LLM Knowledge Cutoff Dates (Every Major Model Fact-Checked, June 2026)
-min.png)
Ask any model when its knowledge stops and you often get a wrong answer. Qwen maintainers have said openly that self-reported cutoff dates are unreliable because models are not trained to identify themselves. Blog posts copy each other. Aggregators lag new releases.
This guide fixes that. I fact-checked LLM knowledge cutoff dates across OpenAI, Anthropic, Google Gemini, Qwen, Kimi, and MiniMax using official API docs, model cards, and provider system prompts first, then cross-checked against OpenRouter and aggregator data. Every row notes the source tier so you know what to trust.
If you track AI visibility or run AEO , cutoff dates matter. A model that stopped learning in January 2025 cannot reliably describe your product launch from March 2026 unless it browses or you feed it context.
TL;DR
- Freshest western frontier cutoffs: GPT-5.5 (December 1, 2025) and Claude Fable 5 / Opus 4.8 (January 2026 reliable knowledge).
- Biggest surprise: Gemini 3.1 Pro and Gemini 3.5 Flash both list January 2025 in official Google docs despite May 2026 releases. Third-party blogs claiming January 2026 for 3.5 Flash are wrong.
- Chinese labs vary: Qwen 3.5+ uses a 2026 cutoff in DashScope system prompts; Kimi K2.6 is April 2025; MiniMax M3 is January 2026.
- Never trust a model's self-reported cutoff. Check the API docs or this table.
How we fact-checked
Tier 1 (official): Provider API documentation, model overview pages, HuggingFace chat templates with embedded cutoff strings, Anthropic Help Center.
Tier 2 (secondary): OpenRouter model metadata, aiknowledgecutoff.com, llm-bento.com, technical reviews citing vendor specs.
Tier 3 (unverified): Model self-reports when asked "what is your knowledge cutoff?" We excluded these from the tables unless no Tier 1 or 2 source exists.
Two cutoff types (Anthropic only): Anthropic publishes both a reliable knowledge cutoff (where knowledge is most extensive) and a training data cutoff (broader date range). Other providers usually publish one date.
Release vs cutoff gap: A model released in June 2026 can still have a January 2025 cutoff. New release does not mean fresh training data.
Checked on June 17, 2026.
Master comparison: freshest model per family
| Family | Flagship model | Knowledge cutoff | Release date | Source tier | Live data option |
|---|---|---|---|---|---|
| OpenAI | GPT-5.5 | December 1, 2025 | April 23, 2026 | Tier 1 | Web search, file search |
| Anthropic | Claude Fable 5 | January 2026 | June 9, 2026 | Tier 1 | Tool use, web (via tools) |
| Gemini 3.5 Flash | January 2025 | May 19, 2026 | Tier 1 | Search grounding | |
| Qwen | Qwen3.7 Max | 2026 (year-level) | May 2026 | Tier 1 (API prompt) | Web search tools |
| Moonshot | Kimi K2.6 | April 2025 | April 20, 2026 | Tier 2 | Tool use (no native browse) |
| MiniMax | MiniMax M3 | January 2026 | June 1, 2026 | Tier 1 (HF template) | Tool use |
OpenAI / ChatGPT knowledge cutoff dates
OpenAI publishes exact dates on each model page at developers.openai.com/api/docs/models .
