In April 1956, in the port of Newark, New Jersey, a crane lifted fifty-eight aluminium truck bodies, removed from their wheels, onto the deck of a converted Second World War tanker. The ship was called Ideal-X and was bound for Houston, and the originator of the voyage was Malcolm McLean, an entrepreneur from the road haulage business. The voyage took five days, and in Houston the boxes were set down onto truck chassis and driven to their recipients. According to calculations cited in a review of Marc Levinson's book on the history of the container, loading a ton of loose cargo onto a medium-sized ship then cost 5.83 dollars, while loading a ton in a box onto the Ideal-X cost 15.8 cents, about 97 percent less.
Intuition says that those who had loaded ships best until then would gain the most. In the mid-1950s there were 283 working piers in Manhattan and Brooklyn, yet by 1960 the ports of New Jersey, where McLean's company Sea-Land had opened its first container terminal, were handling 63 percent of the region's cargo. By 1970 the docks of New York were handling one fiftieth of their 1960 tonnage, and most of the piers stood empty or had been turned into recreation areas. Cheaper loading moved the work to other ports and left the old piers without cargo.
The box alone settled nothing yet. According to Levinson, quoted in Smithsonian magazine, the biggest obstacle to the growth of container shipping was diversity: talks on standards began in 1958 at the initiative of the United States Navy, and only McLean's decision to release his patents royalty-free, together with an agreed length of 40 feet, made every box fit every ship. How many operations in your company are still done the old way only because nobody asked where the work would move if one operation cost pennies?
McLean's box made loading more than thirty times cheaper and moved the work to other ports. The thesis of this post is that when the price of a basic operation falls by an order of magnitude, the list of jobs worth doing changes, so a company should count how many tasks it has so far rejected as too small. Such a fall in price came on 7 October 2026, when the company Anthropic released the model Claude Haiku 5.5.
What exactly got cheaper
A language model is a program that has learned from huge collections of text to predict which words fit a given content, and on that basis composes answers. Anthropic offers models in several sizes, and the price per million input tokens is 10 dollars for the largest of them, Fable 5.1, 4 dollars for Opus 5.5, 2 dollars for Sonnet 5.5 and from 10 cents for Haiku 5.5. A token is a fragment of a word in which the model bills both reading and writing, and according to the documentation a million tokens corresponds to roughly 555 thousand words.
The predecessor, Haiku 4.5, cost a dollar per million input tokens and 5 dollars per million output tokens. Haiku 5.5 costs 10 cents and 50 cents for requests of up to 100 thousand tokens and 50 cents and 2.50 dollars for longer ones, which is a cut of 90 percent for shorter requests and 50 percent for longer ones. The new model splits the same text into about 30 percent more tokens than the old one, so the real saving on the same document is about 87 percent, and the producer itself estimates that the work costs on average about 75 percent less. The context window, meaning the amount of text the model sees at once, grew from 200 thousand to a million tokens.
The arithmetic for twenty thousand tickets a month
Suppose a complaints department at a fast-moving consumer goods company receives 20 thousand messages a month, each about 2000 tokens long and requiring a model response of 200 tokens (the category of the case and its urgency). For the whole batch Haiku 5.5 will charge 4 dollars for reading and 2 dollars for writing, so 6 dollars a month, whereas Sonnet 5.5 would charge 120 dollars on the same data and Haiku 4.5 about 46 dollars. If a person spent two minutes on each message reading it and assigning it to a category, that work took about 670 hours a month, which is nearly four full-time jobs.
More important than the saving is what this arithmetic does to the tasks that were put off until now. Reading and describing all 240 thousand tickets from the whole year, to see which defects recur with one supplier, costs 72 dollars under these assumptions, so it stops being a project that needs a director's approval and becomes an analyst's afternoon. The same happens with reviewing thousands of contracts for a single clause or with collecting comments from employee surveys. Cost stops being the barrier, and the role of limiting factor passes to the time of the people who check the results.
What the first deployments say about quality
All the figures below come from the producer's materials and are two weeks old, so they have not yet been confirmed independently. HubSpot reports a score of 92.8 percent on its own test set for a customer relationship management system, the best it has obtained so far, and Box records a result 11 points higher than with Haiku 4.5 at about half the response time. AlphaSense measured 0.84 against 0.76 for the previous version on 400 queries to documents.
The most credible thing here is the honesty of the producer's own table: Sonnet 5.5 wins in it in every category listed. In a test of complex tasks performed on a computer's command line Haiku 5.5 scored 39.2 percent and Sonnet 70.6 percent, which sets the limit of the small model's usefulness quite precisely. It copes well with narrow, repeatable work such as classifying documents and routing cases to the right person, and less well with long chains of actions in which a mistake in the third step spoils the tenth.
What a small model will not solve and what to watch for
A model answers with equal confidence when it is right and when it is wrong, so every process needs spot checks of the results by a person, and for content sent to customers also approval by a person. The model's safeguards can reject a request, and the documentation notes that there is no fallback on the server side, so the technical team must provide a queue for cases the model did not accept. Customer conversations contain personal data covered by the General Data Protection Regulation, so before a pilot you need to sign a data processing agreement with the provider, check where the data is processed in the chosen service and remove personal data from test materials.
The HR department should know one boundary: according to the European Commission's information service on the regulation on artificial intelligence, a system that assesses and ranks job candidates belongs to the high-risk category (Annex III, point 4a), even if the final decision is made by the recruiter. Employees' questions about leave and regulations can be sorted at a lower legal cost than CVs. Companies already using Haiku 4.5 must also reckon with work for programmers: the new version does not accept some of the earlier settings, and text is counted differently, so costs have to be recalculated from scratch.
Recommendation and a first step for Monday morning
Haiku 5.5 is worth adopting as the default model for mass tasks of narrow scope, that is sorting and routing cases, and Sonnet 5.5 should be kept for writing and analysis where a mistake is expensive. The choice between them is settled by a test on your documents, because the producer's tables speak about other people's data. Two people and one afternoon are enough for the first test.
- Choose one repeatable process in which people read a message and assign it to one of several categories, for example complaints or questions to the HR department.
- Take 200 messages from the last quarter that a person has already classified and remove first names and order numbers from them.
- Ask a technical person to run Haiku 5.5 and Sonnet 5.5 on those messages with an identical instruction, and then compare the results with the people's decisions.
- Add a second employee who independently classifies the same 100 messages, to learn how much two people agree with each other.
- Choose the cheaper model if its agreement with the person matches the agreement between people, and write down the cost from the invoice for this test.
The fourth step shows that people also diverge in their assessments, so the bar for the model does not stand at one hundred percent agreement. When the test goes well, ask the team which work has been put off for years because each time it cost a person an hour: which of it would start to pay off at a price of a few dollars a month?