Meta is closing the AI gap — fast. At an internal town hall on July 2, 2026, Meta’s chief AI officer Alexandr Wang told employees that the company’s next large model, codenamed Watermelon, has caught up to OpenAI’s GPT-5.5 across several key benchmarks. Hours later, Wang took to X to tease a near-term update to Muse Spark that promises major gains in coding and agentic capabilities.
Watermelon: order-of-magnitude compute, frontier performance
Watermelon is the successor to Avocado — the internal codename for Muse Spark, Meta’s current flagship that launched in April 2026 after three delays. According to Wang, Watermelon “uses an order of magnitude more compute than Avocado,” signaling Meta’s willingness to spend massive infrastructure dollars to compete at the frontier.
The benchmarking parity with GPT-5.5 is significant. GPT-5.5, released by OpenAI in April 2026, scored 58.6% on SWE-Bench Pro — a standard coding benchmark where the current Muse Spark reached 52.5%. If Watermelon is matching GPT-5.5 across multiple benchmarks, Meta is finally erasing the performance deficit that has dogged its models since the Llama era.
The timing matters. OpenAI has already previewed GPT-5.6 to the U.S. government in late June 2026, though general release remains pending. Anthropic’s Claude Opus 4.8 still leads coding benchmarks at 69.2% on SWE-Bench Pro. Google’s Gemini 3.5 Pro is expected this month. The frontier isn’t standing still — but Meta is no longer watching from the sidelines.
Muse Spark update: coding and agents, “pretty soon”
Wang’s X post on July 3 confirmed that a Muse Spark update is imminent, with specific improvements in coding and AI agent capabilities. When asked by a user when Meta would have a coding model on par with Anthropic’s Claude Opus, Wang replied: “pretty soon.” He added that users would “like what we have cooking.”
This is more than rhetoric. Meta has acknowledged since Muse Spark’s launch that coding remains a persistent weak spot. The upcoming update appears designed to close that gap, potentially bringing Muse Spark closer to Claude Opus territory on SWE-Bench and Terminal-Bench 2.0.
The bigger picture: $145 billion bet starting to pay off
The revelations come at a pivotal moment. At the same July 2 town hall, CEO Mark Zuckerberg candidly admitted that the company’s AI reorganization “hasn’t come to fruition yet” and that agent development “hasn’t really accelerated in the way we expected.” Meta’s stock is down 11.7% year-to-date, and the company is spending up to $145 billion on AI infrastructure in 2026.
But Watermelon’s benchmark parity with GPT-5.5 — and the accelerated pace of model iteration it implies — provides the clearest signal yet that Meta’s massive capex cycle is beginning to deliver competitive models. Combined with the Meta Compute cloud business announced July 1, which aims to monetize excess infrastructure, Meta is building multiple paths to ROI on its AI spending.
The next milestones to watch: the Muse Spark coding update’s actual benchmark scores, Watermelon’s formal launch timeline, and Meta Connect 2026 in September — where the company traditionally ships product. For now, the gap is closing faster than most observers expected.
Sources: Business Insider, SiliconANGLE, Alexandr Wang on X, andrew.ooo