Knowledge cutoffs and model versioning

Short answer: a model is like an extremely well-read colleague who left the school last year and has not read a newspaper since. Everything he remembers stops on his last day, and he will not think to mention that.

The first week of September

Early in September, a Secondary 6 form teacher was preparing a university admissions briefing and asked the AI what was new in this year's joint admissions arrangements.

The answer was beautifully organised: application windows, subject selection strategy, results day arrangements, with a bulleted list of things to watch. She almost printed it for the class. Checking against the Education Bureau website, she found two of the arrangements described the old system, which had since changed.

Her question was blunt: how can something this capable not know about this year?

The colleague who left last year

Picture a senior colleague at your school. Twenty years of teaching, knows the curriculum, excellent memory, writes well. But he resigned on a particular day last year and has had no contact with school matters or the news since.

Ring him today and ask how to write this year's admissions briefing. He will be delighted to help and will produce something very plausible from what he knew before leaving. He will not say that he has been away and does not know the current arrangements, because in his memory what he knew is still correct.

That colleague's last day is the model's knowledge cutoff.

The reason for it was already explained in lesson 6: pre-training runs thousands of GPUs for weeks to months, so data collection has to stop on some date. Nothing after that date is in the parameters.

The cutoff and the hallucination of the previous lesson are twin problems. The model does not know the new arrangements, but it contains a machine that always fills the blank — so it builds a plausible current version from the patterns of the old system.

The actual dates, and how to check them yourself

Cutoffs are not guesswork. Major providers publish them. As of September 2026:

• OpenAI's API model page lists the knowledge cutoff for GPT-6 Astra as 30 April 2026, and for the GPT-5.6 family (Sol, Terra and Luna) as 16 February 2026.
• Anthropic's models overview lists Claude Opus 5's reliable knowledge cutoff as May 2026 and Claude Sonnet 5's as January 2026, while the older Claude Haiku 4.5 sits back at February 2025.

Three things worth noticing.

First, different tiers from the same provider can be months or more than a year apart. Using the newest vendor does not mean using the freshest knowledge.

Second, the cutoff is usually several months before the release date. Training, evaluation and safety testing all take time.

Third, do not ask the model for its own cutoff. It may well get it wrong, because that date was written into documentation after training finished. Check the provider's page.

Cutoffs are not the only clock: versions and retirement

Two other things can affect your materials mid-year.

Version updates. A model under the same name may be quietly updated, and after an update the same prompt template can produce different output. The quality usually improves, but the formatting and length often shift. Panel prompt templates deserve a re-test each school year.

Retirement. Providers publish retirement dates per tier. Anthropic's documentation, for instance, gives each current model a "not sooner than" commitment. For a school this means that if a workflow depends deeply on one tier's specific behaviour, that dependency will eventually break.

So one line is worth putting into a procurement specification: can the platform switch providers? A platform that can switch turns a model change into a re-test rather than a redeployment.

An example from the classroom

A STEM panel built a thorough AI planning workflow last year: twelve prompt templates covering lesson plans, worksheets, assessment papers and parent circulars, all kept in a shared drive.

At the start of the new school year three colleagues separately reported the same thing. The same templates were producing a different shape of output — some longer, and a template that used to generate a table now produced a bulleted list. Nobody had edited anything.

The cause was that the platform had changed its default model tier over the summer. Output quality had in fact improved, but the formatting change meant worksheets that used to be print-ready now needed reformatting.

What they did afterwards is worth copying. Each template gained a three-line header: date last tested, the model tier it was tested on, and the three formatting requirements the template depends on. Then, at the August panel planning meeting, they spend half an hour re-running all twelve and fixing whatever has drifted.

That half hour saves far more than it costs. What it really does is convert "the model will change" from an unpleasant surprise into a scheduled routine.

Try it yourself

To see the cutoff with your own eyes, paste this into any AI tool:

Tell me the significant education-related news in Hong Kong last week, listing your sources and their dates.
Then state clearly whether this answer came from what you remember from training or from searching the web just now. If it came from memory, say roughly when your knowledge stops.

If search is not enabled, the model will usually admit it cannot cover last week. If it produces "last week's news" without being able to name a source, that is the hallucination from the previous lesson. The most useful thing about this prompt is that it tells you whether the interface in front of you has search at all.

What this means for your classroom

1. Verify anything about this school year yourself. Examination dates, admissions arrangements, curriculum updates, Education Bureau circulars — the official website is the authority, and the AI is for organising and rewriting.
2. Put school information in the knowledge base rather than waiting for the model to catch up. Teaching schedules change every term; models update roughly once a year. A knowledge base edited today takes effect today.
3. Give prompt templates a version header. Date last tested, model tier used, formatting requirements relied on. Three lines will save a term's worth of confusion.
4. Ask during procurement whether providers can be switched. Retirement is a certainty, and switchability means the school is not tied to one vendor's timetable.

Next

That completes module 2. You now know how a model is made, and you know its three structural limits: it does not know your school, it fills every blank, and its knowledge stops on a date. Module 3 changes the angle, moving from inside the model to where models come from — the open-weights world, how to read a model card, and why a small model may suit a school better.

Continue with Hugging Face, the GitHub of AI. Previous lesson: where hallucination and bias come from. To compare current tiers and specifications across providers, see the AI model database.