{"id":6818,"date":"2026-07-03T08:50:50","date_gmt":"2026-07-03T03:50:50","guid":{"rendered":"https:\/\/cifrum.kz\/meta-watermelon-gpt-5-5-what-is-known\/"},"modified":"2026-07-03T23:46:07","modified_gmt":"2026-07-03T18:46:07","slug":"meta-watermelon-gpt-5-5-what-is-known","status":"publish","type":"post","link":"https:\/\/cifrum.kz\/en\/meta-watermelon-gpt-5-5-what-is-known\/","title":{"rendered":"Meta says Watermelon matches GPT-5.5, but has not disclosed the tests"},"content":{"rendered":"\n

Menlo Park, United States.<\/strong> Meta artificial intelligence chief Alexandr Wang told employees that a model under development, codenamed Watermelon, had matched OpenAI\u2019s GPT-5.5 on a set of industry benchmarks. Business Insider reported the remarks<\/a>, citing two people familiar with the internal town hall.<\/p>\n\n\n\n

The claim cannot yet be treated as a verified model comparison. Watermelon remains in training, Meta has not named the benchmarks, released scores or provided the system for independent testing. The company has issued no official public announcement about Watermelon.<\/p>\n\n\n\n

What Alexandr Wang reportedly said<\/h2>\n\n\n\n

Wang reportedly described Watermelon as the model following Avocado, the internal codename associated with Muse Spark. He also said Watermelon was using \u201can order of magnitude\u201d more compute than the earlier project. In ordinary numerical usage, that suggests an increase of roughly ten times.<\/p>\n\n\n\n

Business Insider did not specify what form of compute was being compared: total operations, accelerator count, training duration or a product of those measures. Without a methodology, the figure cannot be directly converted into model size, development cost or expected quality.<\/p>\n\n\n\n

Meta and OpenAI did not provide public comments on the comparison. The most accurate description for now is that Meta leadership reported an internal milestone that remains to be verified.<\/p>\n\n\n\n

Why matching benchmarks does not establish full parity<\/h2>\n\n\n\n

A score depends on task selection, dataset version, reasoning settings, the number of attempts and whether cost and response speed are counted. Two models can be close in mathematics while differing substantially in coding, tool use, multimodality or long-horizon reliability.<\/p>\n\n\n\n

Watermelon\u2019s training status adds another uncertainty. The final system can change after further training, behavioural tuning and safety evaluation. Until a technical report appears, it is not known whether the comparison used an intermediate checkpoint or a near-release configuration.<\/p>\n\n\n\n

Cifrum.kz has previously explained why matching an AI rival on selected cybersecurity tests<\/a> cannot be generalized to every capability. The same limit applies to the Watermelon claim.<\/p>\n\n\n\n

\"Infographic
Watermelon is still training, the benchmarks have not been named and no public independent verification exists. Infographic: Cifrum.kz.<\/figcaption><\/figure>\n\n\n\n

What is known about GPT-5.5<\/h2>\n\n\n\n

OpenAI officially released GPT-5.5 on 23 April 2026<\/a>. It published results covering agentic coding, computer use, professional tasks, scientific evaluations and cybersecurity. GPT-5.5 became available through ChatGPT, Codex and the API.<\/p>\n\n\n\n

OpenAI\u2019s published figures are also developer-reported and benefit from outside replication. However, researchers have benchmark names, evaluation conditions and access to the model. No comparable evidence package is available for Watermelon.<\/p>\n\n\n\n

GPT-5.6 has already moved the target<\/h2>\n\n\n\n

On 26 June, OpenAI began a limited preview of the GPT-5.6 series<\/a>. Its flagship Sol model is reported by the company to improve on GPT-5.5 across several agentic, biology and cybersecurity tasks. Broader availability is planned later, while the initial preview is restricted to a small set of partners.<\/p>\n\n\n\n

That makes GPT-5.5 a clear but no longer newest OpenAI reference point. Even if Meta\u2019s internal parity result is reproducible, it places Watermelon relative to the April model rather than the competitor\u2019s entire current lineup.<\/p>\n\n\n\n

From Muse Spark to Watermelon<\/h2>\n\n\n\n

Meta introduced Muse Spark on 8 April<\/a> as the first model in a new series from Meta Superintelligence Labs. The company describes it as a compact, fast system for complex reasoning and multimodal tasks. Muse Spark powers Meta AI, and a larger next generation was officially in development at the time of the announcement.<\/p>\n\n\n\n

The link between the Avocado codename and Muse Spark, as well as the Watermelon name, comes from media reports about internal projects. Meta\u2019s official pages do not yet give the next generation a public name, release date or model card.<\/p>\n\n\n\n

What ten times more compute could mean<\/h2>\n\n\n\n

More compute can support a larger model, more training data or a longer training run. Scaling does not guarantee a proportional quality gain. The outcome also depends on architecture, data, optimization algorithms and post-training.<\/p>\n\n\n\n

Meta\u2019s own explanation of its computing infrastructure<\/a> describes model training and serving as a combination of GPUs, custom chips, networks and data centers. A \u201ccompute\u201d figure without a unit is therefore a scale indicator rather than a technical specification.<\/p>\n\n\n\n

Meta expects up to $145 billion in capital expenditure<\/h2>\n\n\n\n

In its first-quarter results, Meta raised expected 2026 capital expenditure to $125\u2013145 billion<\/a>. It attributed the revision to higher component prices and additional data-center costs needed for future capacity.<\/p>\n\n\n\n

The range includes principal payments on finance leases and covers company infrastructure broadly. It should not be described as Watermelon\u2019s budget or assigned entirely to generative AI.<\/p>\n\n\n\n

What would verify Meta\u2019s claim<\/h2>\n\n\n\n