| Model | API ID | Knowledge cutoff | Release | Context | Notes |
|---|---|---|---|---|---|
| GPT-5.5 | gpt-5.5 | December 1, 2025 | Apr 23, 2026 | 1.05M | Current OpenAI flagship |
| GPT-5.5 Pro | gpt-5.5-pro | December 1, 2025 | Apr 24, 2026 | 1.05M | Higher reasoning tier |
| GPT-5.4 | gpt-5.4 | August 31, 2025 | Mar 5, 2026 | 1.05M | Frontier workhorse |
| GPT-5.4 Pro | gpt-5.4-pro | August 31, 2025 | Mar 5, 2026 | 1.05M | Pro variant |
| GPT-5.4 mini | gpt-5.4-mini | August 31, 2025 | Mar 17, 2026 | 400K | Fast, cheaper |
| GPT-5.4 nano | gpt-5.4-nano | August 31, 2025 | Mar 17, 2026 | 400K | Lowest cost GPT-5.4 |
| GPT-5.3 Chat | gpt-5.3-chat | August 31, 2025 | Mar 3, 2026 | — | ChatGPT conversation model |
| GPT-5.3 Codex | gpt-5.3-codex | August 31, 2025 | Feb 24, 2026 | — | Agentic coding |
| GPT-5.2 | gpt-5.2 | August 31, 2025 | Dec 11, 2025 | 400K | Previous frontier |
| GPT-5.2 Pro | gpt-5.2-pro | August 31, 2025 | Dec 10, 2025 | — | Pro variant |
| GPT-5.2 Chat | gpt-5.2-chat | August 31, 2025 | Dec 10, 2025 | — | Chat variant |
| GPT-5.1 | gpt-5.1 | September 30, 2024 | Nov 12, 2025 | — | Older generation |
| GPT-5 | gpt-5 | September 30, 2024 | Aug 7, 2025 | — | First GPT-5 release |
| GPT-4.1 | gpt-4.1 | June 1, 2024 | Apr 14, 2025 | 1M | Still available via API |
| GPT-4o | gpt-4o | October 1, 2023 | May 13, 2024 | 128K | Retired from ChatGPT Feb 2026 |
Fact-check note: GPT-5.5 adds roughly four months of knowledge over GPT-5.4. If ChatGPT defaults to GPT-5.5 for your tier, anything after December 2025 needs browsing or pasted context.
Anthropic / Claude knowledge cutoff dates
Anthropic distinguishes reliable knowledge cutoff from training data cutoff. Source: Models overview and Help Center .
Claude Fable 5 and Mythos class (above Opus)
| Model | API ID | Reliable cutoff | Training cutoff | Release | Notes |
|---|---|---|---|---|---|
| Claude Fable 5 | claude-fable-5 | January 2026 | January 2026 | Jun 9, 2026 | Mythos-class, GA. Default in Claude Code Pro/Max |
| Claude Mythos 5 | claude-mythos-5 | January 2026 | January 2026 | Jun 9, 2026 | Limited availability (Project Glasswing) |
Fable 5 sits above the Opus line. Same January 2026 cutoff as Opus 4.8, with safety classifiers that can refuse requests (Opus 4.8 does not).
Opus family
| Model | API ID | Reliable cutoff | Training cutoff | Release |
|---|---|---|---|---|
| Claude Opus 4.8 | claude-opus-4-8 | January 2026 | January 2026 | May 28, 2026 |
| Claude Opus 4.7 | claude-opus-4-7 | January 2026 | January 2026 | Apr 16, 2026 |
| Claude Opus 4.6 | claude-opus-4-6 | May 2025 | August 2025 | Feb 5, 2026 |
| Claude Opus 4.5 | claude-opus-4-5 | May 2025 | August 2025 | Nov 24, 2025 |
| Claude Opus 4.1 | claude-opus-4-1 | January 2025 | March 2025 | Aug 5, 2025 (deprecated) |
Sonnet and Haiku (current)
| Model | API ID | Reliable cutoff | Training cutoff | Release |
|---|---|---|---|---|
| Claude Sonnet 4.6 | claude-sonnet-4-6 | August 2025 | January 2026 | Jan 17, 2026 |
| Claude Haiku 4.5 | claude-haiku-4-5 | February 2025 | July 2025 | Oct 15, 2025 |
| Claude Sonnet 4.5 | claude-sonnet-4-5 | January 2025 | July 2025 | Sep 29, 2025 |
Fact-check note: Fable 5 and Opus 4.8 share the freshest Anthropic cutoff at January 2026. Sonnet 4.6 has a newer training data cutoff (January 2026) but a lower reliable cutoff (August 2025), which means it may know recent facts unevenly.
