All insights

Weekly18 September 20261 min read

This week in AI: new model numbers, same question

OpenAI has started rolling out its GPT-6 models and Google has new real-time voice models. A bigger version number is not a reason to switch. Your own tests are.

What happened

OpenAI has begun rolling out its GPT-6 family this month, starting with paying customers. Google released new real-time speech models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, for developers and across its own products, and a new Flash model that it says reasons and codes better at the same speed and cost as the last one.

Our take

Every few weeks a new model arrives with a new number, and every few weeks we get the same question: should we switch?

Our answer is always the same: test it. Keep a set of examples from your own work, such as real questions residents ask, real documents your team reads, and real letters you send, with the answers you would accept. When a new model arrives, run the set and compare accuracy, cost and speed with what you use now. If it is better on your work, switch. If not, wait.

This sounds slow, but it is quicker than the alternative. Without your own tests you are relying on benchmarks that may not reflect your tasks, or on impressions from a few tries.

The voice models are worth watching separately. Real-time speech that can reason before it replies could make phone-based services much better, as long as people can always reach a human when they need one.

Keep reading

Related insights