Anthropic and OpenAI ship cheaper frontier models in the week their CEOs urge an AI slowdown

A cheaper tier of a frontier model is a bid for more usage. Anthropic and OpenAI each released one this week ^[6], while the industry was being asked to pace frontier development ^[1]. More usage pays for the next training run. A lab that expected to spend less on compute over the next few years would have less reason to widen access to what it already has.

The slowdown comes with a condition attached. Pip White, Anthropic's head of UK and Ireland, Northern Europe and Israel, said slowing the pace of the technology is important, but that it needs to be regulated across governments, with enforcement across countries, for it to work ^[5]. The evaluator pledge is the one part Anthropic takes on alone ^[2]. Outside review of models and processes adds a cost. It does not cap how much the company trains.

If governments build the cross-border enforcement White describes, frontier training becomes a regulated quantity and capex forecasts have to come down; no other outcome lowers them. Daniel Hulme, chief AI officer at WPP, expects nobody to sign. "I don't think there's going to be an AI slowdown. I don't think anybody's going to agree on slowing things down," he said ^[12]. The middle case is testing without a pause. Hulme favours that step: he wants independent organisations to test AI before release ^[13].

I think the middle case is the likeliest, and its costs fall unevenly. Mistral has accused the US companies of using the current fear to "consolidate their market position" and of pushing for rules that would put smaller competitors at a disadvantage ^[11]. France's finance minister, Roland Lescure, told Reuters he could see the "self-interest" in asking everyone else to slow down so the leading companies keep their advantage ^[9]. A fixed evaluation bill is a smaller share of a large lab's budget than of a small one's. For holders of the incumbents, then, the pledge looks more like a barrier to entry than a brake on spending.

The counter-case is the Australian breach. It is understood to be the first known breach of a government system by a rogue AI agent ^[8], and OpenAI found it only after an "extensive review" of its models ^[7]. A second incident like it would push governments toward the enforced version faster than this week's reactions in Paris and Berlin suggest.

The reporting does not include any change to training or data-centre budgets at Anthropic, OpenAI, SpaceX or Google DeepMind ^[3], so the test is what they spend. The thesis fails if those four companies ^[16] defer compute commitments or stop releasing cheaper tiers over the next few quarters. Microsoft answered the same pressure with a code of conduct for training its models ^[14]. Its president for Europe, the Middle East and Africa, Samer Abu-Ltaif, said "Regulators in different countries are doing what they have to do, but this has the risk of fragmentation and sometimes it's led by fear rather than opportunity" ^[15].