Knowledge Cutoff

A knowledge cutoff is the date after which a language model has no training data. Anything published later is not part of the model's parametric knowledge, so the model can only know about it if it searches the web, reads a provided document, or uses another form of retrieval.

Vendors publish cutoffs in their model documentation, and the dates have moved much closer to release than they were a few years ago. As of October 2026, the official figures for current flagship models are:

• OpenAI GPT-6 Sol: April 20, 2026. GPT-6 Astra: April 30, 2026. GPT-6 Luna: May 18, 2026.
• Anthropic Claude Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1: June 2026 (Claude help center).
• Google Gemini 4 Argon and Gemini 3.8 Flash: March 2026, with Google noting that some domains may be less current (January 2025 in places for 3.8 Flash).
• xAI Grok 4.7: May 2026.
• DeepSeek V4 family: not published. The DeepSeek-R1 paper gave July 2024 for the older V3 base model.

Two details make cutoffs less precise than a single date suggests. First, Anthropic distinguishes a 'reliable knowledge cutoff', the date through which the model's knowledge is most complete, from a 'training data cutoff', which can be later. Coverage thins out in the final months before any cutoff because the web has not yet written much about recent events. Second, there is always a gap between cutoff and release. GPT-6 Sol's cutoff is April 20, 2026, but it reached the API on September 22, 2026, so it launched about five months behind the world.

For GEO, the cutoff splits a brand's AI visibility into two layers. What the model learned before the cutoff shapes how it describes you when it answers without searching: which category it puts you in, which competitors it names next to you, which prices and features it remembers. Everything after the cutoff (a rebrand, a new product line, updated pricing) exists only if the assistant searches and finds it. Assistants that search by default, such as Perplexity, Google AI Overviews and ChatGPT for many query types, reduce the problem, but they still start from the model's prior beliefs when deciding what to search for. Grounding queries generated from an outdated mental model can miss your newest pages entirely.

The practical consequences are concrete. A company founded after a model's cutoff has no parametric presence in that model at all and depends entirely on retrieval: crawlable pages, access for AI crawlers such as OAI-SearchBot and Claude-SearchBot, and third-party coverage that search engines index. An established brand faces the opposite risk, where the model confidently repeats old facts. Test both by asking each assistant about your brand with web search off (where the product allows it) and on, and compare the answers. When a new model generation ships with a later cutoff, rerun your tracked prompts: changes in how the model describes you often show up then, not gradually.