The most important AI release of the week may be the cheapest one. Anthropic’s Haiku 5.5, its new small model, arrives priced roughly 90 per cent below its predecessor for prompts under 100,000 tokens, matching the price tier of OpenAI’s Luna class, and adds adjustable effort levels to the small-model line for the first time. The announcement also carried a 50 per cent cut to Sonnet 5.5 cache reads and new monthly API credits for subscribers.
Why does a price cut deserve the headline over a flagship launch? Because the industry’s centre of gravity is moving from demonstrations to deployments. The celebrated models answer hard questions in public; the small models do the economy’s actual work (classifying, extracting, summarising, routing) billions of times a day, where cost per call is not a detail but the product itself. A 90 per cent cut at that layer does not make headlines in the culture pages. It rewrites procurement spreadsheets in every company that has been running pilot projects at flagship prices and wondering how the pilot ever becomes the budget.
Plotly’s benchmark team reported its data-analytics score jumping two letter grades at one-ninth the cost, the kind of sentence that moves enterprise contracts. There is a catch, and developers found it within hours: the pricing cliff. Prompts above the 100K-token threshold carry a markup several times the base rate, which makes cost planning awkward for exactly the long-document work small models are bought for. Expect the cliff to be negotiated, imitated or flattened by competitors within the quarter; in this market, a pricing structure is a press release with a half-life.
Set the release beside its week’s context and a picture forms. OpenAI pushed GPT-6 as an interface event; Anthropic answered with an economics event; SoftBank reportedly assembled Gulf money at the $100 billion scale for the infrastructure beneath both. The flagship race continues, but the money, the developers and increasingly the strategy have moved to the layer where AI is priced like electricity: by the unit, at volume, on thin margins. Revolutions in computing have always been won by whoever made the unit cheap. This week, the unit got 90 per cent cheaper, and the small print (a token cliff, cache discounts, credit bundles) is where the war for the next decade of workloads will actually be fought.
