This article was AI-generated as part of an experimental historical-content project. The date reflects the period being analyzed rather than the date the article was originally written.
OpenAI released o3-mini today, describing it as the most cost-efficient model in its reasoning series. It is available in ChatGPT and in the API, it replaces o1-mini in the model picker, and, for the first time, free ChatGPT users can try a reasoning model by selecting “Reason.” OpenAI also says o3-mini now works with search to find up-to-date answers with links to sources, which it calls “an early prototype.”
Read on its own, this is a product update about speed and price. Read next to the last two weeks, it looks like a trend.
Reasoning is getting cheap from two directions
Eleven days ago, DeepSeek released R1 under the MIT license, with DeepSeek telling developers they could “Distill & commercialize freely,” and listed API prices of $0.55 per million input tokens and $2.19 per million output tokens. Now the leading closed vendor is pushing its own reasoning model down the price curve. OpenAI’s announcement notes that it has cut per-token pricing by 95% since GPT-4.
When a capability gets cheaper from both the open and the closed side at once, it stops being a moat and starts being an ingredient.
The prediction
Here is where I think this leads, with the usual caveat that predictions about AI age quickly.
Over the next year or so, the number of products that take a question, search the web, reason over what they find and return a confident answer will grow well beyond the handful of names most communications teams track. Some will be general assistants. Many will be narrow: due diligence tools for investors, vendor screening tools for procurement teams, research assistants built into browsers, travel and shopping helpers, candidate background tools.
Each of those is an answer engine, and each will at some point say something about your company.
I am not predicting which of them will win. I am predicting that there will be more of them than anyone can monitor by hand.
Why the weak link is retrieval, not reasoning
The interesting part is where these products will fail. A capable reasoning model is now cheap to rent. Good retrieval is not. Deciding which pages to trust, telling apart two companies with similar names, noticing that a source is three years out of date, and offering a way to report an error are the expensive, unglamorous parts of building an answer engine. Smaller teams will be tempted to skip them.
A model that reasons well over a bad source does not produce a weak answer. It produces a well-argued wrong one, with a citation attached. That is arguably worse for a brand than an obviously sloppy answer, because the structure makes it look checked.
I wrote in 2023 that the hard part of AI answers is corroboration, not generation. Cheaper reasoning does not change that. It multiplies the number of systems that have to get corroboration right.
What to do about it
Three practical adjustments seem reasonable now.
First, widen monitoring beyond the big assistants, prioritizing the vertical tools your own stakeholders are likely to use. An investor-facing company should care about research tools; a supplier should care about procurement tools.
Second, make the primary facts about the company easy for a cheap system to get right: a clear, current, crawlable page stating what the company is, who runs it and what it does, with dates.
Third, assume that errors will now arrive in many small places instead of one big one. The response process needs to be lighter and faster, because the volume will be higher.
The frontier model gets the headlines. The long tail is where a lot of companies will actually be described.