Cheaper only if you choose

Key findings

1. 01

   Inside one maker's price list, the strongest model costs up to a hundred times the lightest. On 13 November 2026 OpenAI listed a million tokens of output at USD 50 on its strongest general model and USD 0.50 on its lightest. Anthropic listed USD 40 and USD 0.50, and Google USD 10 and USD 2.50¹³⁴.

2. 02

   The makers advise against running everything on the top model. OpenAI's guide says to "keep the lightest setting that meets your quality bar"². Google's says to "pick the lowest level that meets your quality bar"⁵. Anthropic names its lightest model for simple tasks, and its second tier as the place to start³⁶.

3. 03

   AI gets cheaper only for a buyer who keeps choosing. The price of a fixed level of capability has fallen about thirteenfold a year since 2023. The estimate assumes a user who always takes the cheapest model that can do the task, and its authors add that real users "do not switch models so often"⁷.

4. 04

   Spending moved up the price list. In an investor's survey of nearly 400 US enterprise decision-makers, the top models' share of AI spending rose from 35% to 63% in a year, and their share of tokens from 18% to 42%. Reported spending averaged 4.7 times the year before, against a planned 1.9 times⁸.

5. 05

   People match the model to the task only loosely. Among paying users of one assistant, 51% of all use ran on the top model, and 55% of the use for computer and mathematical tasks⁹. On the same maker's developer service, a task that cost 1% more was used 0.29% less¹⁰.

6. 06

   In tests, choosing the model for each request saves anything from nothing to 98% of the cost, and the newest and strictest test finds at most a third. A 2026 benchmark of ten methods on 33 models found up to 31.7% saved at the accuracy of the best single model, and several well-known methods saved nothing¹¹.

7. 07

   Most of the gain comes from knowing the kind of work. The same benchmark traces a large part of it to telling broad kinds of task apart¹¹, and a study of the cost of a correct answer found a different class of model cheapest for each kind of task¹².

8. 08

   A lighter model is not cheaper when it fails. On 175 simulated office tasks of many steps, one maker's lighter model cost USD 0.6 a task and completed 11.4% of them. Its stronger model cost USD 4.2 and completed 30.3%¹³.

9. 09

   Nobody has published what a business saved. Among these sources there is no bill before and bill after, for a company of any size. Every saving above is a test result or an estimate.

How far apart are the prices?

Up to a hundredfold inside one maker's list. Makers price by the token, a piece of a word. A page of text is several hundred tokens, and what a model reads and what it writes are priced separately. On 13 November 2026 OpenAI listed its strongest general model, GPT-6 Astra, at USD 10 for a million tokens read and USD 50 for a million written, and its lightest, GPT-6 Luna, at USD 0.10 and USD 0.50¹. Anthropic listed Claude Fable 5.2 at USD 8 and USD 40, and Claude Haiku 5.5 at USD 0.10 and USD 0.50³. Google listed Gemini 4 Carbon at USD 2 and USD 10, an introductory price that doubles on 1 December 2026, and Gemini 3.5 Flash-Lite at USD 0.30 and USD 2.50⁴. These are the makers' list prices. SalexHub's list of models carried the same ones that day.

What that is for one task. Take a task that reads 2,000 tokens, about three pages, and writes 500. At the list prices above it costs USD 0.045 on OpenAI's strongest model and USD 0.00045 on its lightest. Ten thousand such tasks a month come to USD 450 or USD 4.50. The arithmetic is ours. A real bill will differ from it, for reasons the makers print themselves.

• Thinking is billed as writingGoogle's list prices output "including thinking tokens"⁴: the reasoning a model does before it answers is paid for, shown or not. In a study of 151 benchmark runs with reasoning switched on, a median 94.7% of what the generated tokens cost went on thinking¹⁴. Those were hard exam-style tests, so read it as a ceiling.
• A price per token is not a price per taskAnthropic's own guide says to compare models on the cost of a completed task. On one of its research benchmarks its strongest model cost 4.3 times as much as its third tier "for no measurable gain in score". On a coding benchmark its second tier, at twice the token price of the third, solved about the same share of tasks for about 7% more per solved task⁶. The maker calls these results internal and directional.
• Work that can wait costs halfAnthropic and Google list work that can wait up to a day at half price, and Anthropic lists text a model has read before at a tenth of the price or less³⁴. In records of model calls from more than a thousand engineering teams, 28% of the calls that could have used that discount did¹⁵.

