPT-5.5 caliber would fundamentally shift procurement conversations. The question is no longer whether open-source can compete with closed-source on capability. It is how long the remaining gap lasts before training runs close it entirely.
What Open-Weight Parity Means for Enterprise Procurement
The practical implications extend beyond pricing. Open-weight models give enterprises something closed APIs cannot: full control over deployment, fine-tuning, and data residency. A GPT-5.5-equivalent model that runs on-premises eliminates the compliance risk that has kept regulated industries like healthcare, defense, and financial services cautious about frontier AI adoption.
Meta’s prior releases in the Llama family already demonstrated this pattern. Llama 3.1 became the default choice for companies that needed frontier-class reasoning without sending proprietary data to a third-party API endpoint. A Watermelon-class release would extend that logic to the very top of the capability curve, potentially unlocking use cases that compliance teams previously blocked.
The Compute Arms Race Continues
Meta’s announcement arrives alongside a broader industry pattern. Microsoft committed $2.5 billion this same week to a new AI implementation unit. OpenAI proposed giving the U.S. government a 5 percent equity stake worth $42.6 billion. The frontier AI industry is simultaneously consolidating capability and distributing access at a pace that makes quarterly planning nearly impossible for every adjacent sector.
Watermelon’s release timeline has not been disclosed. But with 600,000 GPUs already running and over a million accelerators planned for annual procurement, Meta has the infrastructure to push training runs that few organizations on Earth can match.