Google Gemini knowledge cutoff dates
Google publishes cutoff dates on each model page. Source: Gemini API models .
| Model | Model ID | Knowledge cutoff | Release | Context | Notes |
|---|---|---|---|---|---|
| Gemini 3.5 Flash | gemini-3.5-flash | January 2025 | May 2026 (GA) | 1M | Flagship Flash, thinking on by default |
| Gemini 3.1 Pro Preview | gemini-3.1-pro-preview | January 2025 | Feb 19, 2026 | 1M | Also gemini-3.1-pro-preview-customtools |
| Gemini 3.1 Flash Lite | gemini-3.1-flash-lite | January 2025 | Mar 3, 2026 | 1M | Cost-optimized |
| Gemini 3 Flash Preview | gemini-3-flash-preview | January 2025 | Dec 17, 2025 | 1M | Preview predecessor to 3.5 Flash |
| Gemini 3 Pro Image Preview | gemini-3-pro-image-preview | January 2025 | Nov 20, 2025 | 65K | Image generation |
| Gemini 2.5 Pro | gemini-2.5-pro | January 2025 | Mar 25, 2025 | 1M | Previous Pro generation |
| Gemini 2.5 Flash | gemini-2.5-flash | January 2025 | Mar 25, 2025 | 1M | Previous Flash generation |
Fact-check correction: Several third-party articles claim Gemini 3.5 Flash has a January 2026 cutoff. Official Google documentation and the Gemini 3.5 Flash FAQ both state January 2025. The entire Gemini 3.x line shares the same static cutoff regardless of release date.
For post-January 2025 facts, Google expects you to use Search Grounding, not fresher weights.
Qwen knowledge cutoff dates
Alibaba does not publish a single cutoff page. Cutoffs vary by model generation and are embedded in DashScope API system prompts. Qwen maintainers warn that models hallucinate their own cutoff dates when asked directly.
| Model / series | Example API IDs | Knowledge cutoff | Source | Notes |
|---|---|---|---|---|
| Qwen3.7 Max | qwen3.7-max, qwen3.7-max-2026-05-20 | 2026 (year-level) | Tier 1 (DashScope prompt) | Flagship, May 2026 |
| Qwen3.7 Plus | qwen3.7-plus | 2026 | Tier 1 (DashScope prompt) | Cost-effective tier |
| Qwen3.6 Max Preview | qwen3.6-max-preview | 2026 | Tier 1 (DashScope prompt) | Apr 2026 preview |
| Qwen3.6 Plus | qwen3.6-plus | 2026 | Tier 1 (DashScope prompt) | Apr 2026 |
| Qwen3.6 Flash | qwen3.6-flash | 2026 | Tier 1 (DashScope prompt) | Fast tier |
| Qwen3.5 Plus | qwen3.5-plus, qwen3.5-plus-2026-02-15 | 2026 | Tier 1 (DashScope prompt) | Feb 2026 |
| Qwen3.5 Flash | qwen3.5-flash, qwen3.5-flash-2026-02-23 | 2026 | Tier 1 (DashScope prompt) | Vision-language Flash |
| Qwen3.5 397B A17B | qwen3.5-397b-a17b | 2026 | Tier 1 (DashScope prompt) | Open-weight flagship |
| Qwen3 Max | qwen3-max, qwen3-max-2026-01-23 | June 30, 2025 | Tier 2 (OpenRouter) | API tier, Sep 2025 |
| Qwen3 Next 80B (Instruct / Thinking) | qwen3-next-80b-a3b-* | September 30, 2025 | Tier 2 (OpenRouter) | MoE next-gen |
| Qwen3 2507 series | qwen3-235b-a22b-instruct-2507, qwen3-30b-a3b-*-2507 | June 30, 2025 | Tier 2 (OpenRouter) | 2507 = July 2025 release tag |
| Qwen3 Coder Plus / Flash | qwen3-coder-plus, qwen3-coder-flash | June 30, 2025 | Tier 2 (OpenRouter) | Coding-focused |
| Qwen3 Coder 480B A35B | qwen3-coder-480b-a35b | June 30, 2025 | Tier 2 (OpenRouter) | Largest coder model |
| Qwen3 VL series | qwen3-vl-* | March 31, 2025 | Tier 2 (OpenRouter) | Vision-language |
| Qwen3 dense / MoE base | qwen3-235b-a22b, qwen3-32b, qwen3-30b-a3b | March 31, 2025 | Tier 2 (OpenRouter) | Original Qwen3 launch |
| Qwen2.5 series | qwen2.5-* | June 30, 2024 | Tier 2 (OpenRouter) | Previous generation |
How to read Qwen version tags: 2507 in a model name means a July 2025 release snapshot, not a July 2025 knowledge cutoff. The 2507 Instruct models carry a June 30, 2025 cutoff per OpenRouter metadata.
