Alexandr Wang: The Man Tasked with Bringing Meta to the Forefront of AI

When Mark Zuckerberg last year renamed the research division to Meta Superintelligence Labs and appointed Alexander Wang to lead it, he sent a clear signal — Meta is deadly serious about artificial intelligence. Wang, who previously served as CEO of Scale AI (a company specializing in training data for AI models), now leads a team of elite researchers known by the codename TBD.

It was Wang who, at Thursday's internal town hall meeting, publicly announced for the first time — though so far only to employees — a significant milestone: the upcoming Watermelon model achieves GPT-5.5 performance in internal tests, OpenAI's flagship from April 2026. The information was first reported by Business Insider, which relies on two sources familiar with the meeting.

From Avocado to Watermelon: Meta's Fruit-Themed Model Hierarchy

In April 2026, Meta introduced the Muse Spark model series, internally codenamed Avocado. The model performed solidly on benchmarks but couldn't match the absolute top represented by OpenAI or Anthropic. Watermelon is its direct successor — and according to Wang, it uses an order of magnitude (ten times) more computational power.

"Watermelon, our next model after Avocado, is currently in training. It uses an order of magnitude more computational power than Avocado," Business Insider quoted Wang's words from the internal meeting. Wang did not specify the exact benchmarks on which Watermelon achieved GPT-5.5 level — this is a crucial unknown that Meta will have to clarify over time.

An interesting detail: Wang hinted on network X on the same day that an update to the existing Muse Spark model will come "very soon" and bring significant improvements in programming and agent capabilities. When asked when Meta would have a coding model comparable to Anthropic's Claude Opus, he replied cryptically: "Quite soon. You'll like what we have planned."

125 to 145 Billion Dollars: The Cost of Catching Up

Catching up to the AI forefront is not cheap. Meta announced to investors this year that it expects capital expenditures of 125 to 145 billion dollars (approximately 2.7 to 3.2 trillion CZK) — most of this sum is directed towards chips, data centers, and other infrastructure for artificial intelligence. This is an increase from the original estimate of 115–135 billion, which the company justified by rising component prices and further investments in data centers.

Meta is also aggressively hunting for talent. According to earlier Business Insider reports, it offers top AI researchers compensation in the hundreds of millions of dollars. This approach resembles an arms race — with the difference that this time the weapons are graphics processors and machine learning PhDs.

Meta Caught Up to GPT-5.5 — But OpenAI Already Has GPT-5.6

Here comes the key context. GPT-5.5, against which Watermelon is reportedly measured, was released in April 2026. In late June, OpenAI introduced GPT-5.6, its most powerful model to date — though so far only in a limited preview version, at the request of the US government, which wants to security-vet the models before widespread deployment.

In other words: Meta with Watermelon is catching up to a leader that is already a step ahead. This doesn't mean the success isn't real — if Watermelon truly achieves GPT-5.5 performance, it's the biggest leap in Meta's model quality since Llama 3. At the same time, however, it shows how quickly the AI race is progressing.

What This Means for Users — and for the Czech Republic

Meta models are not just an academic matter. They are integrated into the Meta AI assistant, which runs across Facebook, Instagram, WhatsApp, and Messenger. These platforms are used by tens of millions of people in the Czech Republic. If Watermelon (or its commercial version) arrives in these services, it could mean a significant leap in the quality of the AI assistant for ordinary Czech users.

Currently, Meta AI is not officially available in Czech — Meta primarily makes its AI assistants available in English and several other world languages. However, with the approaching effective date of the EU AI Act (fully effective from August 2026), we can expect Meta, like other major players, to have to adapt its models to the European market — whether it's language localization or meeting regulatory requirements for transparency and safety.

Moreover, it's not just about a chatbot. Wang explicitly mentioned agent capabilities — meaning the AI's ability to independently perform multi-step tasks, such as booking an appointment, searching for and comparing products, or managing a calendar. If Watermelon truly elevates these capabilities to the GPT-5.5 level, the AI assistant in Meta applications could start competing with specialized agents from OpenAI or Anthropic.

Open versus Closed Path

The strategic question is: will Watermelon follow the tradition of open models like Llama, or will Meta go the route of a proprietary solution? From Wang's statements, it's not yet clear, but the fact that Watermelon uses ten times more computational power suggests that its operational costs will be enormous — which speaks more for a closed deployment within its own services.

This would be a departure from Meta's current strategy, which has long relied on the open-source community with its Llama models. For European researchers and startups building their products on open models, this would mean a significant loss.