Meta says its next model matches GPT-5.5: the key is what Alexandr Wang doesn't say

By Business Insider (via Hacker News) · July 3, 2026.

Alexandr Wang, head of Meta Superintelligence Labs, told an internal employee meeting that the company's next model —codenamed Watermelon— has reached the level of GPT-5.5, OpenAI's flagship model launched in April. According to two sources cited by Business Insider, Wang based the claim on benchmarks 'closely tracked' by the industry, without specifying which ones. Watermelon succeeds Avocado (the internal name for Muse Spark, the model Meta unveiled in April with disappointing results) and, according to Wang, is trained with 'an order of magnitude more compute.' In parallel, in a public message on X, Wang announced an update to Muse Spark with improvements in coding and agentic capabilities, and when asked whether Meta would match Anthropic's Claude Opus in coding, he replied it would be 'soon.'

Context matters as much as the announcement itself. Meta has long been trying to convince developers and clients that its models belong in the top tier, without fully succeeding: Muse Spark fell short of OpenAI and Anthropic in the relevant comparisons. To reverse that, the company has launched a talent offensive with packages worth hundreds of millions of dollars per researcher and has raised its infrastructure spending forecast from $115-135 billion to $125-145 billion for this year, on chips and data centers alone. It's Zuckerberg's most expensive bet to date, and Wang's message to staff —whether or not it's backed by public evidence— serves an obvious function: sustaining internal morale and curbing talent flight amid an all-out hiring war.

Here it's worth applying the same skepticism we demand when China claims to have 'caught up' with the Western frontier: a statement made in an employee meeting, with no named benchmarks, no independent verification, and Meta itself declining to confirm it to the press, is a statement of intent, not a measured fact. What's more, the target keeps moving: OpenAI already unveiled GPT-5.6 last month —though it hasn't yet released it broadly, at the request of the US government— so 'catching up' to GPT-5.5 could mean arriving late to the previous generation. It's the usual pattern in this race: each lab announces it has closed the gap right as its rival has already taken the next step.

Our take is that the competitive noise among Meta, OpenAI, Anthropic, and Google —however inflated in the short term by marketing and talent retention— is, on balance, good structural news. The more labs that approach the frontier, the faster the cost of intelligence falls and the harder it becomes for a single actor to control access to transformative capabilities; that pushes toward the abundance scenario we advocate for the long term. The short-term risk isn't that Meta is lying about its benchmarks, but that these announcement wars —spending figures nearing $150 billion a year on infrastructure alone— accelerate a concentration of resources and power among a handful of companies able to afford that race, leaving out those without that capital muscle. Technical progress is worth celebrating when backed by verifiable data; in the meantime, Wang's statement should be read for what it is: a signal of strategic intent, not a measurement of capability.