Fact-check note: Qwen 3.5 and above claim 2026 knowledge in official API system prompts, making them among the freshest open models. Older Qwen3 variants stop at March or June 2025. Do not ask the model what it knows. Check the API ID you are calling.
Kimi (Moonshot AI) knowledge cutoff dates
Moonshot does not publish a centralized cutoff table. Kimi model docs at platform.moonshot.ai list capabilities but omit cutoff dates.
| Model | API ID | Knowledge cutoff | Release | Source | Notes |
|---|---|---|---|---|---|
| Kimi K2.7 Code | kimi-k2.7-code, kimi-k2.7-code-highspeed | Not published | Jun 12, 2026 | — | Coding-focused, thinking always on |
| Kimi K2.6 | kimi-k2.6 | April 2025 | Apr 20, 2026 | Tier 2 (vendor specs via reviews) | Multimodal, 256K context, agent swarm |
| Kimi K2.5 | kimi-k2.5 | Not published | Jan 27, 2026 | — | Visual agentic model |
| Kimi K2 Thinking | kimi-k2-thinking | Not published | Nov 6, 2025 | — | Open reasoning variant |
| Kimi K2 0905 | kimi-k2-0905 | December 31, 2024 | Sep 2025 | Tier 2 (OpenRouter) | Original K2 release |
| Kimi K2 0711 | kimi-k2-0711 | December 31, 2024 | Jul 2025 | Tier 2 (OpenRouter) | Earlier K2 snapshot |
Fact-check note: K2.6 specs cited in multiple technical reviews (training on 15.5T tokens, April 2025 cutoff) come from Moonshot release materials, not a standalone docs page I could fetch. OpenRouter still lists the original K2 snapshots at December 2024. Some API directories incorrectly list K2.6 as January 2025. I side with April 2025 for K2.6 based on consistent vendor-spec reporting.
K2.5 and K2.7 Code cutoffs are unconfirmed as of June 2026. Treat as April 2025 or newer until Moonshot publishes official dates.
MiniMax knowledge cutoff dates
MiniMax embeds cutoff dates in HuggingFace chat templates for some models. The release notes do not list cutoffs.
| Model | API ID | Knowledge cutoff | Release | Source | Notes |
|---|---|---|---|---|---|
| MiniMax M3 | MiniMax-M3 | January 2026 | Jun 1, 2026 | Tier 1 (HF chat template) | 1M context, multimodal |
| MiniMax M2.7 | MiniMax-M2.7, MiniMax-M2.7-highspeed | Not published | Mar 18, 2026 | — | Self-evolving agent model |
| MiniMax M2.5 | MiniMax-M2.5, MiniMax-M2.5-highspeed | January 2025 | Feb 12, 2026 | Tier 2 (llm-bento, reviews) | SWE-bench 80.2% |
| MiniMax M2.1 | MiniMax-M2.1, MiniMax-M2.1-highspeed | June 2025 | Dec 23, 2025 | Tier 2 (OpenRouter, langmart) | Polyglot coding |
| MiniMax M2 | MiniMax-M2 | September 2025 | Oct 27, 2025 | Tier 2 (llm-bento) | First M2 open release |
| MiniMax M1 | MiniMax-M1 | June 30, 2024 | Aug 2025 | Tier 2 (OpenRouter) | Long-context reasoning |
| MiniMax-01 | MiniMax-01 | March 31, 2024 | Jan 15, 2025 | Tier 2 (OpenRouter) | Text + vision, 4M context |
Fact-check note: M3 is the only MiniMax model with a Tier 1 cutoff (January 2026) baked into its official chat template. M2.5 at January 2025 means it is stale relative to M3 despite strong coding benchmarks. M2.7 cutoff is unpublished; likely between M2.5 and M3 based on release timing.
