AI assistants write and summarize very well. Their weaknesses are twofold, and both are often mixed up: fabrication and knowledge that is limited to a certain date. The issue of fabrication has been discussed in the article about the materials attached to AI. This article addresses the second issue, as you can verify the date yourself in a minute.
What is knowledge cutoff
Language models are trained on data up to a certain date. Events after that date are unknown to them unless they are given access to search or documents.
Anthropic distinguishes between two dates. Training data cutoff is the limit of their training data. Reliable knowledge cutoff is the limit of the most extensive and reliable knowledge. For Claude Haiku 4.5, its reliable knowledge extends to February 2025, its training data goes up to July 2025, and the model was released on October 15, 2025. From February to July 2025, its knowledge is no longer as reliable as before, and after July 2025, it is completely empty.
New models do not necessarily have new knowledge
The dates below are taken from the official page of the creators, as of October 3, 2026. For the two Claude models, the reliable knowledge cutoff is listed.
| Model | Official knowledge cutoff | Distance to October 2026 |
|---|---|---|
| Claude Opus 5.5 | June 2026 | about 4 months |
| GPT-6 Sol | April 20, 2026 | about 5 months |
| Claude Haiku 4.5 | February 2025 | about 20 months |
| Gemini 3.5 Flash | January 2025 | about 21 months |
Gemini 3.5 Flash is in stable release status, and its documentation page was updated on September 23, 2026. Its knowledge cutoff remains January 2025. Therefore, the right question is not "is this model new or old," but rather "when is its cutoff date."
The quickest to become outdated are usually rates and prices, newly revised regulations, software versions, schedules, and who holds what position.
Why models rarely answer "I don't know"
Research from OpenAI published in September 2025 points to how models are evaluated. If the evaluation only counts correct answers, guessing is always more advantageous. For example, with birth dates: a random guess has a 1 in 365 chance, while "I don't know" is guaranteed to score zero.
The SimpleQA test in the same report shows the consequences:
| Model | Withholding answers | Correct | Incorrect |
|---|---|---|---|
| OpenAI o4-mini | 1% | 24% | 75% |
| gpt-5-thinking-mini | 52% | 22% | 26% |
The accuracy is almost the same. The difference is that the first model almost never holds back, resulting in nearly three times as many errors.
This research also explains the origin of fabrications. Rare facts without patterns, such as someone's birth date, cannot be guessed from language patterns. Dates, numbers, and names are also precisely the inputs you usually ask about the latest matters.
Search features help, but are not automatic
Google's documentation for Grounding with Google Search explains the sequence: the model first weighs whether searching will improve the answer. If it decides to search, its answer includes source links per sentence fragment. This means that having the search feature enabled does not guarantee that the answer is a result of the search.
Five checks before using answers about recent matters
- Ask yourself: could this have changed in the last two years? If so, the AI answer is just a starting point.
- Look for the cutoff date of the model you are using on the official page of the creator.
- Check the source links. Answers without links come from the model's memory.
- Allow the model to answer "I don't know." Anthropic's guidelines state that this simple method can drastically reduce incorrect information.
- Request quotes from the documents you attach, then strike out claims that lack citations.
If you prefer to see the explanation directly, there is a clip.
Sources
- Anthropic, Models overview, Claude Platform Docs, accessed October 3, 2026: https://platform.claude.com/docs/en/models/overview
- Anthropic, Claude Haiku 4.5, model specification page, Claude Platform Docs: https://platform.claude.com/docs/en/models/haiku-4-5/overview
- Anthropic, Reduce hallucinations, Claude Platform Docs: https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations
- OpenAI, Why language models hallucinate, September 5, 2025, along with its paper (Kalai et al., arXiv 2509.04664): https://openai.com/index/why-language-models-hallucinate/
- OpenAI, GPT-6 Sol Model, OpenAI API Docs: https://developers.openai.com/api/docs/models/gpt-6-sol
- Google AI for Developers, What's new in Gemini 3.5 Flash, updated September 23, 2026: https://ai.google.dev/gemini-api/docs/whats-new-gemini-3.5
- Google AI for Developers, Grounding with Google Search, updated September 23, 2026: https://ai.google.dev/gemini-api/docs/google-search