Labs race on models, policy, and cost controls

xAI has put Grok 4.5 into private beta at SpaceX and Tesla on a 1.5T V9 foundation model, with Cursor data added in supplemental training and internal claims of performance close to Opus. Meituan released LongCat-2.0 under an MIT license, pairing 1.6 trillion total parameters with 48 billion activated per token on average and a native 1-million-token context window. The model was also revealed as Owl Alpha, an anonymous OpenRouter model that had already gained developer traction.

Anthropic is tightening enforcement against Chinese company access to Claude after reported workarounds involving Singapore subsidiaries, VPNs, and overseas cloud infrastructure. Court emails also sharpened its dispute with the Department of Defense, showing Dario Amodei rejected contract language that could allow domestic mass surveillance and fully autonomous weapons. Anthropic separately added Claude Enterprise admin analytics, spend alerts, and a HackerOne bug bounty for Fable 5 cyber jailbreaks.

OpenAI has reportedly discussed giving the US government a 5% stake valued at approximately $42.6 billion as part of a broader standards and IPO strategy, while Crunchbase said global VC funding hit $510 billion in the first six months of 2026, with OpenAI and Anthropic taking $217 billion, or 43%. Meta’s Watermelon model reportedly matches GPT-5.5 on internal evaluations, and the UN and ITU launched an AI for Good Global Commission focused on developing-world benefits.