The Market Caught Up. A Year After GPT-5
How Turkey’s university entrance exam confirmed a prediction I made a year ago. The next AI race is no longer about intelligence. It is about delivering the most value…
Exactly one year ago, I published an article titled: “The GPT-5 Curveball Didn’t Top the IQ Chart. That Was Never the Point.” In that article, I argued something that sounded controversial at the time.
The future of AI would not be decided by who won benchmark leaderboards. It would be decided by who solved real problems most effectively.
Subscribe to stay updated whenever I publish
This week, I came across an excellent benchmark by Alican Kiraz, who evaluated 27 frontier language models using the actual 2026 Turkish university entrance exam (TYT). His analysis compared models not only by exam score, but also by inference cost and token consumption. His results reinforce the exact argument I made a year ago.
👏👏 Credit to Alican Kiraz for putting together one of the most practical AI benchmarks I have seen recently.
🎓 For readers outside Türkiye, the TYT is part of the national university entrance examination taken by millions of students every year. Together with the Advanced Proficiency Test (AYT), it determines admission to the country’s most selective universities, including Boğaziçi University, Middle East Technical University (METU/ODTÜ), Istanbul Technical University (İTÜ), Koç University and Yıldız Technical University (YTU). Every additional correct answer can change a student’s future.
I know this system well. I earned my bachelor’s degree in Mechanical Engineering from YTÜ and later completed my master’s degree in Industrial Engineering at METU/ODTÜ. Both required highly competitive entrance scores, and these universities consistently admit students from the very top of the national rankings.
That is why Alican’s benchmark is particularly interesting. It is not measuring models on abstract puzzles or synthetic datasets. It is comparing them on the same examination that determines which students earn a place at some of Türkiye’s most competitive universities.
The Leaderboard Isn’t the Story 📈
At first glance, the results look familiar. GPT-5.6 Sol, Grok 4.5, Fugu Ultra and Gemini 3.6 Flash all finished at the top.
The difference? Almost none…
The top four models answered virtually the same number of questions correctly. If someone only looked at the leaderboard, they would conclude that these models are essentially tied. But the real story starts when you move one column to the right.
🏁 Every one of the leading frontier models would likely earn admission to one of Türkiye’s most selective universities. The competition is no longer about proving intelligence. It is about proving economic value.
Intelligence Is Becoming a Commodity 💰
One model required around $35 to complete the benchmark. Another achieved the same score for less than $2. Gemma 4 31B, an open-source model, scored only a few questions behind the leaders while costing just a fraction of the price.
That is a much more interesting observation than who ranked first. Because users rarely ask:
Which model has the highest benchmark score?
Instead they ask:
- Which model gives me the best result?
- Which one is fast enough?
- Which one is reliable?
- Which one fits inside my infrastructure?
- Which one makes economic sense?
Those questions create businesses. Benchmark rankings do not.
The Benchmark Gap Is Shrinking 📉
A year ago, many discussions focused on a single question.
“Which model is smarter?”
Today, that question matters much less. The gap between frontier models has become surprisingly small across many knowledge-intensive tasks.
Get Altan "Atabarezz" Atabarut’s stories in your inbox
Join Medium for free to get updates from this writer.
Once multiple models reach a similar capability level, intelligence stops being the primary differentiator. Everything else starts to matter more.
- Cost
- Latency
- Reliability
- Context handling
- Tool use
- Product integration
- Developer experience
- Enterprise deployment
This is exactly what happened in cloud computing. Nobody buys the fastest CPU anymore. They buy the infrastructure that delivers the best business outcome. AI is following the same path…
A Good Benchmark. But It Can Go Further. 🔬
I actually like the idea of using a real university entrance exam. It measures reading comprehension, mathematical reasoning and general knowledge under realistic conditions.
That makes it far more meaningful than many synthetic benchmark datasets. But it still measures only one dimension.
Real-world AI should also be evaluated on tasks like:
- Writing production-quality code
- Completing multi-step agent workflows
- Using external tools correctly
- Reviewing contracts
- Creating investor presentations
- Long-context document analysis
- Reliability across repeated runs
- Cost per successful task
These are the problems companies actually pay for.
The Most Interesting Number Wasn’t the Score 💡
The biggest takeaway from Alican’s benchmark is not who finished first. It is how close everyone has become. When four different frontier models produce nearly identical academic performance, the competitive advantage moves somewhere else.
Execution. Product. Economics.
Exactly what I argued one year ago. The market didn’t suddenly become less interested in intelligence. Intelligence simply stopped being scarce.
The interesting question is no longer whether AI can outperform excellent students. We already know it can.
The interesting question is why four different frontier models, all capable of earning admission to Türkiye’s most selective universities, are separated by only a handful of questions while their operating costs differ by an order of magnitude.
The Next Competitive Frontier 🚀
We are entering a new phase of AI. The winners will not necessarily be the companies with the smartest model.
They will be the companies that combine capable models with superior products, better workflows, stronger agent systems and lower operating costs.
Model intelligence is converging. Execution economic value is becoming the differentiator. That may be the biggest shift in AI over the past year.
All the best
Altan
Subscribe to stay updated whenever I publish
References:
- My previous article:
“The GPT-5 Curveball Didn’t Top the IQ Chart. That Was Never the Point.”
https://atabarezz.com/the-gpt-5-curveball-didnt-top-the-iq-chart-but-a9f4e9315e8b?sk=1b596d349dd50b629e1b82f8df677af6 - Alican Kiraz’s benchmark and analysis:
https://x.com/AlicanKiraz0/status/2082014156849631258
