← Back to feed

Meta’s next frontier model is still in training and has already matched GPT-5.5 on internal benchmarks, according to Alexandr Wang, the chief of Meta Superintelligence Labs. Wang made the disclosure in an internal company town hall reported by Business Insider, citing two sources familiar with the meeting.

The model is codenamed Watermelon — the successor to Avocado, which is Meta’s internal name for Muse Spark. Wang told employees Watermelon uses an order of magnitude more compute than its predecessor. Meta declined to comment. OpenAI did not respond.

What Is Known

Wang described benchmark parity with GPT-5.5 but did not specify which benchmarks were used. That gap is material: internal benchmarks chosen by the lab running them carry different weight than external evaluations on standard leaderboards. The claim is notable as a signal of trajectory, not as a confirmed capability number.

The compute claim is more concrete. An order of magnitude means roughly 10x. Avocado shipped in April and debuted at number four on the Artificial Analysis Intelligence Index before settling below the frontier tier. If the 10x compute figure translates into proportional capability gains, the training mathematics would land Watermelon in the GPT-5.5 range — which is consistent with what Wang reportedly described.

GPT-5.5 launched in April at 82.7% Terminal-Bench 2.0 and 58.6% SWE-Bench Pro. It is currently the frontier ceiling that Anthropic’s Claude Fable 5 and Claude Opus 4.8 compete against. Parity at that level would mark the first time Meta’s own-weight frontier model has traded benchmarks with the leading US labs.

The Sequence

The model lineage matters for reading the claim:

• Muse Spark (Avocado): April 2026, debuted at AA Intelligence Index #4, fell short of GPT-5.4 and Claude Opus 4.6 in independent evaluations
• Next model (Watermelon): Still in training, 10x Avocado compute, Wang says GPT-5.5-class on internal tests
• GPT-5.5: April 2026, current benchmark reference at frontier tier
• GPT-5.6 Sol: June 2026, government-only access, higher than GPT-5.5

Watermelon claiming parity with GPT-5.5 would position it ahead of where Muse Spark landed but still behind GPT-5.6 and Anthropic’s Mythos-class models. The frontier has moved since GPT-5.5 shipped in April.

Why It Is Still Significant

Meta has spent roughly $145 billion on AI infrastructure in 2026 and burned through multiple talent acquisition cycles without fielding a model that competes with the leading frontier labs on neutral ground. Muse Spark was a credible Llama successor but did not close the gap.

If Watermelon delivers on the internal benchmarks — and more importantly, if those benchmarks map to external evaluations when the model ships — it would represent the first time Meta’s AI investment has produced a head-to-head frontier competitor. At 10x Avocado’s compute, the training budget alone signals this is not an incremental release.

Zuckerberg acknowledged at the same town hall that Meta’s broader AI ambitions are taking longer than expected. The Watermelon claim is the internal counter-signal: the next model in training is tracking toward the numbers the market has been waiting for.

Wang’s claim remains unverified by external evaluation. Public benchmark results will be the actual test.