The sentence that sums it all up
In January 2025, on Microsoft's official forum, a disappointed user wrote about Microsoft Designer: "It's exactly like dealers getting people addicted to drugs." He was describing a phenomenon that now affects every major AI service: unlimited generations at launch, followed by drastic limits once users have become dependent.
This strategy has a name: bait and switch (you lure people in with an enticing offer, then replace it with a less favourable one once the customer is committed). It is nothing new in the software world, but generative AI has pushed it to an unprecedented level. Here is how it works, across five concrete cases from 2024-2026.
Case 1: Microsoft Designer, the bait-and-switch manual
Microsoft Designer (Microsoft's visual creation tool, the in-house equivalent of Canva) is the textbook case. Launched in 2024 with unlimited, free generations, the service attracted millions of small businesses and independent creators. Many built entire workflows around the tool: creating Instagram visuals, business cards, flyers. The product was used daily by some companies.
Then in January 2025, a window appeared: "You have AI credits!". 15 boosts per day. Then 60 generations per month if you were a Microsoft 365 subscriber (Microsoft's paid office suite, which includes Word, Excel, PowerPoint and Outlook). For users used to generating 10 images in a row to find the right one, this was unworkable. And Microsoft offered no "Designer only" plan: you had to subscribe to the full Microsoft 365 to get more credits.
As we documented in our article on the great AI failures, Microsoft eventually killed off Designer in February 2026, after vainly trying to monetise the dependency it had created. The bait and switch worked in the short term, and failed in the long term.
Bait and switch exploits a documented cognitive bias: loss aversion. Once you have tasted a service without constraints, losing access to it is psychologically more painful than never having had it. Users would rather pay to keep what they got for free than look for an alternative. Every tech company knows this. And they apply it methodically to AI.
Case 2: ChatGPT and the silent degradation
OpenAI applies bait and switch in a more subtle way. ChatGPT Plus launched in February 2023 at 20 dollars a month with "priority" access to GPT-4 (the fourth generation of the GPT language model, OpenAI's conversational AI). At first, users had a high quota. Over the months, quotas were gradually reduced: 50 messages per 3 hours, then 40, then 25 during certain periods of high demand. With no public announcement. No press release.
More subtly still, OpenAI introduced GPT-4 Turbo in November 2023, then GPT-4o in 2024, then GPT-5 in 2025. At each transition, Plus users were switched to the new model, sometimes without being able to go back to the old one. The problem: "new" models are systematically faster and cheaper to serve, but not always better for every task. Several power users (heavy, professional users) complained that GPT-4o gave less in-depth answers than GPT-4. OpenAI eventually admitted the problem in 2024.
The bait and switch here is not in the price (still 20 dollars), but in the quality of the service delivered for that price. The service sold in 2023 is not the same as in 2026.
Case 3: Midjourney and the end of the unlimited plan
Midjourney (one of the main AI image generation services, accessible via Discord) had a "Pro" plan at 60 dollars a month that included an unlimited "Relax Mode". You could generate as many images as you wanted, provided you accepted a longer queue. For creators and freelance illustrators, it was the deal of the century.
In 2024-2025, Midjourney gradually introduced limitations on Relax Mode, then turned it into "Yearly Relax Hours". True unlimited disappeared. To maintain their productivity, many users had to move up to the "Mega" plan at 120 dollars a month, or even pay for additional "fast hours". The cost increase for the same volume of work often exceeded 100%.
Case 4: Claude Pro and the hidden "usage caps"
Anthropic, OpenAI's direct competitor founded in 2021 by Dario and Daniela Amodei (former OpenAI researchers), is no exception. Claude Pro is sold at 20 dollars a month with "extended" access to Claude (their conversational AI, the equivalent of ChatGPT). But the word "extended" hides a shifting reality. Depending on the period, Pro users can send 45 messages every 5 hours, or 30, or sometimes fewer. The quota varies with server load, without notice.
The problem here is different from Microsoft: Anthropic is not doing a bait and switch on price. The company simply changes the value delivered by the subscription, without touching the rate. It is a gentler version of the bait and switch, but the result is similar: the user pays more for the same service over time. As we analysed in our article on OpenAI's lock-in, these mechanics are structurally tied to the platform position.
Case 5: Higgsfield and the transparency of the credit-based model
For comparison, let us look at a case that does the opposite. Higgsfield (an AI image and video generation platform, which aggregates more than 30 different models under a single interface), whose pricing system we described in our article on Higgsfield MCP, uses a model of transparent credits. You buy 1,000 credits per month for 49 dollars (Plus plan), and each generation consumes a number of credits known in advance (0.5 credit for a Soul image, 12 for GPT Image 2 in 4K, 135 for a Seedance 2.0 video).
This model is not perfect either (heavy users can quickly exhaust their credits), but it has one major advantage: the user knows exactly what they are buying. No surprises, no hidden degradation, no sudden limitations. This is probably the direction all AI services will eventually take, because it is the only one that stands the test of time and trust.
The six rules to avoid getting trapped
If you adopt a new AI service in 2026, here are six practical rules to avoid falling into the trap of unlimited freemium.
Rule 1: never build a critical workflow on a free or "unlimited" plan. If a company offers you something at a price that seems too good to be true, it is probably temporary. Build on stable paid plans, even if it costs more at the outset.
Rule 2: prefer credit-based models to message-based models. A transparent credit system (like Higgsfield) is more predictable than a system that talks about "priority messages" or "extended access" without specifying quotas.
Rule 3: read the terms of use, especially the clauses that allow the company to modify the service unilaterally. All tech companies reserve this right, but how aggressively they use it varies from one firm to another.
Rule 4: always keep a credible alternative identified. If you use ChatGPT, identify which other service you would use instead tomorrow. If you use Cursor (an AI code editor), know which IDE (Integrated Development Environment) you would take as an alternative. The mere fact of knowing that alternative reduces your psychological dependence.
Rule 5: track your real monthly costs over 6 to 12 months. Most cost increases happen in small increments. If you look at your bill month after month, you will see that what cost 30 dollars in January costs 65 dollars in December, without anything in the service having fundamentally changed.
Rule 6: be ready to leave. The best defence against bait and switch is the ability to leave. Export your data regularly. Document your workflows. Periodically test alternatives. A company that knows its users can leave does not allow itself to hold them to ransom.
Will the market discipline itself?
The real question is whether bait and switch practices will continue to spread, or whether competition will eventually impose more transparency. Several signals point in opposite directions.
On one hand, the consolidation of the market around a handful of giants (OpenAI, Anthropic, Google, Microsoft) makes opaque practices easier. When there are few credible alternatives, users suffer. This is exactly what we described in our article on lock-in.
On the other, the emergence of powerful open source models (models whose code and parameters are published freely, accessible to all) (Chinese DeepSeek V4, French Mistral Large, Meta's Llama) creates competition from below that forces commercial players to stay competitive. A user tired of OpenAI's practices can now run DeepSeek V4 on their own servers for a few dollars. That option did not exist two years ago.
The market will probably polarise: on one side, premium services with stable, transparent terms, more expensive but reliable. On the other, free or very cheap services, with terms that shift and friction. Each user must choose their camp.
To go further on these issues, you can read our analysis of faked AI demos, which shows how these same players sell their products in advance, and our article on the AI bubble, which places these pricing practices back within the sector's broader economic pressure. It is better to choose consciously than to accept passively.