The models are starting to specialise
OpenAI released a coding-specific version of GPT-5 on Monday. Even if you never write software, the direction matters to you.
On 15 September 2025 OpenAI released GPT-5-Codex, a version of GPT-5 tuned specifically for software engineering work. It is the same base model, adapted for one job.
If your company does not write software, the release itself is irrelevant. The pattern is not.
What specialisation signals
For three years the industry pursued one direction: a single model that is better at everything. This is the first strong sign of the other direction — the same model adapted for a domain, because general capability has got good enough that fit now beats raw strength.
Coding is first because it is the largest and most measurable commercial market. It will not be last.
What arrives next, and what it means for a buyer
Expect versions tuned for law, for medicine, for customer service, for accounting. Some will be genuinely better at their domain. Some will be the same model with a different name and a higher price, and telling those apart will become a real purchasing skill.
Two questions separate them. What is it tuned on, and what changed? A real specialisation has an answer — a described dataset, a measured difference. A repackaging has a brochure.
And: what does it do worse? Specialisation is a trade. A model tuned for one domain is usually weaker outside it. A supplier who will not name the trade-off has not made one.
The Swiss angle
Be particularly careful with anything sold as tuned for Swiss law, Swiss accounting or Swiss administrative practice. The market is small, the training data is scarce, and the gap between a model that has read Swiss material and a model that has been told it is Swiss is invisible from the outside.
Ask for the two answers above before paying a premium for a national label.