We simulated an AI price war with 45 fictional buyers

The Experiment

Real AI prices have fallen hard this year. OpenAI and Anthropic cut flagship prices by up to 75% (our coverage), and cheap models such as MiniMax M3.1 Flash keep pushing the floor lower. The question nobody can answer from a price page is what happens next to the people paying the bills.

So we ran a thought experiment. We wrote a fictional version of the price war and handed it to MiroFish, an open-source tool that turns documents into a simulated society of AI agents, lets them argue for a while, then interviews them and writes a report.

Nothing in the world we built is real. The labs are invented: Northwind AI and Halcyon Labs play the closed-model leaders, Qilin Open Models publishes free weights, and Tessera Cloud resells everyone’s models to enterprises. We also invented 15 people who buy, build with, sell or are hurt by these models, from a solo developer in Bangalore to a procurement head in Lagos, a freelance technical writer in Osaka and a support team lead in Milan. We kept real companies and real people out on purpose, so no agent would put words in a real firm’s mouth.

MiroFish read our three documents and built 45 agents: our 15 people, the four fictional companies, and a crowd of group voices it pulled from the text, such as “banks”, “writers” and “developers”. The agents posted and replied on a simulated Twitter and a simulated Reddit for 40 rounds, producing about 160 posts, 270 comments and 630 actions in all. Then we asked each of them three questions: what have you changed since the price cuts, who wins and who loses, and what would make you switch providers.

Who Switched

Switching was concentrated in a small, loud group. The solo developer moved her meeting-notes app from Northwind’s flagship to a hosted open model in an afternoon and saw her margin jump from 40% to 75%. The student switched constantly and stayed loyal to nobody. One forum regular routed every request to whichever provider was cheapest that hour.

What the agents added on their own was a sharper idea of what loyalty means in a price war. The founder of a small inference host said his protection was “the exit path, not loyalty”: the weights are open, the configs live in a repository he controls, and he has tested rolling back. A startup founder put it as a design rule: a change of host or model “should touch only the model call.”

That flips the usual picture. In the simulation, the buyers who stayed put were often the ones who had made leaving cheap, not the ones who were trapped.

Who Stayed

Regulated and enterprise buyers barely moved on price. Bank, hospital and European company voices kept saying the same thing: price comes after data location, contract terms and whether the model will still exist in a year. The bank’s data scientist runs an open model on the bank’s own servers and cares more about evaluation paperwork than token prices. The insurer’s procurement head treated the cuts as leverage for bigger discounts and did not care which model sat underneath, “as long as the contract and the audit trail are right.”

The interviews also produced a practical list of what actually triggers a move for the cautious middle:

• A cheaper model that passes the buyer’s own tests on real tasks, so the cost per resolved case falls, not just the price per token.
• A host that changes its terms without notice, or looks like it is subsidising its prices.
• A model scheduled for withdrawal, or a change notice shorter than the buyer’s internal review.
• Data-handling promises the buyer cannot get written into a contract.

The cost of switching kept everyone else in place. One team in our setup re-tested its prompts for two weeks after a switch, found a bug that sent wrong refund amounts to customers, and switched back.

Who Gained and Who Lost

The simulated money moved away from the model layer and toward the people building on top of it, but unevenly.

• Gaining: builders who sell a finished result rather than raw model output, buyers who negotiate hard, and small operators who can now afford to build things that were too expensive a year ago.
• Squeezed: the inference hosts. The host founder admitted that most of his requests “now lose a little money, and I know it,” and warned that a supplier’s subsidised rate is “a timer.” Resellers that simply marked up tokens were hit early too; the sales-assistant founder cut his price 30% and now argues that “customers don’t pay for tokens.”
• Protected for now: the closed labs, through premium reasoning tiers and enterprise contracts, though Northwind’s own pricing manager worried that its cheap lite model was eating the flagship faster than planned.
• Losing: people paid by the hour or by the task. The support team lead watched four colleagues’ jobs disappear when first replies were automated, and the freelance writer’s rates fell by a third because clients assume AI does most of the work. In the simulation, the savings reached employers first and workers not at all.

Will Prices Bounce Back?

The sceptical journalist in our cast argued that this looks like earlier cloud and ride-hailing price wars, where investors paid for cheap prices until competition thinned out. The simulation took that seriously but never saw it happen. Several agents expected hosts to fail, and an open-source maintainer offered warning signs to watch: prices below what the GPUs really cost, rate cards that stop updating, and discounts tied to long commitments.

Its conclusion was that a rebound is hard to make stick while near-flagship open weights can be downloaded and moved to another host in a day. If one host dies, as one commenter put it, you move the same model somewhere else. The buyers most exposed to a future price rise are the ones least able to move: regulated companies whose contracts, data rules and validation work make switching slow.

What This Experiment Can and Cannot Tell You

A simulation like this is a structured way to think, not a forecast. Every agent is a language model playing a role we described, so a good share of the report repeats the backgrounds we wrote. The solo developer checks prices every week because we said so. Our interviews were also incomplete: about half of the 135 answers came back before the batch timed out.

What it is good for is pressure-testing an idea. The “exit path, not loyalty” argument, the switching triggers and the warning signs for subsidised pricing came from the agents, not from us, and they are worth checking against your own setup. If you are choosing models now, our Claude Haiku 5.5 analysis is a real-world starting point.

The Science Behind the Story

MiroFish is a free, open-source project under the AGPL-3.0 licence. It reads seed documents, extracts the people and organisations in them into a knowledge graph, and turns each one into an agent with a persona and memory. The agents then interact on simulated social platforms built with OASIS, the open agent-simulation framework from the CAMEL-AI team, and a report agent searches the results and interviews the agents before writing its findings.

Everything specific in this piece is invented: the four companies, the 15 people, their prices and their quotes. What is real is the tool, the method, and the price cuts that inspired the scenario. We ran the agents on Anthropic’s Claude Haiku 5.5. One practical lesson: the tool makes an agent out of every named entity in its sources, so feeding it real news turns real companies into characters. That is why we built a fictional world instead.

The Question It Leaves You With

If your own AI provider doubled its prices tomorrow, how long would it take you to move, and is that number something you have actually tested?