Does AI get cheaper without anyone choosing?

The price of a fixed capability falls fast. A research group that tracks five benchmarks estimates that the cost of reaching a given score has fallen about 47% a quarter since 2023, which is thirteenfold a year⁷. A second group, working from two public price trackers, puts the fall at five to ten times a year¹⁶.

The fall goes to the buyer who moves. The first estimate follows the cheapest model able to reach each score. Its authors say what that assumes: a user who searches for the most cost-effective model for each task, "when real users do not switch models so often, and therefore do not reap quite the same savings"⁷. The second group adds the other half. Over the same years the price of running the newest top models rose between three and eighteen times a year, because the models grew and reason for longer¹⁶.

The money followed the top models. An investor's survey of nearly 400 US enterprise decision-makers, run in September 2026, estimates that the top models took 35% of enterprise AI spending a year earlier and 63% now, while their share of tokens went from 18% to 42%⁸.

A corporate card company saw the same turn among customers that track their token spending with its software. In April 2026 premium models were 45.8% of the tokens those businesses used and 55.9% of the cost. Ten months earlier premium models had been 5.7% of the cost¹⁷.

The bill outran the plan. In the investor's survey, reported AI spending averaged 4.7 times the previous year's, where respondents had planned 1.9 times. The median was 2.2 times against a planned 1.8⁸. The survey's own reading is that buyers put capability before cost when they pick a model. Our reading is narrower: a price list that falls thirteenfold a year did not lower these bills, because the work moved to the dearest line on it.

Why does a business stay on one tool?

Staying has three layers, and each has its own evidence. One model answers every request. One assistant is bought for every person. One supplier sends the bill.

• One model: the newest becomes the defaultIn a 2025 survey of 150 technical decision-makers, 66% had moved to a newer model of the supplier they already used, 23% had changed nothing, and 11% had changed supplier¹⁸. The move people make is up the same list.
• One assistant: a seat for everyoneAmong more than 85,000 US businesses that are customers of one corporate card company, the median business paying for AI spent USD 16.32 per employee a month in October 2026, less than one seat of an assistant each¹⁹. What that seat buys depends on the model its user picks, and the section on seats below takes it up.
• One supplier: a second is added, the first is keptIn February 2026, 16% of those businesses paid both OpenAI and Anthropic, up from 8% a year before, and about 79% of Anthropic's business customers also paid OpenAI²⁰. The share changing supplier in a month reached 8% in September 2026, the highest the company has recorded. Its page does not define a change²¹.

Teams that build with AI already run several models. In records from more than a thousand engineering teams, over 70% of organizations used three or more models, and two older models were each still in use at about a fifth of them, 19% and 22%, in March 2026¹⁵. Among businesses that track token spending with the card company's software, the median used nine models in April 2026¹⁷. Running several models is one thing, and choosing between them for each kind of work is another. These figures count the first.

Much of the choosing happens outside the business's own decision. In the United Kingdom, 55% of employees said they use AI for work or education, while about 35% of businesses reported using any²². In the European Union in 2025, 19.95% of enterprises with ten or more people used AI: 17% of the small ones, 30.36% of the medium and 55.03% of the large. Among enterprises that had considered AI and did not use it, the reason named most often was a lack of expertise, by 70.89%²³. A business that has nobody to compare models is the one most likely to keep the first it met.

Do people pick the model that fits the task?

A little. One maker sampled a million conversations from its assistant and a million from its developer service in February 2026, and set the model used against what the task would earn if a person did it. For every additional USD 10 of hourly wage, the share of conversations on its top model rose by 1.5 percentage points in the assistant and by 2.8 on the developer service. Among paying users the top model carried 51% of all use, and 55% of the use for computer and mathematical tasks⁹.

Price moves the choice very little. On the same maker's developer service a year earlier, a task that cost 1% more was used 0.29% less, once the kind of task was accounted for. The authors call business use of the service "surprisingly price-insensitive", and the analysis preliminary¹⁰.

The same seat is used very differently. Another maker reports of its enterprise customers that 19% of the people active in a month had never used data analysis, 14% had never used reasoning and 12% had never used search, and that a worker at the 95th percentile sent six times the messages of the median one²⁴.

