Last week Anthropic released Claude Haiku 4.5, and the headline number is worth sitting with: it performs in the neighbourhood of Sonnet 4 – a model that was state-of-the-art in May – at one third of the price and about twice the speed. Five months, in other words, is now roughly how long it takes for frontier capability to become the budget option.
A couple of weeks before that, Sonnet 4.5 arrived pitched explicitly as a model for agents – software that plans, uses tools, operates a computer, and works at a task for hours. Put the two releases together and you get the actual story of this quarter: highly capable AI is simultaneously getting much cheaper and much more autonomous.
Neither trend is news to anyone watching closely. But most governance programmes we see were designed under two assumptions that just quietly expired: that capable AI is expensive enough to be rationed, and that it waits for a human to prompt it.
When capability is scarce, gatekeeping worksLink to this section
Think about why AI governance in most organisations looks the way it does: a pilot group, an approved tool, a licence count, a review board for new use cases. All of that is rationing logic. It made sense when frontier capability cost real money per seat and arrived through procurement.
At Haiku-class prices, the economics invert. Capability this cheap doesn't arrive as a product you evaluate; it arrives as a feature flag in software you already own. Your CRM sprouts an assistant. The meeting tool starts summarising. The document platform quietly adds "ask AI about this file." Every SaaS renewal is now, potentially, an AI deployment – no procurement event, no pilot group, no review board meeting.
Your AI footprint is no longer the list of AI tools you bought. It's the fraction of your existing software that has switched AI on – and that fraction only moves one direction.
The governance consequence: gatekeeping has to give way to defaults. You can't review your way through a hundred embedded copilots one exception at a time. What scales is a small set of defaults applied everywhere – data-class rules that don't care which product is asking, vendor-AI questions built into every renewal (what model, whose infrastructure, is our data trained on, can we turn it off), and a register that tracks embedded AI, not just chosen AI.
The second inversion: AI as a userLink to this section
The agent half of the story deserves its own paragraph, because it breaks a different assumption – that AI only talks.
An agent that can operate a browser and use tools is, functionally, a new class of user in your environment: extremely fast, endlessly patient, credentialed with whatever access it inherited, and possessing no common sense beyond what it was given in writing. Cheap capable models mean you won't have three of these; you'll have hundreds, spun up casually because the marginal cost rounds to zero.
Identity and access teams have spent two decades learning to manage service accounts. Agents are service accounts that improvise. The playbook – least privilege, own credentials, logged actions, human accountability for outcomes – exists. The scale and the improvisation are new. Start small and boring: agents get scoped credentials, agents get logs, a named human answers for each agent's output. Culture first, tooling second.
Training at population scaleLink to this section
Here's where this lands on our desk. Pilot-group thinking produced pilot-group training: coach the fifty people with licences, catch everyone else at induction. When AI is embedded in the default tools, everyone is an AI operator – the coordinator summarising minutes, the bookkeeper's reconciliation assistant, the recruiter's screening helper none of you evaluated.
Training has to make the same move governance does: from artisanal to default. Baseline judgement for every seat – what data goes where, how much to trust output, when to escalate – refreshed as the tools shift under everyone's feet, with completion you can actually evidence. The craft content still matters for power users, but the floor is now the control that matters, because the exposure is now universal.
(Yes, delivering that floor affordably at every-seat scale is a problem we've been working on intensely. More on that another day.)
The planning assumption to updateLink to this section
If you take one line to your next leadership meeting, make it this: frontier AI capability now depreciates to commodity pricing in roughly half a year, and autonomy is following the same curve. Any plan that treats today's capability, price, or usage pattern as stable for a budgeting cycle is already wrong in a knowable direction.
Cheap, capable, increasingly autonomous AI in every tool, operated by every employee. That's not a scenario anymore; it's the current quarter. The organisations that will be fine are the ones that stop governing AI like a rare purchase and start governing it like electricity – universal, metered, fused, and handled by people who've been taught not to stick forks in the socket.