Google took the benchmark lead on Wednesday and then handed the keys to a very small room. Gemini 4 Argon is billed as Google’s most capable model, strongest in long-horizon software engineering, vulnerability remediation and knowledge work. But for now it goes only to partners in its Fairwind security program, who get to find and fix flaws in their own software before anyone else gets the model. Paid users follow after testing. The introductory price is $2 per million input tokens and $10 per million output, doubling later.

That gating is the story as much as the scores. Google’s own comparison reportedly has GPT-6 Astra still ahead on several coding and computer-use tasks, and Claude Opus 5.5 ahead on terminal-agent work, so “retaking the lead” is a lead on a chosen scoreboard. What is unambiguous is the release pattern: the most cyber-capable models now ship defenders-first, a norm that barely existed a year ago.

OpenAI went the other way on volume. At DevDay it announced more than twenty things, led by “dots,” always-on agents with their own computer and browser and connections to thousands of apps, plus GPT-6.1 Sol, an upgrade priced at about a fifth of Astra’s standard rate, and a public-beta Agents API with hosted execution and memory. The subtext, reported alongside the event, is that the newest flagship, Astra, is being withheld over security concerns. Ship the cheaper workhorse, hold back the sharp one. It is a sensible product strategy that doubles as an admission.

The money is not slowing for the caution. OpenAI is reportedly seeking at least $30 billion at roughly $1.4 trillion, a bridge round now that Sam Altman has pushed the IPO to 2027. Talks are early and terms may change. Still, a valuation that nearly doubles in six months, from a company projecting heavy negative free cash flow for years, tells you how the market is pricing the race.

Regulators have noticed the same dissonance. The FTC has opened a broad investigation into the safety of systems from Anthropic, OpenAI and other labs, reportedly with formal demands for information expected within weeks. It is the first formal US regulatory step aimed at the rogue-agent incidents of the past few months. A voluntary industry accord was signed the same day, but a voluntary pledge and a subpoena are different instruments.

The through-line: capability is being rationed (Argon to defenders, Astra to nobody), prices are falling on the tier below, and the people with the checkbooks and the subpoenas are both arriving at once.

A note on today’s sourcing: this edition leans on search-result summaries rather than full page reads, so briefs are fewer and publication dates are day-level.