The maker of the first study says of its own field: "Little is known about how AI users choose between different models"⁹. What is known comes from one maker's users choosing inside one maker's list.

What do the makers themselves advise?

Not to start at the top. The three largest makers each publish a guide to choosing among their own models, and none of them tells the reader to run everything on the strongest.

• Keep the lightest setting that passesOpenAI: compare the same inputs across models and "keep the lightest setting that meets your quality bar"². Google: run your own tests at every level, "then pick the lowest level that meets your quality bar"⁵.
• Start one tier downAnthropic's price list names its lightest model for simple tasks, its third tier for most production work and the tier above that for the most complex reasoning³. Its guide to cost starts the reader on the second tier, which in its tests matched the previous top model for about a fifth of the cost per solved task⁶.
• Tune one model before adding a secondThe same guide warns against building a mix of models too early: most workloads, it says, are settled by adjusting the effort of one model. In its tests, running every task at low effort and repeating only the failures at high effort cost about 40% less than running everything at high, at the same pass rate⁶.
• Test on your own workAll three say that their tables are a starting point and the reader's own tasks decide²⁵⁶.

How much does choosing the model for each request save?

In the first tests, most of the bill. A 2024 journal study sent each request down a chain of models, cheapest first, and stopped at the first answer a small checker accepted. Across five sets of classification and short questions it matched the best single model for 50% to 98% less: 98.3% on a set of news headlines, 52.3% on a set of science questions²⁵. A study presented the same year, which lets a small model check its own answer against a supplied text, reports over 50%. It was run with the large model assumed to cost 60 to 200 times the small one, and its authors place the point where the method starts to pay at a tenfold gap²⁶. A third study reports up to 40% fewer calls to the larger of two models with no drop in quality. That held where the two models were close. Where they were far apart, moving 40% of the requests cost 10.3% of the quality²⁷.

Under one fair comparison, a third at most. In 2026 a benchmark of more than 400,000 test items from 21 sets put ten published methods and 33 models, the current top ones among them, under one protocol. The best methods cut cost by up to 31.7% while matching the best single model. Several did no better than that single model, the early two-model methods among them¹¹. A 2024 benchmark had found the same of the methods it tested: none clearly beat a fixed mix of models that never looks at the request²⁸.

The saving depends on being able to tell a good answer from a bad one. A 2025 study works this out formally and names a good estimate of answer quality as the deciding factor. With the estimates that can be built today for general questions, its best method led the simple ones by at most 1.2%²⁹. A method trained on one kind of traffic chose at about the level of chance on questions unlike it³⁰. A 2026 survey of the field adds that most methods must be retrained when a new model appears, and that there is no agreed way to compare them³¹.

What is sold today. One cloud provider offers to choose between two models of one family for each request. Its product page claims savings of up to 30%, and its user guide says the choice is tuned for prompts in English only and cannot learn from a customer's own results³². That figure is the seller's own.

Nobody has measured it in a business. The nearest thing to a field study we found takes one feature of one product in use and tests whether small models could replace the large one behind it. A panel of people scored the best small model 5.6 out of 10 against 5.9, at an estimated fifth to twenty-ninth of the cost³³. A 2026 paper from inside one large company describes how it would move part of its coding assistant's work to cheaper models. About 98% of that spending sat with two top models, and the paper says it has no live results yet³⁴.

Which work needs which model?

The kind of work predicts it better than anything else measured. The 2026 benchmark found its leading methods hard to tell apart, and concluded that a large part of the gain comes from recognising broad kinds of task, not from judging each request finely¹¹. A study that prices the correct answer, not the token, found a different class of model cheapest for each kind of task¹². Put beside the makers' guides, the evidence sorts work into four kinds.

• Sorting, tagging, pulling values out of textThe lightest tier. OpenAI's guide puts simple data extraction and small edits on its lightest model at low effort². Where every model gets a task right, the cheapest one costs least per correct answer¹².
• Drafts, summaries, answers from material you supplyA middle tier. A small model can check its own answer when the source text is in front of it²⁶. For one short writing feature, small models came within 0.3 points of the large one on a ten-point scale³³.
• Calculation, code, planning in stepsA model that reasons, with the reasoning turned up. Reasoning models were the cheapest per correct answer on complex quantitative problems, for all their higher price¹². Reasoning paid most on step-by-step tasks and little on questions of recall, and in two cases the highest setting lowered accuracy¹⁴.
• Long tasks of many steps that nobody watchesThe strongest model. This is where the lighter one fails most often, and a failed task still has to be paid for and done again¹³⁶.

