--- title: Exploring Claude/GPT Knowledge Cutoffs | daily.dev description: A reverse-engineering analysis of frontier LLM training timelines using three probing techniques: historical quiz scoring, self-reported date estimation, and... canonical: https://daily.dev/posts/exploring-claude-gpt-knowledge-cutoffs-dfbea5aty twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Exploring Claude/GPT Knowledge Cutoffs | daily.dev og:description: A reverse-engineering analysis of frontier LLM training timelines using three probing techniques: historical quiz scoring, self-reported date estimation, and... og:url: https://daily.dev/posts/exploring-claude-gpt-knowledge-cutoffs-dfbea5aty og:image: https://api.daily.dev/og/posts/DFBeA5aty.png og:image:alt: Exploring Claude/GPT Knowledge Cutoffs og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Exploring Claude/GPT Knowledge Cutoffs **[Hacker News](https://daily.dev/sources/hn)** · 6 min read · 3 upvotes · 0 comments ## Summary A reverse-engineering analysis of frontier LLM training timelines using three probing techniques: historical quiz scoring, self-reported date estimation, and self-identity questioning. Key findings include that Anthropic's Opus 4.7+ models share a pre-training cutoff around late December 2025, GPT-5.6 has a distinct checkpoint finishing around late February 2026, and Opus 5 claims a May 2026 cutoff but shows no knowledge beyond January 2026. A notable observation is that Anthropic's Sonnet 5 frequently self-identifies as GPT-4, suggesting older ChatGPT conversation data is present in Claude's training mixture. The analysis also surfaces evidence that labs train on real user chat sessions from their own products. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What does self-identification probing reveal about whether Claude models were trained on ChatGPT outputs? Probing suggests Anthropic's Sonnet 5 model will regularly self-identify as GPT-4 when pushed without grounding context, hinting that older ChatGPT conversation data may have entered Claude's training mixture, possibly carried through the Sonnet 3.5 lineage. In a follow-up test asking models to mimic another model's identity quirks, Claude models reproduced OpenAI quirks at 68% accuracy, while OpenAI models only managed 8% on Claude quirks. _Anyone tracking how training data quietly shapes model behavior can follow deep dives like this on daily.dev._ ### How can you estimate the pre-training cutoff date of a language model without official documentation? One method builds a multiple-choice quiz from daily Wikipedia facts (like a page listing events by date) and measures each model's error rate across time, since the point where accuracy drops sharply approximates when the model's training data loses coverage. This approach suggested Anthropic's Opus 4.7 and later models share a training run cutting off around late December 2025. _Developers evaluating model recency and reliability can find this kind of probing research on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 24 comments across hackernews (as of 2026-09-14). **TL;DR:** Commenters found the reverse-engineering analysis genuinely interesting and largely credible, with an OpenAI employee chiming in to clarify how model versioning actually works; the main debates center on whether labs train on competitor outputs and whether knowledge cutoffs are domain-partitioned rather than uniform. **Sentiment:** 60% positive · 30% mixed · 10% skeptical **The case for** - An OpenAI employee confirmed that API model weights are kept fixed, lending credibility to the post's methodology of probing stable checkpoints. - The finding that Sonnet 5 self-identifies as GPT-4 is seen as plausible evidence of training on public chat-arena data rather than direct distillation. - The analysis surfaces a useful distinction between marketing model names (which may be updated silently) and actual weight checkpoints. **The pushback** - The assumption that a marketing name like 'Opus 5' maps to a single, fixed model is questioned — silent updates without name changes are confirmed to happen in consumer products. - Some commenters suspect knowledge cutoffs may be domain-partitioned (e.g., software changelogs vs. current events) rather than a single uniform date, which the post's methodology may not fully capture. - The post's charitable interpretation of Anthropic not distilling ChatGPT outputs is disputed, with commenters noting public datasets like LMSYS Chatbot Arena make such training data freely available regardless of ToS. **By community** - hackernews (positive): Broadly receptive and engaged, with an OpenAI insider adding rare first-hand clarification, while a subset of commenters pushed back on methodology assumptions and the training-data ethics framing. **Hottest debate:** Whether Claude's self-identification as GPT-4 reflects deliberate distillation of competitor outputs or merely incidental training on publicly available chat-arena datasets. **Open questions** - Do labs update shared knowledge-base snapshots or RAG-style injected data independently of model weights, effectively changing what a 'model' knows without a version bump? - Are knowledge cutoffs truly uniform across domains, or do different topic areas (e.g., software changelogs vs. current events) have meaningfully different effective cutoffs? - How much time elapses between a model's training cutoff and its public release, and can this gap be reliably estimated from the outside? **Highlights** > I can explain how we do it at OpenAI. In the API, we keep the models fixed. There are tiny caveats like rare bug fixes or models like `chat-latest`, but this is spiritually true. Suspicions of models changing over time are either human hallucinations or bugs on our end. However, in ChatGPT, we sometimes update models without changing their names. For example, we recently launched an update to GPT-5.6 Sol in ChatGPT (https://openai.com/index/improving-gpt-5-6-sol-in-chatgpt/). Our goal isn't to be opaque or sneaky, but just to not exhaust people trying to keep track of little changes. When the changes are big, we give models a new name so that people know to expect something different. > — [tedsanders on hackernews · 2 comments](https://news.ycombinator.com/item?id=49249585) > That's an