Meta's next LLM "Watermelon" reaches GPT-5.5 parity

Key Shifts

Meta’s next LLM “Watermelon” reaches GPT-5.5 parity: VP Alexandr Wang confirmed in an internal town hall that Watermelon — the successor to Muse Spark — has caught up with OpenAI’s flagship GPT-5.5, using an order of magnitude more compute than its predecessor. A Muse Spark update with major coding and agentic gains is also coming “soon.” Business Insider · SiliconANGLE
Microsoft launches Frontier Company with $2.5B and 6,000 employees: A new operating business focused on end-to-end AI transformation for enterprise customers — “AI engineering that amplifies and protects your intelligence.” Signals a shift from AI experimentation to measurable business outcomes. Microsoft Blog
China’s GLM 5.2 goes open-source, rivals Anthropic: Zhipu AI’s new model competes with Claude on major benchmarks. With a 1M-token context window, open-source availability, and “no regional limits,” it represents the second Chinese open-source shockwave after DeepSeek. Euronews
Global VC funding hits record $510B in H1 2026: The AI boom drove all-time-high venture investment. IPOs and M&A exits are surging alongside funding, suggesting the exit market is finally thawing. Crunchbase News · SiliconANGLE

Kagi introduces an AI toggle: The search engine’s July 2 update lets users switch AI features on or off. As fatigue grows over forced AI integration, this design pattern — giving users control — could become a competitive differentiator for B2B products. Kagi Changelog
China’s Kling AI raises $2B for AI video: Kuaishou’s AI video subsidiary secured funding from Alibaba, Tencent, and others. AI video generation is becoming a capital-intensive battleground, with implications for creator tools and media platforms worldwide. Bloomberg
US heatwave triggers AI data center energy alarm: Record temperatures are straining power grids already stressed by AI data center demand. Water usage and cooling costs are spiking, making infrastructure sustainability a real bottleneck for AI expansion. Al Jazeera

AI inference economics turn “obviously profitable”: Analysis by Sean Goedecke argues that per-token inference costs have fallen below API pricing, making the unit economics of inference businesses positive. If model providers keep competing on price, startups integrating AI into products will see improving margins. Sean Goedecke
Anthropic Fable export controls lifted, but China goes open-source anyway: The US dropped restrictions on Anthropic’s most powerful models, while China’s GLM 5.2 released with no regional limits. The tension between AI sovereignty and openness is becoming the defining policy question. NYT · WSJ

The “AI toggle” as a UX pattern: Kagi’s opt-in AI toggle shows a middle ground between forcing AI and hiding it. For B2B SaaS, offering AI features as a toggle rather than always-on could boost both user trust and adoption rates. Kagi Changelog
Open-source models signal GPU vendor diversification: GLM 5.2 and other recent open-source models are showing competitive performance on alternative hardware like AMD and Huawei silicon. AI infrastructure startups should evaluate multi-vendor strategies to reduce NVIDIA dependency. Euronews

When will Meta Watermelon actually ship?: Still in internal evaluation. The Muse Spark update comes first, but Watermelon’s real-world performance, API availability, and whether Meta maintains its open-source strategy are all open questions. Business Insider
AI VC bubble vs. sustainable growth: Much of the record $510B in H1 2026 is concentrated in AI. The dot-com comparisons are getting louder — the ability to distinguish AI companies with real revenue and unit economics from the rest is becoming a critical skill for founders and investors alike. Crunchbase News