Reference, verified July 2026
LLM knowledge cutoff dates
A single, plain table of the knowledge cutoff for every major AI model, ChatGPT, Claude, Gemini, Llama, Grok, and Perplexity, plus whether each one can read the live web. Checked against official model docs, with sources on every row.
Latest flagships
live webCutoff is what the model knows unaided. With a search tool, each of these can still read and cite pages published today.
Knowledge cutoff by model
Grouped by provider, newest models first. The web-access column is the important one: a cutoff only limits what a model knows before it searches.
OpenAI (ChatGPT)
OpenAI model docs| Model | Knowledge cutoff | Live web access |
|---|---|---|
| GPT-5.5current | December 2025 | Optional, via search tool |
| GPT-5.4 | August 2025 | Optional, via search tool |
| GPT-4.1 | June 2024 | Optional, via search tool |
| GPT-4o | October 2023 | Optional, via search tool |
| o3 (reasoning) | June 2024 | Optional, via search tool |
| o4-mini (reasoning) | June 2024 | Optional, via search tool |
Anthropic (Claude)
Claude model overview| Model | Knowledge cutoff | Live web access |
|---|---|---|
| Claude Opus 4.8current | January 2026 | Optional, via web search tool |
| Claude Sonnet 5current | January 2026 | Optional, via web search tool |
| Claude Fable 5current | January 2026 | Optional, via web search tool |
| Claude Haiku 4.5 | February 2025 | Optional, via web search tool |
Anthropic publishes a "reliable knowledge cutoff", the date through which the model’s knowledge is most complete. The broader training-data cutoff can be later.
Google (Gemini)
Gemini API model docs| Model | Knowledge cutoff | Live web access |
|---|---|---|
| Gemini 2.5 Procurrent | January 2025 | Optional, via Search grounding |
| Gemini 2.5 Flash | January 2025 | Optional, via Search grounding |
Google’s newer Gemini 3 family is available, but Google does not publish an explicit per-model knowledge cutoff for it, so it is omitted here rather than guessed.
Meta (Llama)
Llama 4 model card| Model | Knowledge cutoff | Live web access |
|---|---|---|
| Llama 4 (Scout, Maverick) | August 2024 | No, open weights, no built-in web tool |
Llama 4 is Meta’s latest open-weight, Llama-branded release. Meta’s most capable frontier work has since moved to a separate closed model line.
xAI (Grok)
xAI developer docs| Model | Knowledge cutoff | Live web access |
|---|---|---|
| Grok 4 seriescurrent | November 2024 | Yes, native real-time X and web search |
Grok has a training cutoff like any model, but in the product it defaults to real-time X and web search, so it routinely answers about current events.
Perplexity
Perplexity docs| Model | Knowledge cutoff | Live web access |
|---|---|---|
| Perplexity (Sonar and partner models)current | Not applicable, answers from live web | Yes, retrieval on nearly every query |
Perplexity is a retrieval product, not a single model. It runs a live web search on nearly every query, so there is no meaningful fixed knowledge cutoff.
Dates were verified against each vendor’s official model documentation in July 2026. Models we could not confirm from a primary source, including some very new releases, are intentionally left off rather than estimated. Cutoffs change often, so treat this as a snapshot and confirm on the linked source before citing a date.
What a knowledge cutoff actually means
The number is widely misread. Three things are worth holding onto, especially if your goal is to be mentioned in AI answers.
A cutoff is not a hard wall
Most current assistants can call a search or browsing tool. When they do, they read live pages and can answer about events long after their training cutoff, and they can cite your page. The cutoff only describes what the model knows without help.
Reliable cutoff vs training cutoff
Some vendors separate the date through which knowledge is most complete from the broader training-data cutoff. The reliable date is the safer number to quote, because knowledge near the very end of training is thinner.
Cutoffs move; this table is a snapshot
Vendors ship new models constantly and revise dates. Everything here was checked against official model pages in July 2026. Always confirm on the linked source before citing a specific date.
The cutoff is why retrieval matters
If models only knew their training data, no new brand could ever be mentioned. The reason you can still get cited is that assistants search the live web, read fresh pages, and quote them, regardless of when the model was trained.
So the practical takeaway from this table is not the dates themselves. It is that your visibility lives on the retrieval path: being one of the pages an engine fetches, trusts, and names at answer time.
Knowledge cutoff FAQ
What is ChatGPT’s knowledge cutoff?
It depends on the model. As verified in July 2026, OpenAI lists GPT-5.5 with a December 2025 cutoff, GPT-4.1 at June 2024, and GPT-4o at October 2023. When ChatGPT uses its search tool, it can still answer about newer events by reading live pages, so the cutoff only describes what it knows without searching. Confirm the current figure on OpenAI’s model documentation.
What is Claude’s knowledge cutoff?
Anthropic publishes a "reliable knowledge cutoff" per model. As verified in July 2026, current Claude models including Claude Opus 4.8 and Claude Sonnet 5 list a January 2026 reliable cutoff, while Claude Haiku 4.5 lists February 2025. Claude can also use a web search tool to read and cite live pages beyond that date.
What is Gemini’s knowledge cutoff?
As verified in July 2026, Google lists a January 2025 knowledge cutoff for Gemini 2.5 Pro and 2.5 Flash. Google’s newer Gemini 3 family does not publish an explicit per-model cutoff, so we do not list a date for it. With Search grounding enabled, Gemini can answer using live results regardless of the training cutoff.
Does a knowledge cutoff mean the AI cannot know recent events?
Not anymore, for most assistants. Without a tool, a model only knows what was in its training data up to the cutoff. But ChatGPT, Claude, Gemini, Grok, and Perplexity can all retrieve live web pages, which is exactly how they answer about current events and cite recent sources. Perplexity and Grok do this by default; others do it when the question calls for it.
How do I get my content into AI answers if models have a cutoff?
Through the retrieval path, not the training path. When an assistant searches the web to answer a question, it reads and cites live pages, so your content can appear even if it was published after the model’s cutoff. The work is making pages easy to retrieve and quote, then measuring which prompts actually cite you.
Related guides
Keep going deeper on AI search visibility across the rest of the guide series.
How AI Search Engines Work
The retrieve, rank, synthesize, and cite pipeline behind ChatGPT, Perplexity, Gemini, and Claude answers.
Read the guideAI Crawlers Explained
GPTBot, ClaudeBot, PerplexityBot, and more: what each AI bot does and how to allow or block it.
Read the guideHow to Rank in ChatGPT
What actually increases the odds of being surfaced, summarized, and cited in ChatGPT.
Read the guide