good answer to the question when taken in strict terms of a model version = weights. But, insofar as: 1. "a model" as presented to the user, isn't just its weights, but also anything else happening on the "business layer" (though this maybe applies more to ChatGPT than "direct" model access via the API); and 2. said business layer has any "knowledge base"-type stuff going on in it (i.e. automatic or tool-call-triggered embedding of results from search of some vector-DB into which has been embedded distilled pre-validated trustworthy info — like per-user memory mechanisms, but searching + injecting from global shared data sources); and 3. said "knowledge base" mechanisms are where most of the up-to-date, fast-changing info a model "knows" (without having to do a web search) is actually coming from; ...then do y'all ever update the pinned knowledge-base data snapshot associated with the model version, without updating the weights themselves? > — [derefr on hackernews · 1 comments](https://news.ycombinator.com/item?id=49250740) > Given how competitive the space is any form of waiting seems like it has more downside than upside to me. If you have the new "best" model you may only have a few weeks of time in the market before some other lab releases a new model that beats yours. This means you should get it out ASAP so you can maximize the time during which your model is "best". Once your model isn't best any more you're going to lose a lot of revenue to the new leader. > — [simonw on hackernews](https://news.ycombinator.com/item?id=49248240) > In my tests, I compiled the changelogs from various projects and checked how much functionality the model could recall over several test runs. In some cases, even with the latest models, the cutoff was around June 2024. Beyond that, it didn't recognize the features. However, it did know the cause of the last pope's death and even which hospital he was in (April 2025). However, the model did not know when Trump was sworn for the second time (Jan 2025) - but it did know THAT he was elected a second time (Nov 2024). > — [tobwen on hackernews](https://news.ycombinator.com/item?id=49249296) > Distilling is only prohibited from chats you prompt. But there are lots in the open like the LMSYS Chatbot Arena data. It's a essentially one of the largest public sources of preference data, maybe still useful for boostraping RLHF and successor techniques (possibly a more highly educated contributor base than the low paid contractors), and was full of GPT-4-era inputs. > — [cma on hackernews · 1 comments](https://news.ycombinator.com/item?id=49253742) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=49244085) · 136 points · 24 comments ## Similar posts on daily.dev - [Opus 4.6 and Codex 5.3](https://daily.dev/posts/opus-4-6-and-codex-5-3-4ynvsyxn5) · Tech Lead Digest · 0 upvotes · 0 comments - [GPT-5.4 vs Claude Opus 4.6: a guide to choosing the right model](https://daily.dev/posts/gpt-5-4-vs-claude-opus-4-6-a-guide-to-choosing-the-right-model-bhtobehwf) · portkey · 1 upvotes · 0 comments - [I gave Claude, Gemini, and ChatGPT the same wrong fact, and only one of them caught it](https://daily.dev/posts/i-gave-claude-gemini-and-chatgpt-the-same-wrong-fact-and-only-one-of-them-caught-it-00ydijucq) · XDA Developers · 0 upvotes · 0 comments - [Why GPT-5.4, Claude, and Gemini can’t agree on basic, real-world facts](https://daily.dev/posts/why-gpt-5-4-claude-and-gemini-can-t-agree-on-basic-real-world-facts-q7pcokdcw) · The New Stack · 0 upvotes · 0 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#llm](https://daily.dev/tags/llm), [#claude](https://daily.dev/tags/claude), [#gpt](https://daily.dev/tags/gpt) [View this post on daily.dev](https://daily.dev/posts/exploring-claude-gpt-knowledge-cutoffs-dfbea5aty) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Exploring Claude/GPT Knowledge Cutoffs","url":"https://daily.dev/posts/exploring-claude-gpt-knowledge-cutoffs-dfbea5aty","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/exploring-claude-gpt-knowledge-cutoffs-dfbea5aty"},"datePublished":"2026-08-10T20:53:35.238Z","dateModified":"2026-09-14T06:04:34.861Z","description":"A reverse-engineering analysis of frontier LLM training timelines using three probing techniques: historical quiz scoring, self-reported date estimation, and...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/59385e7fb95d9ec76f3e36af03d708a2?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/59385e7fb95d9ec76f3e36af03d708a2?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/exploring-claude-gpt-knowledge-cutoffs-dfbea5aty","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,llm,claude,gpt","timeRequired":"PT6M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"Exploring Claude/GPT Knowledge Cutoffs"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/exploring-claude-gpt-knowledge-cutoffs-dfbea5aty#faq","mainEntity":[{"@type":"Question","name":"What does self-identification probing reveal about whether Claude models were trained on ChatGPT outputs?","acceptedAnswer":{"@type":"Answer","text":"Probing suggests Anthropic's Sonnet 5 model will regularly self-identify as GPT-4 when pushed without grounding context, hinting that older ChatGPT conversation data may have entered Claude's training mixture, possibly carried through the Sonnet 3.5 lineage. In a follow-up test asking models to mimic another model's identity quirks, Claude models reproduced OpenAI quirks at 68% accuracy, while OpenAI models only managed 8% on Claude quirks. Anyone tracking how training data quietly shapes model behavior can follow deep dives like this on daily.dev."}},{"@type":"Question","name":"How can you estimate the pre-training cutoff date of a language model without official documentation?","acceptedAnswer":{"@type":"Answer","text":"One method builds a multiple-choice quiz from daily Wikipedia facts (like a page listing events by date) and measures each model's error rate across time, since the point where accuracy drops sharply approximates when the model's training data loses coverage. This approach suggested Anthropic's Opus 4.7 and later models share a training run cutting off around late December 2025. Developers evaluating model recency and reliability can find this kind of probing research on daily.dev."}}]} ```