Like in Your Favorite Supermarket Shelf: The Quiet Arrival of AI Shrinkflation

By The GenerIA Team  |  May 21, 2026  |  10 m.

Share this article:

After grocery aisles, shrinkflation has officially hit the frontier of AI. Some tech giants are quietly trimming the computing power behind your prompts while keeping prices exactly the same...

Consumers have long grown familiar with shrinkflation: the chocolate bar that loses a few grams, the detergent bottle whose shape changes just enough to disguise a missing fluid ounce, the cereal box that retains its height but slims at the base. The price stays the same or sometimes rises but, in every case, the quantity delivered quietly contracts. What was once a marketing footnote has become a defining feature of consumer economics.

A strikingly similar pattern is now surfacing in the most unexpected aisle of the digital economy: frontier AI models. Users are paying the same subscription, invoking the same brand, calling the same version number, yet receiving measurably less of what they purchased. The spring 2026 controversy surrounding a major flagship model has turned this quiet shift into an industry-wide conversation.

They call it "Adaptive thinking"...

Over the past several weeks, professional users of one of the industry's leading frontier models have reported a consistent and unsettling experience. Responses feel shallower. Instruction-following weakens. Long reasoning chains contract into hasty conclusions. Initially dismissed as anecdotal frustration or confirmation bias, the complaints began to harden into evidence when a senior AI director at a major semiconductor company published an analysis of nearly 7,000 agentic coding sessions and more than 230,000 tool calls, conducted on a stable internal workload between January and March 2026. The telemetry showed a roughly 67% drop in reasoning depth by late February, a collapse in the number of files read before code edits (from 6.6 to 2.0) and a sharp rise in premature stop-hooks.

Crucially, the model itself had not been retrained or replaced. The provider's engineering leadership acknowledged two deliberate product changes: the introduction of an "adaptive thinking" mechanism that lets the model self-adjust its reasoning budget, and a reduction of the default "effort level" from high to medium. The underlying weights were unchanged. What changed was how much computation the model was allowed to spend on each request by default. No proactive announcement was pushed to end-users. The change was disclosed in the narrow technical sense, through a changelog line and a dialog box. For the average subscriber, it was invisible.

Same models, reduced portions

Shrinkflation in the food industry operates on a simple principle: preserve the visible attributes of the product (brand, packaging, price point) while silently reducing the unseen dimension that actually determines value (weight, volume, active ingredient). The analogy maps onto large language models with disquieting precision. The visible attributes of an AI product are its name, its version number and its monthly subscription cost. The unseen dimension is the quantity of computation actually spent on each request: reasoning tokens consumed, depth of context reread, number of tool calls allowed, length of the internal chain of thought. Reduce any of these by default and the perceived product degrades without a single line of the user-facing description changing. The subscription tier is identical. The invoice is identical. Only the substance contracts.

It is important to be precise about what did and did not change in this episode. The model itself, its weights, its training, its fundamental capability, was not made weaker. Power users quickly discovered that undocumented commands and configuration flags (a maximum-effort flag, an environment variable to disable adaptive thinking, a carefully worded custom instruction asking for thorough reasoning) restored the earlier behavior almost entirely. If the model had truly been degraded, those switches would not work. The capability is intact; the default allocation of compute per query is what was trimmed. This is not a dumber model. It is the same model, served on a reduced portion. That is exactly the structural signature of shrinkflation, and it is the reason the food-industry analogy, sometimes used loosely in online discussions, fits the situation more tightly than the inflammatory language of "a weaker AI" would suggest.

Something feels off

What makes this dynamic particularly acute in AI is that the "weight" of the product is invisible even to experts. A shopper can weigh a chocolate bar on a kitchen scale. An AI user has no equivalent instrument because inference compute is internal to the provider's infrastructure and is not exposed in any standardized way on the user side. Detecting the change therefore requires either privileged telemetry access or statistically rigorous benchmarking across hundreds of sessions, both of which are out of reach for ordinary users. The result is a vague but persistent sense that "something feels off", precisely the sensation that food shrinkflation is engineered to produce.

