GPT-5 Is Here, and My Clients Are... Underwhelmed
Last week, three different enterprise clients told me the same thing about GPT-5: "Is this it?" Their companies are spending six figures monthly on OpenAI APIs, running thousands of customer interactions daily through agentic interfaces. And they're not happy.
A fintech company I work with built their entire customer service architecture around predictable model behaviour. They had GPT-4 handling complex regulatory questions, GPT-3.5 for basic FAQs, and could switch between them based on load and cost constraints. Now they're getting inconsistent responses to identical prompts, and they suspect GPT-5 consists of multiple models pretending to be one. Another client in insurance noticed that their accuracy metrics actually dropped after the GPT-5 migration enough to matter when processing insurance claims. When they reverted to GPT-4 via Azure, performance improved again.
The only person who was excited was a product manager who loves the new chat interface. He loves switching between "deep research mode" and casual conversation. That's UI design, not model intelligence. When your biggest win involves better buttons rather than better performance, something fundamental has shifted.
This matches a pattern building since early 2023. Back then, MIT Technology Review warned we were approaching a "data wall." The math was simple: we're running out of quality human text to train on. Estimates suggested public text would be exhausted by 2026-2032. But that assumes linear consumption. What actually happened was exponential. Major labs started hoovering up every byte of readable text, from Reddit posts to scientific papers. By 2024, they were already scraping the bottom of the internet barrel.
The technical reality becomes clear when you understand power laws. To get each linear improvement in performance, you need exponential increases in data and compute. The first 10% improvement might cost $1 million. The next 10% costs $10 million. The one after that? $100 million. And we've reached the point where the next meaningful jump would require more text than humans have ever written.
By November 2024, the cracks started showing. Reports leaked that OpenAI's next model, Orion, delivered only modest improvements over GPT-4. The company that once promised exponential progress was struggling to maintain linear gains.
Apple's machine learning team published "The Illusion of Thinking" in June 2025, demonstrating that models couldn't solve problems beyond a certain complexity threshold. While critics argued the tests were flawed, that models were hitting token limits rather than reasoning barriers. But that defense reveals its own problem. If our most advanced AI can't list the steps to solve a Tower of Hanoi puzzle because it runs out of tokens, how can it handle real-world complexity?
The industry response has been telling. Instead of claiming continued scaling success, everyone pivoted to "test-time compute." Translation: we can't make models smarter through training, so we'll let them think longer during inference. OpenAI's o1 and o3 models represent this shift. They use reinforcement learning at runtime to explore multiple reasoning paths. The marketing claims 10x to 100x performance improvements compared to pure pre-training scaling.
But my clients see through this. One CTO told me: "We're paying more compute costs to make the model pretend to think harder about problems it fundamentally doesn't understand. It's like making a calculator show its work. Impressive animation, same limitations."
Even Sam Altman seems to be managing expectations. The same person who testified to Congress about existential AI risks now calls AGI a "meaningless term" and openly admits we're in a bubble. That's a complete reversal of the narrative he's been selling. But this might be the best thing that's happened to enterprise AI!
For the past two years, I've watched companies paralyzed by FOMO. Why invest in building robust systems today when next month's model might make everything obsolete? Why train your team on current tools when the ground keeps shifting? This endless wait for the "next big thing" has prevented real value creation.
Now, with capabilities plateauing, we can finally stop chasing mirages and start engineering solutions. My clients who've invested in proper #LangGraph architectures and systematic #RAG implementations are pulling ahead, not because they have access to better models, but because they've learned to maximize what we already have.
The companies winning right now aren't the ones waiting for GPT-6 to solve their problems. They're the ones who've accepted that current AI - with all its limitations - is incredibly powerful when properly implemented. They're building systematic approaches to prompt engineering, investing in observability, and treating AI integration as an engineering discipline rather than magic.
The technical evolution tells the real story. Companies shifted from pure scaling to what researchers call "portfolio approaches." Instead of throwing more data at pre-training, they're optimizing across multiple dimensions. Synthetic data generation, where models create their own training data. Architectural innovations like State Space Models that process sequences more efficiently than transformers. Hybrid search combining dense and sparse retrieval. Tool use that grounds language models in real computation.
Each approach addresses specific limitations. Can't find more human text? Generate it. Transformers too slow? Build something better. Models hallucinating? Connect them to calculators and databases. None of these solve the fundamental reasoning problem, but together they create systems that actually work.
The slowdown forces us to confront an uncomfortable truth: we've been conflating linguistic fluency with intelligence. These models excel at sounding smart. They've memorized the patterns of human reasoning without understanding the underlying logic. When you ask GPT-5 to solve a complex problem, it performs an impressive puppet show of thinking. Sometimes that's enough. Often it isn't.
The slowdown is a reality check that's long overdue. We're moving from the "AI will solve everything" phase to the "let's solve specific problems well" phase. The companies that thrive in 2025 will be the ones who stopped waiting for AI to get "good enough" and started getting good at AI. The ones who understand that a 10x improvement in implementation beats a 10% improvement in model capabilities every single time. I'm relieved. It's time to stop talking about what AI might do tomorrow and start delivering what it can do today.
At Chelsea AI Ventures, we are implementing bespoke AI for companies. We’ve also developed an AI training programs: Our AI Literacy course helps teams understand what AI can actually do today - not the hype, not the fear, just practical capabilities and limitations. For leadership, our Executive AI Training focuses on strategic implementation: how to identify real opportunities, avoid costly mistakes, and build competitive advantages with current technology.
My book, the second edition of Generative AI with LangChain, covers LangGraph architectures, multi-agent systems, and advanced RAG pipelines that work regardless of whether you're using GPT-4, GPT-5, or whatever comes next. Companies who want to succeed focus on building reliable, testable, observable systems.


