GPT-5.2 just achieved near-perfect scores on almost all industry benchmarks. In the hours that followed, the forums were filled with opposite testimonies…
Impoverished responses, erratic behavior, noticeable drop in conversational quality; both observations are nevertheless true. A tool that performs well in tests and disappoints in real life is not a tool in transition but a tool that has changed owners.
When the score becomes the official lie
Benchmarks measure what they are asked to measure. In this case, formal reasoning skills, code completion, documentary synthesis under controlled conditions. They do not measure day-to-day reliability, consistency over a long conversation, or suitability for real and changing professional use cases.
However, this is precisely where users observe a degradation. Since the GPT-5.x iteration, the same pattern has been reproduced with each update: triumphant announcement, mass adoption, then collective complaint within 24 to 72 hours. This is not a coincidence. It’s a signal.
GPT-5 was never made for you
GPT-5 was not designed for the general public. It was designed to conquer the enterprise market, a territory in which Anthropic had a significant head start. This positioning is not illegitimate in itself, it is consistent with investor pressure and the maturity of the market. What is less clear is the total lack of transparency towards users who continued to pay for their subscription, train their teams, and structure their workflows around a tool whose parameters changed without notice.
GPT-5.2 was even deployed urgently, according to several sector sources, in the form of an “early checkpoint” sent in the face of competitive pressure from Google. Quality sacrificed on the altar of speed to market. It is no longer a product choice, it is an industrial crisis communication decision. You were the product. You are now the excuse.
When users vote with their bank card
Data from 2026 confirm this unambiguously. ChatGPT’s market share fell from 60% at the start of 2025 to less than 45% in the first quarter of 2026. More than 1.5 million subscribers canceled their subscriptions in March 2026 alone. These are not anecdotal figures: they reflect a breakdown in trust, slow to build, quick to lose.
This massive churn movement does not mean that generative AI is going backwards. It means that users began to do what they should have done from the beginning: evaluate, compare, arbitrate. The maturity of the market is measured precisely at this moment when habit gives way to reasoned choice.
The real vulnerability is not technical, it is organizational
What should worry business management is not that a language model occasionally regresses. Models evolve, adjust, reorient themselves, it’s the very nature of this industry. What should be worrying is having left entire workflows, internal training, decision-making processes based on a tool whose publisher changes the parameters without notice, without documentation of the changes, and without any accountability mechanism.
Choosing an AI tool by default, because it is the first, because it is known, because it avoids the friction of comparison, is to entrust the operational stability of your organization to a product roadmap that you do not control. This is not a technology policy. It is a strategic abdication.
Evaluate, not just adopt
The question is no longer “ChatGPT or something else?”. It is: on what objective criteria does your organization evaluate the AI tools it integrates? Consistency over time, stability between versions, transparency of updates, documentation of behavioral changes, but also the real cost of use, hosting and localization of data, regulatory compliance, reversibility. These criteria are as legitimate as benchmark scores. And they are rarely in the technical sheets.
It is moreover on these criteria, and not on the rankings, that arbitrations are made today. Organizations that have taken the time to test several models on their own use cases rarely come to the same conclusions as the leaderboards: a model can dominate the benchmarks in fixed conditions and prove less reliable on a daily basis than a better balanced model. Use value is built over time and is observed in production, not in a launch table.