What the cutoff gap means in practice
| Provider | Flagship | Cutoff | Release | Gap (months) |
|---|---|---|---|---|
| OpenAI | GPT-5.5 | Dec 2025 | Apr 2026 | ~4 |
| Anthropic | Claude Fable 5 | Jan 2026 | Jun 2026 | ~5 |
| Gemini 3.5 Flash | Jan 2025 | May 2026 | ~16 | |
| Qwen | Qwen3.7 Max | 2026 | May 2026 | ~0 to 5 |
| Moonshot | Kimi K2.6 | Apr 2025 | Apr 2026 | ~12 |
| MiniMax | MiniMax M3 | Jan 2026 | Jun 2026 | ~5 |
Gemini has the largest gap between release hype and static knowledge. GPT-5.5 and Claude Fable 5 are the freshest western frontier models. Qwen 3.5+ and MiniMax M3 compete on cutoff freshness among Chinese labs.
When cutoff dates matter for SEO and AEO
Training data vs live answers: ChatGPT with browsing, Gemini with Search Grounding, and Perplexity with native search can answer post-cutoff questions. The cutoff still matters for default answers, API calls without tools, and which pages get cited from memory versus retrieval.
Brand visibility: If your product launched after a model's cutoff and the model names a competitor instead, the fix is usually content plus retrieval, not waiting for retraining. See how to improve brand visibility in AI search engines .
Do not trust one run: AI answers are non-deterministic. A model may know a pre-cutoff fact on one run and miss it on the next. Measure trends, not snapshots.
Frequently asked questions
What are LLM knowledge cutoff dates?
LLM knowledge cutoff dates mark the point after which a model's training data ends. The model has no built-in knowledge of events, products, or news after that date unless it uses browsing, search grounding, or you provide context in the prompt.
Which LLM has the most recent knowledge cutoff?
As of June 2026, Claude Fable 5, Claude Opus 4.8, Qwen 3.5+, and MiniMax M3 claim January 2026 or full-year 2026 cutoffs. GPT-5.5 is close behind at December 1, 2025. Gemini 3.5 Flash and 3.1 Pro remain at January 2025 despite recent releases.
What is ChatGPT 5.5's knowledge cutoff?
GPT-5.5 has a knowledge cutoff of December 1, 2025, confirmed on the official OpenAI model page . GPT-5.4 and GPT-5.2 use August 31, 2025.
What is Claude Fable 5's knowledge cutoff?
Claude Fable 5 has a reliable knowledge cutoff of January 2026 and a training data cutoff of January 2026, per Anthropic's models overview and Help Center. It released June 9, 2026.
What is Gemini 3.5 Flash's knowledge cutoff?
Gemini 3.5 Flash has a knowledge cutoff of January 2025, per official Google documentation . Blog posts claiming January 2026 are incorrect. Use Search Grounding for recent facts.
What is Gemini 3.1 Pro's knowledge cutoff?
Gemini 3.1 Pro Preview (gemini-3.1-pro-preview) has a knowledge cutoff of January 2025, same as the rest of the Gemini 3.x family.
Do Qwen models know their own cutoff date?
Usually not reliably. Qwen maintainers state models are not trained for self-identification and will give inconsistent cutoff answers. Use the API model ID and provider documentation instead of asking the model.
What is Kimi K2.6's knowledge cutoff?
Kimi K2.6 is widely reported at April 2025 based on Moonshot release specifications cited in technical reviews. Moonshot has not published a standalone cutoff page. Original Kimi K2 snapshots (0711, 0905) use December 31, 2024 per OpenRouter.
What is MiniMax M3's knowledge cutoff?
MiniMax M3 has a knowledge cutoff of January 2026, embedded in its official HuggingFace chat template. MiniMax M2.5 uses January 2025 and MiniMax M2 uses September 2025 per secondary sources.
How often do knowledge cutoffs update?
Major western providers update roughly once or twice per year per model generation. Chinese labs release faster but publish cutoff data less consistently. A new model name does not guarantee a newer cutoff.
Get new research on AI search, SEO experiments, and LLM visibility delivered to your inbox.
Powered by Substack · No spam · Unsubscribe anytime