Key Highlights
- Alexandr Wang, Meta’s superintelligence chief, informed staff that the company’s forthcoming Watermelon AI model has achieved performance parity with OpenAI’s GPT-5.5.
- The Watermelon model utilizes “an order of magnitude” greater computational resources compared to Avocado, Meta’s internal designation for the Muse Spark model released recently.
- Despite encouraging internal developments, Meta’s stock declined 4.90% after the news broke.
- The company has raised its 2026 AI infrastructure spending projection to $125–$145 billion, surpassing the previous estimate of $115–$135 billion.
- Wang hinted at significant improvements in coding and agentic features arriving in a forthcoming Muse Spark release.
Meta (META) stock experienced a 4.90% decline on Friday, despite indications that the company’s artificial intelligence development is gaining momentum. During an internal town hall meeting, superintelligence chief Alexandr Wang informed employees that Meta’s forthcoming model, internally designated Watermelon, has achieved performance equivalence with OpenAI’s leading GPT-5.5 system — a development that, if confirmed, would represent a significant achievement for the company’s AI initiatives.
Wang referenced AI model benchmarking results to support his assertion, although the specific benchmarks utilized remain undisclosed.
Watermelon represents Meta’s subsequent model following Avocado — the code name assigned to Muse Spark, which launched in April. While Muse Spark demonstrated solid benchmark performance upon release, it remained behind the leading systems from OpenAI and Anthropic.
According to Wang, Watermelon employs “an order of magnitude” additional computational power compared to Avocado. Increased computational resources generally yield enhanced performance, though they also result in elevated operational expenses.
Enhanced Coding and Agentic Capabilities Incoming
Wang also shared some developments publicly. In a statement posted on X, he announced that an upgraded version of the current Muse Spark model will arrive shortly, featuring substantial enhancements to coding and agentic functionalities.
When a user inquired about when Meta would deliver a coding model comparable to Anthropic’s Claude Opus, Wang responded that it would happen “pretty soon” and suggested users would appreciate what the company is developing.
Meta has invested years attempting to narrow the performance gap with OpenAI, Google, and Anthropic. Substantial investments in processing chips, data infrastructure, and elite AI researchers have yet to establish market dominance — at least in the public sphere.
Zuckerberg brought Wang aboard last year to helm what became Meta Superintelligence Labs. Wang directs an elite research division internally called TBD, in addition to the company’s comprehensive AI and hardware operations.
Capital Expenditures Continue Rising
Meta has compensated leading AI researchers with packages worth hundreds of millions of dollars to attract them to the organization. The financial commitments extend beyond personnel.
The company informed investors that it currently anticipates spending between $125 billion and $145 billion in 2026 on processors, data centers, and supporting infrastructure. This represents an increase from the prior projection of $115–$135 billion, attributed to escalating component prices and expanded data center construction plans.
For perspective, OpenAI launched GPT-5.5 in April. The organization subsequently unveiled GPT-5.6 late last month — characterized as its most capable model to date — though widespread access has been restricted, allegedly at the US government’s behest.
Wang’s statements have not received independent confirmation, and Meta declined to provide commentary. OpenAI did not reply to requests for comment.
