The top things worth knowing about in AI today.
OpenAI moved GPT-5.5-Cyber out of preview on 22 June, limiting access to verified security teams at firms including Cisco, CrowdStrike and Palo Alto Networks. The model scored 85.6% on the CyberGym benchmark, up from 81.8% for standard GPT-5.5, and can trace a vulnerability through a codebase, test a patch and prepare it for human review. OpenAI paired it with Patch the Planet, a program with Trail of Bits and HackerOne to fix flaws in widely used open-source projects. The same tooling that speeds up defenders also helps attackers, which is why access stays gated.
Read moreGemini co-lead Noam Shazeer left for OpenAI on 18 June, and AlphaFold lead John Jumper, a 2024 Nobel laureate, left for Anthropic on 20 June. Alphabet shed roughly $250 billion in market value on 22 June, its worst session in more than a year, as investors questioned whether Google can hold its most important AI staff. The exits land while DeepMind is seen as trailing Anthropic and OpenAI in AI coding tools.
Read moreSpaceX agreed to supply up to $6.3 billion of computing power from its Colossus data centre to Reflection AI, a startup building open-weight models as a Western alternative to China's DeepSeek. Reflection is separately in talks to raise $2.5 billion at a $25 billion valuation, up from $545 million a year ago, with backing from Nvidia and JPMorgan. The pitch: governments and banks that won't use closed US labs or Chinese open models want frontier-grade AI they can run themselves.
Read moreMicrosoft released Copilot Cowork worldwide on 16 June, letting Microsoft 365 customers hand multi-step work such as research, drafting and data tasks to AI agents that run in the background. Billing began the same day, with earlier Frontier-program tenants billed from 1 July. It moves Copilot from answering prompts to finishing longer tasks with less supervision, the direction most enterprise AI tools are heading.
Read moreChinese lab Moonshot AI released Kimi K2.7 Code under a permissive licence, a mixture-of-experts model with 1 trillion total parameters that activates 32 billion per token. It cuts reasoning-token use about 30% versus its predecessor and posts double-digit gains on coding benchmarks, with weights free to download from Hugging Face. Capable open models from China keep pressuring the pricing of closed US coding tools.
Read morexAI released Grok Imagine Video 1.5 on 17 June, generating 720p clips with synchronised audio, including dialogue, music and sound effects, in a single pass, and took the top spot on the Image-to-Video Arena leaderboard. A six-second clip renders in about 25 seconds, and pricing is $4.20 a minute against $30 for OpenAI's Sora 2. The cost gap matters for anyone producing video at volume.
Read moreMandatory requirements for how federal agencies use AI came into force on 15 June, part of the National AI Plan, with further rules due in December. The government has committed A$29.9 million to stand up an AI Safety Institute and is moving to ban AI-enabled dark patterns under proposed unfair-trading laws. Agencies now have to document and govern their AI use rather than adopt it informally.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.