OSASI.ORGCommentary · Ideas for our profession
All briefs← Back to OSASI

It's fall. How do you budget AI for 2027?

September 24, 2026 · OSASI Commentary

In June, The Artificial Intelligence Show asked whether the era of affordable AI is over. Its answer: nobody knows how to price this yet, heavy users are being carried by light ones, and falling costs may rescue the flat-price plan. The host ended on a question for the fall: how do you even begin to budget AI for 2027? The fall has arrived.

How a brief is made

A member reads an article or listens to a podcast and sends the snippets, key thoughts or concerns tied to a passage or a timestamp. An AI assistant, chosen for the task and the setup (Claude, ChatGPT, Codex, Gemini or another), searches the full article or transcript, pins the times, checks the quotations, and assembles a draft. The draft goes through rounds of review, and nothing is published until a person approves that exact version. The cues on each post start the player at the passage discussed. Commentary is OSASI's perspective, offered for discussion, not a summary of the source. Each brief is co-authored by a person and an AI assistant. AI assistant for this brief: Claude (Anthropic).

Source podcast · The Artificial Intelligence Show

This episode prompted the commentary below. It is one perspective, offered for discussion.

The Artificial Intelligence Show · June 16, 2026
#219: Claude Fable 5, OpenAI IPO, Apple Siri AI Finally Unveiled & Is the Era of Affordable AI Over?

Listening highlights · Paul Roetzer and Mike Kaput. The segment is a five-minute rapid-fire item late in the episode.

−1:15:05 to −1:14:00 remainingIntroductions · Paul RoetzerWho is speaking: the host, his co-host Mike Kaput, and the episode recorded on June 15
−31:22 to −29:25 remainingThe subscription test · Mike Kaput$200 plans that deliver thousands in tokens at API prices; Meta curbing use; OpenAI weighing price cuts
−29:21 to −28:08 remainingWho pays for whom · Paul RoetzerPer-seat prices likely to rise, and light users who may be funding heavy ones
−27:57 to −27:00 remainingMinus 900 percent, and falling costsWhy a plan can be hugely profitable or deeply loss-making, and why cheaper models may save it
−27:00 to −26:28 remainingBudgeting for 2027 · Paul RoetzerOpen models, a moving target, and the budget season now under way

Tap a time above to start the player there. If Spotify shows a card about the show first, close it and the player carries on at that time. Times are approximate transcript times; ads and episode editions may shift them. This plays the full episode, not an extracted clip.

Who is talking

Paul Roetzer hosts The Artificial Intelligence Show; he introduces himself as founder and CEO of SmarterX and Marketing AI Institute, and his co-host Mike Kaput as SmarterX's chief content officer. The episode was recorded on June 15 and led with Anthropic's Fable 5; this brief concerns a five-minute rapid-fire item near the end, "Is the Era of Affordable AI Over?"

The show's own message

Kaput sets out the evidence. The research firm SemiAnalysis bought the consumer plans from Anthropic and OpenAI and ran long coding tasks until each hit its weekly limit: "a $200 a month Claude plan delivered up to roughly $8,000 a month worth of tokens at API pricing, a $200 ChatGPT plan went up to about $14,000". Used to the full, he says, those plans are "heavily, heavily subsidizing" their power users.

Roetzer's reading is that nobody yet knows how to price this. "I think you're just gonna see an overall increase in the per seat pricing," he says, with plans moving to credits or tokens. He then offers a guess, and says it is one: "I have no idea on this", but perhaps the top five or ten percent of users consume "90% plus" of the compute, so the subsidy for heavy users "might be actually be funded by people paying their 20 bucks a month who use the thing like five times a month".

Kaput adds the other side. At full use a top plan runs at what he calls "a negative 900% return", but at a small fraction of use it is highly profitable, and SemiAnalysis expects that falling costs will soon let the labs "profitably serve something like Opus 4.8 level models for 20 bucks a month."

June to September

Two items Kaput mentioned in passing have since played out. The Wall Street Journal reported in June that OpenAI was considering "drastic price cuts" in a coming contest with Anthropic; on September 22 it halved the API prices of its Sol and Luna models (OSASI's brief on that release: $600 to $7,000 per day). And in June The Information reported that Meta was moving to curb its employees' AI use as internal costs reached billions; by September OpenAI was publishing its own researchers' daily token use, valued in the hundreds and thousands of dollars. Two large companies, two opposite answers to the same bill.

Set the two sets of numbers side by side and the scale becomes clear. SemiAnalysis found that a $200 plan, pushed to its weekly limits, could use up to about $14,000 of tokens a month at list price. OpenAI says its median researcher uses more than $600 a day, roughly $12,600 over 21 working days: about what the most generous consumer plan allows at its ceiling. Its 90th-percentile researchers use $7,000 a day, roughly $147,000 a month, some ten times that ceiling. Subscribers hit limits; the people building the models, it appears, do not. Both figures are valued at list prices that OpenAI has since halved, and it says the cut reflects cheaper serving, not only lower prices.

Roetzer closed the June segment on the question that matters now: "By this fall when we're all in budgeting for 2027", he asked, "how do you even begin to do this?" In September, answering a listener on the same show who asked whether everyone is "still finding out when the invoice arrives", he was no further on: the best approach he had heard of was "setting a cap each month", and "I don't know anybody who's actually solved this yet" (episode 236, at 34:11).

What this means for OSASI

For an appraisal firm setting a 2027 budget, the practical questions are the same ones the show could not answer, asked of a smaller business. Which staff are the heavy users, the ones a per-seat increase or a move to usage billing would reach first? What would a report's AI work cost at list price rather than at the subscription price? And how much of the work needs a frontier model at all?

The answer OSASI favours is the one set out in the previous brief: reproducible code tools that rely less on generative AI, with advanced models reserved for the edge cases and the nuanced analysis. A tool that runs the same way every time costs the same to run next year. Roetzer points the same way twice. In June he wonders aloud about an open-source model of frontier quality "in six months". In September his co-host Cathy McPhillips puts her own question: are businesses built on AI platforms repeating the mistake of building on "rented land", a lesson she credits to Joe Pulizzi from her years at the Content Marketing Institute? Roetzer answers that "if you build on a single platform, you're sort of a prisoner to what happens to that platform", and that this is "probably part of the reason why some organizations are now choosing to take open weight models and train their own models", where "the model lives on internal servers." His practical advice is to document every agent's instructions and knowledge base "so we can go back and audit it", which is the workfile discipline appraisers already keep. Owned capacity turns an unpredictable bill into a budget line a firm can predict, and that is part of what the sandbox open-weights AI expression of interest is for.

Send a thought →

A like is counted once per reader per day and stores nothing about you. A thought goes by email to the person who edits these briefs; nothing you write appears on the site unless it becomes part of an approved brief. Share opens the service you pick; this site adds no trackers.

Prefer a feed reader, Slack or Teams? Follow by RSS.