For three years, one model has done everything: read a document, search the web, rank the results, decide what to do next, then write the answer. Jaya Gupta, in a long post on X, argues that era is ending, and that the split will reshape who makes money in AI.
Her case starts with what got lost in that convenience. Extracting text, ranking options, checking a browser action, and solving a hard reasoning problem all ran through the same expensive model, even though they cost wildly different amounts to do well. Gupta calls the coming shift the “Great Unbundling of Intelligence”: instead of sending every task to the priciest model available, software will start routing each small decision to whichever system handles it cheaply enough, and save the expensive model for the questions nothing else can answer.
Agents are what forces the change, in her telling. A single task an agent handles can involve hundreds or thousands of small decisions (should I stop here, is this the right tool, did that click work), and at that volume, shaving fractions of a cent off each step adds up to real money. Gupta points to a specific example: a tool called Jev, built by a company called TypeSafe, which she says answers narrow yes/no or multiple-choice questions directly instead of generating a full sentence and then parsing it back into a decision, a step she labels the “decoder tax.” That is her characterization of the product; TypeSafe’s own claims about it are unverified here.
The financial stakes she draws out are specific to one company. Gupta writes that Anthropic “reportedly” runs gross margins above 80 percent, a figure she attributes to reporting rather than to Anthropic itself, and argues that margin depends on cheap tasks like classifying an email as urgent being billed at the same rate as genuinely hard reasoning. If routing tools get good enough to peel that routine work away and send it to smaller, open-weight models, the expensive model is left holding only the tasks nobody else wants: the ambiguous, long-horizon problems that may also demand more computing time per answer.
That is a real bet against a business model, not just a technical prediction, and it is the strongest part of her argument: a shrinking pool of harder requests could keep frontier models profitable per query while cutting the number of queries that reach them at all, a combination Gupta does not attempt to model with numbers.
Her more concrete claim is about what gets built once cheap judgment exists. She argues that companies currently sample a small fraction of their agent’s actions or customer conversations for review because checking everything is too costly. Once a decision costs a fraction of a cent, she says, that changes: a company could check every action an agent takes instead of a sample, moving quality control from occasional to constant.
Gupta does not name her employer or affiliation in the post, and none is stated in the source, so none should be assumed here. Her argument is her own framing of where agent economics are heading, built from a mix of trend observation and one product example, not a data set she publishes for scrutiny.
For any team already running agents in production, the practical test is narrower than her thesis: audit which steps in your agent’s loop are simple classification or ranking calls dressed up as full model prompts, and price out whether a smaller model or a dedicated classifier would do that step for less without a quality drop worth measuring against the savings.
Jaya Gupta, in a post on X on September 20, 2026.