The economic incentive structure is also familiar. Inference compute is expensive, GPU supply is constrained and demand for agentic AI is accelerating faster than data centers can be built. For a hyperscale provider facing capacity pressure, quietly trimming the default reasoning budget across tens of millions of sessions yields enormous cost savings without requiring any visible change to the commercial offer. It is the textbook move of a mature consumer-goods category under margin pressure, now replayed in a sector that still markets itself as the frontier of innovation.

Is this market maturity?

This is not the sign of a bad actor; it is the predictable outcome of a market structure in which the provider controls every variable that determines product value while the customer has no contractual or technical visibility into those variables. The providers involved have generally acted within the letter of their disclosure obligations. A configuration change was published. A dialog box appeared. A changelog entry was written. But disclosure buried in release notes is not the same as communication to the people whose workflows depend on stable behavior.

Community-sourced workarounds, maximum-effort commands, environment variables and custom instructions such as "treat every request as complex, reason thoroughly, never optimize for brevity at the expense of quality", have proliferated on forums. These folk remedies are revealing in two ways. First, they confirm that the underlying capability is still present in the model; only the default allocation has changed. Second, they place the burden of reconstituting the originally marketed product on the user, who must now learn an undocumented incantation to receive what the subscription nominally provides. This inversion of the implicit contract is what makes the situation a trust issue rather than a mere pricing issue.

The problem is not that a provider tuned its defaults. Tuning is legitimate, and cost-optimizing inference for the average user is a defensible engineering choice. The problem is that the tuning occurred below the threshold of user-visible communication, on a product whose inner workings are contractually opaque to begin with, and on subscriptions that had been sold partly on the basis of the previous behavior.

The case for proprietary AI

The structural answer is not to demand that hyperscalers behave more transparently, an outcome their business model does not strongly incentivize, but to reconsider the architecture of professional AI deployment. An AI system that an organization owns, hosts, tunes and observes is, by construction, immune to silent shrinkflation. There is no remote provider to quietly reduce the effort budget, because the effort budget is set locally and logged locally. There is no ambiguous default, because defaults are owned by the organization. There is no invisible drift, because telemetry is continuous and inspectable. This is the operational premise of bespoke, sovereign AI: replace an opaque service contract with a transparent technical stack whose behavior is fully accountable to its owner.

GenerIA has consistently argued, in prior essays on frugality, observability and data lifecycle management, that scale is not synonymous with value, and that the industrial future of AI in regulated and professional environments lies in carefully engineered, domain-specific systems rather than in ever-larger generalist models. The shrinkflation episode adds a new dimension to this argument. Beyond the already-documented issues of environmental cost, data exfiltration risk and model collapse, organizations now face a form of “product risk” that is uniquely native to hyperscale AI. The risk that the service delivering their business logic today is not configured the same way it was three weeks ago, with no recourse and no meaningful notice. Frugal, sovereign systems, built on curated data, deployed under organizational control and continuously observable end-to-end, do not expose their owners to this risk. Their behavior is a function of deliberate engineering choices, not of a distant provider's quarterly margin calculus or capacity-planning cycle.

Conclusion

Shrinkflation thrives wherever the product's true dimension is invisible to the buyer. In groceries it took regulators and consumer associations decades to impose unit pricing, standardized labels and comparative transparency. AI is at the beginning of that same curve, and the first large-scale episode has just unfolded in public view.

The lesson for professional users is not that one provider behaved badly, it did not, in any conspiratorial sense, but that any model consumed as an opaque service is structurally exposed to this category of drift. The engine may be unchanged while the fuel ration is cut, and from the outside the two are indistinguishable. Intelligence that an organization can measure, audit and control is the only intelligence that can be trusted over time. The alternative, paying a premium subscription for a product whose portion size is redefined quarterly by someone else, is a bargain whose terms will continue to tilt against the buyer.

References

(VentureBeat) Is Anthropic 'nerfing' Claude? Users increasingly report performance degradation as leaders push back