Google shipped Gemini 3.8 Flash and a locked-down security twin

That plays to its one structural advantage. Google designs its own TPUs and runs its own data centres, so its cost per token is not set by Nvidia's margin. Every competitor buying accelerators — OpenAI, Anthropic, Mistral, and the compute providers serving them — pays a supplier whose most recent quarter was $96.2 billion in revenue at gross margins the whole industry funds.

The capability is competitive and the distribution is not. Ramp data cited by TechCrunch put Google at roughly 6 percent of US enterprise AI spending in August, against Anthropic's 43.5 percent and OpenAI's 39.7 percent, though Google notes those figures exclude large strategic deals. The Accenture joint unit announced this week, training up to 1,000 consultants as forward-deployed engineers, is the attempt to fix that.

Everyone has arrived at the same structure

The gated security variant is now an industry convention, established in a single week.

Anthropic shipped Mythos 5.1 on 1 September — the same model as Fable 5.1, released only to vetted cybersecurity and life-sciences organisations. OpenAI released Astra on 3 September to its Daybreak cybersecurity cohort first, having rated it critical for cyber capability under its Preparedness Framework. Google shipped Flash Cyber through Fairwind on 2 September.

Three labs, three vetting programmes, one week. The convergence reflects a real problem — the capability that finds vulnerabilities for defenders finds them for attackers — and a real commercial opportunity, since defenders will pay for a model too capable to release openly.

It also creates a tier of capability available only to organisations that can pass a vetting process, which is a meaningful change in who gets access to the frontier. Hugging Face's July incident response illustrates the cost: its engineers were blocked by closed-source models refusing to analyse attack telemetry, and fell back to open weights.

The claims are self-reported

Both headline figures — more than 70 percent on vulnerability discovery, 2.6 times more correct Chrome patches — are Google's own, on Google's own internal evaluation, with no published methodology.

The Chrome comparison is the weaker of the two as evidence, because Google owns Chrome. A model evaluated on patching vulnerabilities in a codebase its developer maintains, with access to that codebase's history, test infrastructure and internal issue tracker, has advantages that do not transfer to a customer's code.

That is not to say the number is wrong. It is to say it measures something narrower than it appears to, and Google patched an actively exploited Chrome JavaScript engine zero-day the following day.