Knowledge cutoffs and when models stop updating

Week two of term, a Secondary 6 careers teacher prints an AI-generated JUPAS timetable for the notice board. A colleague from the next class points at one line: "That deadline changed on the exam authority site last month. Your sheet is the old version."

The parameters compress a vast slice of public knowledge, but that slice stops on a fixed day. This lesson closes module 2 on structural limits, following where hallucination and bias come from on stale facts, and opening the door to module 3 on open-weight models.

---

The calendar photo in the safe

Picture the school office pinning a paper calendar in the corridor, photographing it, and locking the camera in a safe. Later the Education Bureau moves an exam date, joint admissions shifts a deadline, and a grant scheme extends its application window. The photograph inside the safe does not update itself.

That is what a model does with the public world at the moment training data collection stops. Lesson 6 already explained why: pre-training runs thousands of GPUs for weeks to months, so collection must halt on some date. The awkward part is that the model will not announce the photo is old. It rebuilds something that reads like this year from last year's patterns, which is why cutoffs and hallucinations so often arrive together.

Hard dates on the provider's site

Cutoffs are not guesses, and they are not reliable if you ask the model. Checking major providers in 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, Claude Sonnet 5's as January 2026, and the older Claude Haiku 4.5 back at February 2025.

Three practical notes. Different tiers from the same provider can sit far apart, so "using the newest vendor" is not the same as "using the freshest knowledge". The cutoff usually predates release by several months because training, evaluation and safety testing all take time. Do not ask the model for its own cutoff; that date was written into documentation after training finished, and the model may simply get it wrong. Use the provider page.

Same name, different weights

Besides "when memory stopped", two other clocks can strike mid-year.

Quiet version updates

A tier may keep the same public name while the weights underneath change over the summer. Prompt templates your panel tuned last year can suddenly grow longer, shrink, or turn a table into bullets. Quality may improve, but your print layout needs redoing.

Retirement dates

Providers publish retirement dates per tier. Anthropic's documentation, for instance, gives current models a "not sooner than" commitment. If a workflow depends deeply on one tier's tone or formatting, that dependency will eventually break. One procurement line worth writing down: can the platform switch providers? Switchability turns a tier change into a re-test rather than a redeployment.

Search is an add-on, not a reset

Some interfaces advertise web search, but that sits on top of the cutoff rather than removing it. Ask for "significant education news in Hong Kong last week" and three things may happen: the interface actually searched, you pasted relevant material earlier in the chat, or the model guessed from old patterns and looked as if it searched. The third is the dangerous one because the output shape matches the first two.

The cutoff governs what it remembers; search governs whether it looks anything up right now. Those are two separate questions.

For anything about this school year — exam dates, admissions, curriculum changes, bureau circulars — treat the official website as authority and use the AI only to organise what you have verified. School-specific material belongs in the knowledge base, edited today and effective today, without waiting for a new model generation.

Fifteen minutes in July for the English panel

One secondary school's English panel built nine AI prompt templates last year: reading worksheets, vocabulary lists, parent circulars and DSE oral practice guides. At the start of the new year two colleagues reported the same problem. Nobody had edited the templates, yet every output had "drifted" — some gained two paragraphs of background, and a "exactly three sentences" rule became five.

The cause was a default tier change in July. The IT coordinator had not flagged it at panel level because the vendor emailed only the administrator.

Their fix: a three-line header on each template — date last tested, tier tested on, three formatting rules relied on (for example "must output a table", "under 150 words", "three bullet points"). Each July, at the panel planning meeting, they spend fifteen minutes re-running all nine and fixing drift on the spot. The panel head said: "It used to be a surprise. Now it is on the calendar."

Common misreadings

• The claim: A polished timetable or grant deadline must be current because it looks complete. What actually happens: Structure and plausibility do not guarantee the year is right. Self-check: do the dates in the output match the exam authority site?
• The claim: A "connected" interface removes the need to verify. What actually happens: Search can fail or mix with memory. Self-check: ask whether the answer came from training memory or live retrieval, and demand links.
• The claim: Version tracking is IT's job when something breaks. What actually happens: Tier changes are routine. Self-check: does your shared template folder include a "date last tested" field?

Try it once

Time: about twelve minutes. Materials: any AI tool.

Task: paste the prompt below and record the full answer:

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.

Hand in: four lines of notes: (1) did it give last week's news; (2) could it list clickable sources; (3) did it claim memory or retrieval; (4) does your interface actually offer search.

Done right: you can classify the answer as genuine retrieval, material from the chat, or stale memory plus hallucination, and name one verification step you would use for "this school year" tasks (for example always opening the exam authority site).

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