On those 175 office tasks a weaker model also took more steps: one open model cost nearly twice as much to run as a stronger closed one and completed fewer tasks¹³. Within a kind of work, difficulty still matters. In the one company's tests, a cheaper model passed every easy and every medium task of one coding category and 47% of the difficult ones³⁴.

What does a seat buy?

About the same price everywhere, for an amount of work nobody states. On 13 November 2026 OpenAI listed a business seat of its assistant at USD 20 a user a month billed annually and USD 25 billed monthly³⁵. Anthropic listed its team seat at the same two prices, and a premium seat at USD 100 and USD 125³⁶. Microsoft listed its assistant at USD 30 a user a month, billed annually, on top of the office plan it requires³⁷.

How far a seat goes depends on the model its user picks. No maker promises a number of messages. OpenAI publishes estimates for its work agent on a business seat: between 5 and 45 messages in five hours on its strongest model, and between 350 and 3,000 on its lightest³⁵. Anthropic's page says how fast a user reaches a limit depends on the model used³⁶.

Heavy work is moving off the flat fee. Anthropic's enterprise plan is a seat price plus usage billed at its developer rates³⁶. Microsoft bills its more advanced agents by usage, in credits³⁷. OpenAI sells further credits when a seat's allowance is spent³⁵. A seat is becoming a floor under a metered bill.

Spending is very uneven. Among the card company's customers that pay for AI, the top 1% spent a median of USD 10,199.20 per employee a month in October 2026, more than six hundred times the median business¹⁹.

Who should choose the model?

Our reading: the business, as a rule, by kind of work. The evidence for a system that judges every request is weaker than its headlines. The tested savings shrink under a fair comparison, depend on traffic the method has already seen, and have never been shown on a business's own bill. The evidence for sorting by kind of work is stronger and simpler: it is where the benchmark places most of the gain, and it is what the makers themselves describe.

• By kind of workDecide once which tier does background sorting and extraction, which does drafting and conversation, and which does long tasks. That is the split the evidence supports best¹¹¹².
• By roleThe top model's share of use rises with what a task is worth, by 1.5 to 2.8 points for each USD 10 of hourly wage⁹. A rule can make deliberate what people now do faintly: the strongest model for the roles whose tasks need it.
• By the task in handLeave a person the choice for a single hard task. The makers' advice assumes someone who can try the same input on two settings².
• By the billA rule needs a bill that shows who spent what, on which model. In a 2026 survey of 1,192 practitioners of technology cost management, 98% said they now manage AI spending, up from 31% two years earlier, and the capability they most wanted and did not have was fine-grained monitoring of it³⁸.

What this does not show

• Nobody has measured the overspend itself. No source here states what a business paid against what it would have paid with each task on the cheapest model able to do it. The price lists show how large the gap can be. They do not show how much of a real bill it is.
• The savings are test results. Every percentage saved comes from benchmarks or from an estimate. We found no before-and-after on a company's own work, and no study that ties the choice of model to a business result.
• The spending data comes from sellers and investors. The figures on spending, on the top models' share and on changing supplier come from one corporate card company's customers, who lean towards technology firms, and from an investor with stakes in makers of models. The card company also sells software for managing AI spending.
• Use and choice are measured inside single makers. What people use AI for, and whether they match the model to the task, is known from each maker's own users on that maker's own models.
• Small businesses are missing from the money. Official statistics say how many small businesses use AI. No source here says what they pay, for how many tools, or on which models.
• The prices are one day's. List prices are those of 13 November 2026. One of them doubles on 1 December 2026, and the makers change them often. The names of the models will age faster than the argument.
• Tokens are not one unit. A maker's newer models may count the same text as more tokens than its older ones, and a stronger model may finish a task in fewer steps. A ratio between two list prices overstates or understates the ratio between two bills.
• One source was read in abstract only, the preprint on the rising price of the top models. Its reference says so.
• We sell software in which a business chooses its models. Read our selection, and our reading of it, with that in mind.