Beyond ChatGPT: How Do You Choose Between AI Models?

OpenAI introduced ChatGPT on 30 November 2022 as a free research preview. More than three years later, on 27 February 2026, it announced that the service had reached 900 million weekly users, up from 800 million in October 2025. The figure suggests a field settled in one company's favour, yet what happened over the same years says otherwise.

From monopoly to crowd

Stanford's 2025 AI Index shows that the performance gap between the first- and tenth-ranked models on Chatbot Arena (where users vote on the answers of competing models) narrowed from 11.9% to 5.4% in a single year. On benchmarks, the gap between the best American and Chinese models shrank on MMLU from 17.5 points at the end of 2023 to 0.3 at the end of 2024, and on HumanEval from 31.6 to 3.7. The 2026 index gives two Arena scores: 1,503 for the top American model and 1,464 for the top Chinese one. That arena measures the preferences of voters in general conversation, not the model's performance on your task.

For background on these models, see the article "How Do Large Language Models (LLMs) Like ChatGPT Work?".

What did DeepSeek show, and what did it not?

On 27 January 2025 Nvidia lost about 589 billion dollars of market value in a single day, the largest one-day loss in US stock-market history according to Maginative. This followed the Chinese company DeepSeek's announcement of R1, an open-weights reasoning model, which its founder claimed matched OpenAI's o1.

A cost of "5.6 million dollars" is often quoted in this connection, but it belongs to a different model, DeepSeek-V3. Its team reported in its technical report that full training took 2.788 million H800 GPU-hours, which it priced at about 5.576 million dollars assuming a rental rate of two dollars per hour, excluding the cost of prior research and ablation experiments. It is therefore the cost of running one particular training run, not the cost of building a laboratory or of the research that came before. The plausible lesson is that efficiency can narrow the gap between leader and follower, not that building advanced models has become cheap.

How do you choose a model?

Nobody can say from a distance which model suits you. Practical questions help:

• Test it on your work. Gather twenty real tasks and compare the models' answers with answers you know; published benchmarks are one thing and your performance is another.
• Look at your data. If it is sensitive, ask where it is processed and whether an open-weights model can run on your own infrastructure.
• Count the cost at scale, not the subscription price alone.
• Test reliability. No model is free of inventing plausible-sounding information, so keep a human reviewer in medicine, law and finance (see the article "Responsible Artificial Intelligence: Four Principles for Governance and Trust").
• Test Arabic and your local context with examples from your own language and regulations (see the article "Arabic Artificial Intelligence: Do We Make or Consume? Sovereignty and Arabic Language Models").
• Look at the provider. Will its programming interface stay stable? What is its privacy policy?

Build so you can swap

Leadership changes, and the distance between the leaders has narrowed to this extent. Whoever builds their work on one model ties their fortunes to one company. The wiser course is to keep your application separate from the model so that you can replace it, to keep a small test set that you rerun whenever a new model appears, and to invest in what lasts: writing clear instructions, reviewing outputs and governing data.

What we have written here is a snapshot from October 2026 of a fast-moving market, so check its figures before you base a decision on them. These tools are a useful blessing, and part of gratitude for a blessing is using it well and reviewing what it produces.