# Meta Says Its Unreleased 'Watermelon' Model Has Caught OpenAI's GPT-5.5

The claim that it matches GPT-5.5 is internal, based on a model still in training, and unverified, and it uses roughly ten times the computing power of Meta's Muse Spark.

- Published: 2026-07-05T10:42:30.164Z
- Canonical: https://polylog.news/ai/2026-07-05/meta-says-its-unreleased-watermelon-model-has-caught-openai
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [Meta AI](https://ai.meta.com/blog/introducing-muse-spark-msl/)

Meta's next model, internally codenamed Watermelon, has reportedly caught up with OpenAI's GPT-5.5 on closely watched benchmarks, according to [AI Post](https://t.me/aipost/7425), which cites remarks that Meta's AI chief, Alexandr Wang, made to employees. AI Post says the model is still in training and uses roughly ten times the computing power of [Muse Spark](https://ai.meta.com/blog/introducing-muse-spark-msl/), the smaller model that Meta's superintelligence lab released earlier.

Every important detail here is internal and preliminary. The comparison is Meta's own, the benchmarks are unnamed, the model is not finished, and no outside party has reproduced the result. A claim that an unreleased system matches a rival, delivered to staff by the lab's leader, is the kind of statement to treat as a directional signal rather than a measured result.

The compute figure is the more concrete data point. Ten times the training compute of Muse Spark, if accurate, is a large bet on scale at a moment when rivals are emphasizing efficiency gains that separate capability from raw computing power. It suggests Meta is spending its way back to the frontier rather than improving the training method itself.

## What this means

Internal claims that unreleased models match rivals are a recurring feature of the funding and talent cycle, useful for morale and recruiting long before any weights or numbers are public. The signal worth tracking is whether Meta's heavy spending on computing power for Watermelon produces an independently verified result, or whether efficiency-focused rivals reach the same capability for far less.

## What to watch

- A public release of Watermelon with named benchmarks and third-party evaluation, the only thing that would confirm the claim of parity with GPT-5.5.
- Whether Meta discloses the actual training compute and cost, which would show how much scale it took to close the gap.