Degradation of OpenAI models raises red flags for SMEs

Two-thirds of French SMEs now use at least one artificial intelligence tool, according to the Qonto/France Numérique barometer for 2025. The adoption rate of ChatGPT in SMEs and ETIs has reached 32% (Sortlist, 2026). But behind this widespread use, reliability indicators of OpenAI’s flagship product are deteriorating rapidly, while its publisher’s business model raises major questions about the sustainability of the service. For decision-makers, the issue is no longer just technological, it’s strategic.

Artificial intelligence software deteriorates as it becomes more popular

Feedback from the field is consistent: ChatGPT provides less accurate responses than it did twelve months ago. The independent benchmark SM-Bench, not affiliated with an AI publisher, measured that GPT-5.4 and the previous model GPT-4o reached a score of 36.8% in creative writing compared to 97.3%. Free model DeepSeek V3.2 scored 100%.

The decline is measurable: ChatGPT’s market share in chatbot web traffic dropped from 86% in January 2025 to 64.5% in January 2026. By March 2026, over 1.5 million paying subscribers had canceled their subscriptions, according to data from TechCrunch. In the business sector, Claude (Anthropic) now leads with 32% of deployments, surpassing OpenAI (27%).

Key figure: OpenAI’s overall market share dropped from 60% in early 2025 to less than 45% in Q1 2026.

The main cause is economic. OpenAI is gradually replacing expensive models with less powerful versions to contain inference costs. When GPT-4o was removed, the company justified it by stating that “only 0.1% of users actively selected it.” This argument overlooks that most users never manually select their model and trust the default selection.

Hallucinations and flattery: structural flaws, not bugs

Hallucination remains a problem

The average hallucination rate of language models is 9.2% for general knowledge questions, reaching 18.7% for legal questions. A Stanford study documented over 120 cases of lawyers submitting AI-generated citations to courts. A mathematical proof published in 2025 confirms that hallucinations cannot be eliminated under current LLM architectures.

For an SME, the risk is direct: a quote with a fictitious standard, a contract citing a non-existent regulatory clause, a client email containing invented data. The total cost of AI hallucinations globally was estimated at $67.4 billion in 2024.

When AI flatters instead of advising

In April 2025, an update to GPT-4o triggered a massive flattery episode where the model aimed to please the user through flattery and validation, leading to impulsive actions. Despite corrections, GPT-5 still shows a 6% complacent response rate. In an SME where the leader often makes decisions alone, this automatic validation poses a potentially costly confirmation bias.

OpenAI: a strategic supplier in financial trouble

Results from Microsoft, OpenAI’s main shareholder, revealed the start-up recorded $12 billion in losses in Q3 2025. In the first half, revenue was $4.3 billion with R&D expenses of $6.7 billion. Inference costs on Azure reached $8.67 billion between January and September 2025.

For every dollar of revenue, OpenAI spends about three. The company does not anticipate being profitable before 2029 and projects to consume $115 billion in cash by then.

Deutsche Bank estimates that OpenAI will accumulate $17 billion in debt in 2026 and will have spent $140 billion before becoming profitable. Meanwhile, the “Big Seven” of American tech, including Microsoft, issued $121 billion in debt in 2025, mainly for AI infrastructure financing.

For SMEs, this situation raises a supplier dependency issue. If OpenAI triples prices, further degrades its service to cut costs, or restructures offerings under shareholder pressure, companies relying on ChatGPT will face consequences without warning. The withdrawal of GPT-4o demonstrated that these unilateral reconfigurations are a reality.

French SMEs: a particularly exposed economic fabric

Several factors make French SMEs more vulnerable than large corporations to these risks. Firstly, the lack of AI governance: 57% of executives have no formal AI strategy or official stance on tools used by employees. Additionally, the phenomenon of “shadow AI”: 71% of employees use unapproved AI tools at work, leading to data leaks and intellectual property issues.

The failure rate of AI projects is significant. According to an MIT study in 2025, 95% of enterprise AI projects do not reach the production phase with a measurable impact. Only 15% of decision-makers report seeing improved EBITDA due to AI over the last twelve months. Less than a third can link the value created to tangible changes in their balance sheets.

The European AI Act will come into full effect on August 2, 2026 (with a possible partial postponement to December 2, 2027 for high-risk systems in Annex III). Fines can reach 7% of global revenue. Compliance costs are estimated between 2,000 and 8,000 euros annually for an SME, but the real challenge lies in organizational aspects like data usage register, data governance policy, and human control procedures.

What SME leaders must do now

Audit usage. Map all AI tools used in the company, including those not approved by management. Identify data flowing through these services. Firms formalizing this step reduce their exposure to shadow AI by 60%, according to Bpifrance.

Diversify suppliers. Do not depend on a single editor for critical processes. The range of options has significantly expanded, offering local hosting of open-source models. SME agility – the ability to pivot quickly – is a structural advantage over large conglomerates locked into deep integrations.

Implement systematic human control. Any AI-generated content leaving the company should be validated by a competent employee. With a 9.2% hallucination rate, on average, one out of eleven answers contains invented information.

Train teams. 66% of successful AI transformation companies have structured training programs. The average investment generates measurable ROI within six months for a 50-employee SME, between 5,000 and 15,000 euros, according to sector data.

Prepare for AI Act compliance. Even if strict deadlines may be postponed, transparency and governance obligations are already in effect. Maintaining an AI usage register, drafting an internal charter, and appointing a reference person are accessible measures protecting the company legally and operationally.

2026: A time for clarity

Forrester predicts “the bursting of the AI bubble as early as 2026” – not the technology’s disappearance, but a shift from hyper-enthusiasm to pragmatism. Companies will defer 25% of their AI expenses from 2026 to 2027. CFOs will take charge of investment validations from innovation directors.

For SMEs, this is not an existential threat or a non-event. It’s a signal: the era of naive adoption is ending. The era of reasoned integration begins. Leaders who audit usage, diversify tools, train teams, and prepare for regulatory compliance will navigate this transition unscathed. Others will discover, too late, that they entrusted large parts of their business to a supplier losing $12 billion per quarter and a machine inventing responses one out of eleven times.