Training data
The training-data cutoff and release date for 28 tracked models, newest cutoff first. A model cannot know about anything published after its cutoff unless you put it in the prompt.
Dates as published by each provider, pulled from the AI Gateway catalog on September 3, 2026. Lag is computed from the two published dates.
| Model | Provider | Knowledge cutoff | Released | Cutoff → release | Context |
|---|---|---|---|---|---|
| Claude Fable 5.1 | Anthropic | Jun 1, 2026 | Aug 31, 2026 | 2 mo | 1M tokens |
| Claude Opus 5 | Anthropic | May 2026 | Jul 24, 2026 | 2 mo | 1M tokens |
| Gemini 3.5 Flash-Lite | Mar 2026 | Jul 21, 2026 | 4 mo | 1M tokens | |
| Gemini 3.6 Flash | Mar 2026 | Jul 21, 2026 | 4 mo | 1M tokens | |
| Gemini 3.7 Flash | Mar 2026 | Aug 13, 2026 | 5 mo | 1M tokens | |
| GPT-5.6 Luna | OpenAI | Feb 16, 2026 | Jul 9, 2026 | 5 mo | 1.1M tokens |
| GPT-5.6 Sol | OpenAI | Feb 16, 2026 | Jul 9, 2026 | 5 mo | 1.1M tokens |
| GPT-5.6 Terra | OpenAI | Feb 16, 2026 | Jul 9, 2026 | 5 mo | 1.1M tokens |
| Claude Fable 5 | Anthropic | Jan 31, 2026 | Jul 1, 2026 | 6 mo | 1M tokens |
| Claude Opus 4.7 | Anthropic | Jan 31, 2026 | Apr 16, 2026 | 3 mo | 1M tokens |
| Claude Sonnet 5 | Anthropic | Jan 31, 2026 | Jun 29, 2026 | 5 mo | 1M tokens |
| Muse Glimmer 30B | Meta | Jan 4, 2026 | Aug 10, 2026 | 7 mo | 131k tokens |
| Claude Opus 4.8 | Anthropic | Jan 2026 | May 28, 2026 | 4 mo | 1M tokens |
| GPT-5.5 | OpenAI | Dec 1, 2025 | Apr 24, 2026 | 4 mo | 1M tokens |
| Claude Sonnet 4.6 | Anthropic | Aug 31, 2025 | Feb 17, 2026 | 6 mo | 1M tokens |
| GPT-5.4 | OpenAI | Aug 31, 2025 | Mar 5, 2026 | 7 mo | 1.1M tokens |
| Claude Opus 4.6 | Anthropic | May 31, 2025 | Feb 5, 2026 | 9 mo | 1M tokens |
| DeepSeek V4-Flash | DeepSeek | May 2025 | Apr 23, 2026 | 11 mo | 1M tokens |
| DeepSeek V4-Pro | DeepSeek | May 2025 | Apr 23, 2026 | 11 mo | 1M tokens |
| Gemini 3.1 Pro | Jan 2025 | Feb 19, 2026 | 13 mo | 1M tokens | |
| Gemini 3.5 Flash | Jan 2025 | May 19, 2026 | 16 mo | 1M tokens | |
| Kimi K2.7 Code | Moonshot | Jan 2025 | Jun 12, 2026 | 17 mo | 256k tokens |
| GPT-5.2 | OpenAI | Oct 2024 | Dec 11, 2025 | 14 mo | 400k tokens |
| Llama 4 Maverick | Meta | Aug 2024 | Apr 5, 2025 | 8 mo | 128k tokens |
| Llama 4 Scout | Meta | Aug 2024 | Apr 5, 2025 | 8 mo | 128k tokens |
| DeepSeek R1 | DeepSeek | Jul 2024 | Jan 20, 2025 | 6 mo | 128k tokens |
| GPT-4o | OpenAI | Sep 2023 | May 13, 2024 | 8 mo | 128k tokens |
| GPT-4o Mini | OpenAI | Sep 2023 | Jul 18, 2024 | 10 mo | 128k tokens |
Cutoffs published as a month are shown as the month. Models without a published cutoff are not listed.
The most recent cutoff each provider currently ships.
The knowledge cutoff is the date after which a model's training data stops. Events, releases, prices, and documentation published after that date are not in the model's weights, so it can only know about them if you put them in the prompt or let it search the web. Providers publish the cutoff as a month or a specific day; this page shows it exactly as published.
Claude Fable 5.1 (Anthropic) has the most recent published cutoff of the 28 models tracked here: June 1, 2026. 13 models have a cutoff in 2026: Claude Fable 5.1, Claude Opus 5, Gemini 3.5 Flash-Lite, Gemini 3.6 Flash, Gemini 3.7 Flash, GPT-5.6 Luna and others.
Across the 28 models here with both dates published, the median gap from knowledge cutoff to release is 6 months. That is the time providers spend on training, safety evaluation, and deployment. The gap is computed from the two published dates, not estimated.
Only for questions about recent events. A model with a 2024 cutoff can still out-reason one with a 2026 cutoff on math, code, or writing. Use the cutoff to decide whether you need to supply recent information yourself; use benchmarks and hands-on tests to judge capability.
Three ways: paste the source material into the prompt (documents, release notes, data), enable a web-search or retrieval tool if the provider offers one, or use a model whose cutoff covers the period you care about. For anything that changes often, such as pricing, treat the model's built-in knowledge as stale and supply the current figure.
GPT-4o Mini has the oldest published cutoff on this page: September 2023. 6 tracked models stop before 2025, including GPT-5.2, Llama 4 Maverick, Llama 4 Scout, DeepSeek R1, GPT-4o. Older models can still be the right pick when they are cheaper and the task does not depend on recent knowledge.