AI knowledge cutoff dates (updated September 2026)

A knowledge cutoff is the date after which an AI model has no built-in information. The model was trained on text gathered up to that date, so it has not memorized anything published afterwards. Ask about an event from after the cutoff and the model either says it does not know, or it searches the web (if the product supports browsing) to answer.

It depends on the model behind ChatGPT. GPT-5.6 is still the default chat model (Luna for Free and Go, Sol for Plus and Pro), with a February 16, 2026 cutoff, so that is the date that applies to most ChatGPT users today. OpenAI's newest model, GPT-6 Astra (September 2026), has an April 30, 2026 cutoff and appears as "GPT-6 Pro" for Pro, Business and Enterprise users. GPT-6 Sol (April 20, 2026) and GPT-6 Luna (May 18, 2026) run in ChatGPT Work and Codex. GPT-5.5, the previous default, has a December 2025 cutoff. GPT-5.2 has August 2025, GPT-5 has September 2024, GPT-4o has October 2023, and the original GPT-4 had September 2021. ChatGPT can also browse the web, so current answers may use live results.

Each GPT-6 model has its own cutoff. GPT-6 Astra, released September 3, 2026, has an April 30, 2026 knowledge cutoff. GPT-6 Sol (April 20, 2026) and GPT-6 Luna (May 18, 2026) followed on September 22, 2026. Luna has the newest cutoff of the three, even though Astra is the most capable. All three come from OpenAI's developer model pages.

GPT-5.5 has a December 2025 knowledge cutoff and was released on April 23, 2026. It was the default ChatGPT model from May 2026 until GPT-5.6 took over in August 2026, so it is still the cutoff that applied to most ChatGPT answers for much of 2026.

GPT-4o has an October 2023 knowledge cutoff and was released on May 13, 2024. OpenAI retired GPT-4o from ChatGPT on February 13, 2026, in favor of the GPT-5 series (after briefly pulling it in August 2025 at the GPT-5 launch, then restoring it following user backlash).

GPT-4 was released on March 14, 2023. Its original training cutoff was September 2021. The default gpt-4 alias in the API now resolves to a newer snapshot with a knowledge cutoff around December 2023.

It varies by version. Claude Opus 5.5 and Claude Fable 5.1, both released in September 2026, have a June 2026 reliable knowledge cutoff. Claude Opus 5 has May 2026, while Claude Sonnet 5 and Claude Fable 5 have January 2026 cutoffs. Claude 3.5 Sonnet had April 2024, and Claude 3 had August 2023. We list Anthropic's reliable knowledge cutoff rather than its later training-data cutoff.

Claude Opus 4.8 has a January 2026 reliable knowledge cutoff and was released on May 28, 2026. Claude Fable 5 and Claude Sonnet 5 share that January 2026 cutoff, while Claude Opus 5 advances the reliable cutoff to May 2026 and Opus 5.5 to June 2026.

It depends which Gemini. Gemini 3.8 Flash, released September 2, 2026, has a March 2026 cutoff per Google's model card, with the same warning as 3.6 and 3.7 Flash that some domains still stop at January 2025. The Pro line (Gemini 2.5 Pro, 3 Pro, and 3.1 Pro) carries a January 2025 cutoff; Gemini 3.1 Pro, the newest Pro, is still a preview model. Gemini can also use Google Search grounding for current answers.

Grok 4 has a November 2024 knowledge cutoff and was released on July 9, 2025. xAI's newest model, Grok 4.7 (September 21, 2026), has a May 2026 cutoff — the only one stated on xAI's current models page. Grok 4.6 was published with a February 1, 2026 cutoff. xAI has not published one for Grok 4.1, Grok 4.3, or Grok 4.5. Grok also searches the live web and X by default, so it can answer about events after its cutoff.

DeepSeek does not publish a knowledge cutoff on any model card or product page. The one exception: its R1 research paper states a July 2024 cutoff for DeepSeek-V3-Base, the pretrained checkpoint both DeepSeek-V3 and DeepSeek-R1 are built from. Every other DeepSeek model, including V4, has no cutoff anywhere in DeepSeek's own material.

Alibaba does not publish a knowledge cutoff for any Qwen model. Its top model, Qwen3.8-Max, launched on August 3, 2026, followed by Qwen3.8-27B (August 14) and Qwen3.8-Flash (August 26). Cutoffs shown on third-party sites, or stated by Qwen when you ask it, do not come from Alibaba's documentation.

Z.ai (formerly Zhipu AI) does not publish a knowledge cutoff for any GLM model. The newest, GLM-5.3, launched on August 14, 2026 and is built on the same base model as GLM-5.2 (June 16, 2026). Some API resellers list cutoffs for older GLM models, but those dates do not appear in any Z.ai documentation.

Perplexity does not publish a knowledge cutoff for any Sonar model. Sonar answers from live web search rather than memory, so for most questions the cutoff matters far less than it does for a chat model answering offline. Note that the Sonar Chat Completions API is now legacy and retires on September 27, 2026, replaced by Perplexity's Agent API.

The training data cutoff is the last date that any data was included in training. The reliable knowledge cutoff is the date through which the model answers dependably, which is usually earlier because the most recent months are thin and less reliable. Release date is different again: it is when the public could first use the model, typically several months to over a year after the cutoff.

As of September 2026, Anthropic Claude Opus 5.5 and Fable 5.1 have the most recent published reliable cutoff at June 2026. OpenAI GPT-6 Luna follows at May 18, 2026, with xAI Grok 4.7 and Claude Opus 5 at May 2026. GPT-6 Astra is April 30, 2026. Google's Gemini 3.6, 3.7 and 3.8 Flash are March 2026, though Google warns that some domains still stop at January 2025. These dates move as new models ship, so we review the source documents monthly.

Not entirely. When a model browses the web (as ChatGPT, Gemini, Perplexity, Copilot, and Grok can), it can answer questions about events after its cutoff. But browsing only fires for some queries, the model still leans on its trained knowledge for everything else, and its judgment about what to trust is shaped by what it learned before the cutoff. The cutoff still matters for how a brand is understood.

Different cutoffs are a big reason. A model trained through January 2026 knows about products, funding rounds, and rebrands that a model frozen at October 2023 has never seen. The models also differ in training data, retrieval behavior, and how confidently they answer, so the same prompt produces different results across ChatGPT, Claude, Gemini, and the rest.

A model with an earlier cutoff has no memory of it. Until that content is either picked up by the model's web-browsing step or included in a future training run, the model answers as if your content does not exist. This is exactly why newer brands and pages struggle to get mentioned by AI engines, and why tracking your visibility across engines matters.

Three levers. First, publish content that browsing-enabled models can fetch live (clean pages, an llms.txt, fast load). Second, build the third-party signals (reviews, mentions, structured data) that models trust. Third, monitor which engines actually mention you, since each has a different cutoff and retrieval behavior. FixAEO's free AI visibility checker shows where you stand across major AI search engines.

Every few months. The major labs ship a new flagship model two to four times a year, and each one usually advances the cutoff. Because this is a fast-moving, current-facts topic, we review and update this page monthly.