The Artificial Intelligence (AI) price war is shifting attention from model names to the cost of usable performance. OpenAI is reportedly using steep discounts to attract Anthropic customers, while companies are learning that inexpensive tokens do not automatically produce an inexpensive AI application. This matters to anyone buying AI tools, adding them to workflows, or trying to justify their recurring cost.
Why list prices are only the beginning
OpenAI reportedly offered some large customers discounts of more than 30 percent and, in one case, cost reductions of 80 to 90 percent. A report on the price competition between OpenAI and Anthropic names Replit, DataCamp, and Rippling as examples. Rippling increased OpenAI’s share of its coding work from 16 to 52 percent, while Anthropic’s share fell from 57 to 18 percent.
Replit switched after receiving a discount of more than 30 percent on a contract worth over $100 million annually. A US bank with more than $7 billion in assets received 500 rather than 250 ChatGPT seats at no additional charge. According to the report, this brought the price below $7 per user instead of the usual level above $20.
These deals show that a publicly displayed list price may say little about what a large customer actually pays. They should not be treated as general market prices, however: the terms come from individual negotiations and have not been independently confirmed as permanently available rates. For smaller businesses and Swiss organizations, it therefore remains unclear whether comparable discounts are available or in which currency and contract structure they would be offered.
OpenAI is reportedly gaining market share through this strategy while sometimes accepting lower margins. The report also attributes price cuts for the Sol and Luna model families to more efficient models and infrastructure, although that explanation comes from the provider. Additional pressure is said to come from Chinese models by Deepseek and Kimi, which allegedly offer comparable performance at a fraction of the cost; that comparability is not independently established in the supplied report.
What token prices actually measure
A token is a billing unit for the amount of information processed by an AI model. When a company accesses a model through an application programming interface (API), a connection that allows applications to exchange requests and results, it often pays according to the number of processed tokens. Inputs, generated answers, and extra context can all increase consumption.
The overview of token costs in business AI applications cautions against equating lower unit prices with lower total spending. If more employees use a tool, requests become longer, or the system handles more complex tasks, overall consumption can rise despite cheaper tokens. The dry version of cost control is that a cheaper meter does not help much when it starts running considerably faster.
AI tokenomics describes the economics of using, delivering, and billing for AI models. The relevant issue is not simply how few tokens an application consumes, but what business value those tokens create. According to Netzwoche, the Tokenomics Foundation, established by the Linux Foundation in August 2026, aims to develop standards and methods for comparing actual AI costs with economic value.
A useful comparison therefore needs at least four elements: the unit price, real consumption, the cost of the surrounding infrastructure, and the usable result. For an internal assistant, that result could be the number of useful answers. For a document-processing system that runs overnight, the more relevant measure may be how many documents are processed reliably rather than how little one request costs.
Why the surrounding system changes the price
An AI model does not operate alone in a production application. A harness is the technical environment around the model, including instructions, tools, and management of conversational context. For coding agents, this environment enables a general chat model to work inside a collection of code files.
A preprint study of coding agents summarized by The Decoder found that switching among several harnesses changed results by little more than simply repeating the same test. Costs, however, increased by as much as three times. Because the result comes from a preprint that has not completed final review, it should not be generalized to every AI application.
The study still points to a practical risk: longer system instructions, specialized search tools, and automatic compression of earlier conversation can consume additional tokens without producing a corresponding increase in success. Repeated runs of the same setup also varied by as much as six points on the coding benchmark used in the study. Small differences on a leaderboard may therefore be less meaningful than their decimal places suggest.
A sponsored article in MIT Technology Review uses harness in a broader sense, referring to the full serving environment, including memory management, scheduling, and optimization. It argues that a customer-facing assistant designed for fast responses has different requirements from a document pipeline that can run overnight. An internal knowledge tool and a high-volume customer service platform may also use the same foundation model while generating very different infrastructure costs.
The two sources do not directly contradict each other, but they create a useful tension. The preprint warns against assuming that a more elaborate agent environment automatically improves results. The sponsored article argues that an appropriate serving environment can be necessary for reliability, speed, and scale. For buyers, the implication is that added complexity should be measured against the actual workload, not admired for the size of its technical feature list.
Pros and Cons of aggressive AI discounts
Pros:
- Lower entry costs – Discounts can make pilot projects or broader adoption within a company more affordable.
- Stronger competition – Price pressure among OpenAI, Anthropic, and other providers can reduce the cost of each unit of usage.
- Room to negotiate – The reported enterprise contracts show that list prices are not necessarily final when sufficient volume is involved.
- Better model selection – Companies can use different models for different tasks instead of assigning an expensive flagship model to everything.
Cons:
- Uncertain duration – A steep introductory discount says nothing about pricing after the contract ends or consumption grows.
- Shifted costs – Higher usage, longer context, and elaborate harnesses can consume the savings achieved on the model price.
- Provider dependence – Moving a large share of work to one vendor can complicate later negotiations and technical changes.
- Unproven returns – Lower AI costs do not automatically create higher revenue, margins, or measurable productivity gains.
What this means in practice
If you are getting started, choose one narrowly defined workflow and compare offers using the same task. Record actual consumption, the number of usable results, and every recurring cost for each test. One practical comparison would measure the cost of processing a batch of documents reliably rather than merely placing the token prices of two models side by side.
More advanced users can divide work into task categories. An analysis of the financial value generated by AI describes companies selecting different models according to the task instead of using an expensive flagship model by default. This approach, known as model routing, can reduce costs, but it must also justify the added work required for operation and quality control.
That distinction matters because about one-third of the German companies surveyed in a Bitkom poll cited by the article said they had been surprised by high AI fees. Token consumption had also risen considerably. The finding does not provide a separate rate for Switzerland, but it suggests a relevant purchasing question for Swiss companies: Does the contract clearly state spending limits, usage reporting, and the terms that apply after a discount expires?
Measuring value remains the harder part. Apollo chief economist Torsten Slok found no apparent improvement in profit margins or revenue per worker among what he calls the S&P 493, meaning the S&P 500 without seven large technology companies. TechCrunch summarizes his assessment as no financial payoff so far for corporate America outside the technology sector.
t3n reports the same analysis alongside the contrasting result that the seven large technology companies have benefited measurably. Their revenue per employee and margins have increased substantially since early 2023; margins are reported at about 30 percent, roughly three times their level through 2020. This observation does not prove that AI alone caused the gains, and it cannot be transferred directly to other industries.
An inexpensive AI model can make financial sense when it produces sufficiently good results for a defined task and the surrounding system remains lean. Large discounts improve that calculation, but they do not replace it. The unresolved risk is whether special pricing will last and whether rising token and infrastructure consumption will outpace the value that can actually be demonstrated.
Sources
- KI-Preiskampf: OpenAI lockt Anthropic-Kunden mit Rabatten von bis zu 90 Prozent ab – The Decoder, 2026-10-25
- Wie Unternehmen die Token-Kosten ihrer KI-Anwendungen unter Kontrolle halten – Netzwoche, 2026-10-25
- AI delivers ‘zero financial payoff’ to most companies, Apollo economist says – TechCrunch, 2026-10-25
- „Magnificent Seven“: Welche Unternehmen laut Apollo-Chefökonom bisher vom KI-Boom profitieren – t3n, 2026-10-25
- Aufwendige Harnesses machen Coding-Agenten teurer, aber nicht besser – The Decoder, 2026-10-24
- Designing an AI serving harness built to last – MIT Technology Review, 2026-10-22
Image: Kindel Media via Pexels