META JULY 2 TOWN HALL — TWO CONTRADICTORY MESSAGES
● Zuckerberg said: AI agents "hadn't accelerated in the way we expected" over 4 months. Reorg behind 8,000 layoffs "hasn't come to fruition yet." Expects results in 3-6 months.
● Wang said (minutes later): Watermelon, Meta's in-training model, has "caught up" with GPT-5.5. Uses "an order of magnitude more compute" than Avocado (Muse Spark).
● What's confirmed: Watermelon is in training. Wang's claim is based on internal benchmarks — specific benchmarks not named publicly.
● What's not confirmed: GPT-5.5 parity. No independent benchmark. Meta declined to comment.
● Market verdict: META stock -4.9% to $582. Investors trusted the confession over the promise.
● Meta 2026 capex: $125-$145 billion. The spending is real. The frontier results are still unverified.
The Town Hall That Told Two Stories at Once
Mark Zuckerberg told Meta employees at an internal town hall on July 2 that the company's work on AI agents had not accelerated in the way executives expected over at least the last four months, according to a recording heard by Reuters. He said the company's bets on its new structure "haven't come to fruition yet" and told staff Meta had miscalculated the timing of the changes. He said he expected Meta to see more substantial benefits from its AI spending within the next three to six months — placing a meaningful payoff toward the end of 2026, more than a year after Meta created its Superintelligence Labs unit.
Minutes later, in the same room, Alexandr Wang — Meta's superintelligence chief — told employees a very different story. According to two people familiar with the meeting, Wang said Meta's in-training model codenamed Watermelon had caught up with OpenAI's GPT-5.5 on closely watched AI benchmarks. "Watermelon uses an order of magnitude more compute than Avocado," he said, using the internal name for Muse Spark, Meta's April 2026 release. He added that a Muse Spark update was coming soon with improvements in coding and agentic capabilities. When asked when Meta would have a model comparable to Anthropic's Claude Opus, Wang replied: "pretty soon."
What Watermelon Actually Is — and What the Claim Requires You to Believe
Watermelon is Meta's next major model after Muse Spark (internally codenamed Avocado), currently in training. Wang's comparison to GPT-5.5 is based on internal benchmarks — the specific evaluations used have not been publicly named, and neither Meta nor OpenAI has confirmed the claim. To believe the "caught up to GPT-5.5" assertion at face value requires trusting internal benchmark results from the company making the claim about a model that has not been released.
The choice of GPT-5.5 as the benchmark target is also telling. OpenAI released GPT-5.5 in April 2026 — two months ago. The world has largely moved on. GPT-5.6 Sol/Terra/Luna exists and is in government-restricted preview. If Watermelon has caught GPT-5.5, Meta is claiming parity with a model that was already the benchmark two months ago, not with where the frontier sits today. Analyst Bindu Reddy noted on X: "The problem — the world has already moved on to GPT 5.6 Plus unlike Grok or Gemini, no one can even verify their model performance."
Why the Agents Stall Matters More Than the Watermelon Claim
Zuckerberg's admission about agent development is the more significant data point. Meta made roughly 8,000 layoffs and reassigned 7,000 employees to AI-focused teams, framing the reorganization as a way to fund AI infrastructure and capture efficiency gains from AI-assisted work. The CEO told his staff that the coding tools — specifically citing Anthropic's Claude Code — were expected to "speed development faster than they have." The agents did not perform as projected over four months.
This is the most honest public statement from a major AI company CEO about the gap between AI agent promise and AI agent reality in 2026. Zuckerberg is spending $125-145 billion this year. He reorganized his company around AI agents. He told his staff those agents haven't delivered what he expected. The three-to-six-month timeline for a payoff makes Meta's Q2 earnings call — expected in late July — a critical event: it will be the first earnings report where Zuckerberg's agent timing admission can be tested against actual business results.
Also: Tesla Caps AI Tool Spending at $200/Week — Except xAI
The Information reported on July 2 that Tesla will cap employee spending on outside AI tools at $200 per week starting July 6 — with one set of tools exempt: beta versions of Elon Musk's own xAI Grok. Tesla engineers can use any model they want as long as the invoice stays under $200 per week, unless it's Grok, which has no cap. The spending restriction applies to Claude Code, Codex, Cursor, and all third-party AI coding tools. The exception for xAI is not framed as a productivity decision — it is a loyalty-enforcing mechanism that routes Tesla's AI tool usage toward Musk's own company.
Sources: Reuters, Business Insider, Implicator.ai July 2-3, 2026 · Related: Grok 4 vs GPT-5.5 vs Claude Sonnet — who actually wins → · Claude Code vs Cursor →