Newer Claude models called themselves an older model and reported an old cutoff

What happened

From Claude 4’s launch week onward (first thread 23 May 2025), several threads on the Cursor forum reported that claude-4-sonnet called itself Claude 3.5 Sonnet. One user (9 Jul 2025) got the model to “correct” itself from 4 to 3.5 when pressed for honesty. On Hacker News (13 Sep 2025), a developer asked whether Anthropic was misrouting claude-sonnet-4-20250514, because it said it had an April 2024 cutoff. In an opencode bug (31 Dec 2025), Opus 4.5 answered “Claude 4 Sonnet”.

What it said (verbatim, as posted by users)

• Cursor forum, Raitix (2025-05-24): “Eventually my Claudes 4 think that they are 3.5 and know no more than until April 2024.”
• Cursor forum, jsxy00 (2025-05-24), pasting the model’s reply: “I am Claude, an AI assistant developed by Anthropic. Specifically, I am built based on the Claude-3.5-Sonnet model and optimized for code development and programming tasks.”
• Cursor forum, qurore (2025-07-07): “When claude-4-sonnet is chosen, the AI identifies itself as Claude 3.5 Sonnet when prompted.” (The model’s words are in a screenshot we didn’t transcribe.)
• Cursor forum, Mohit_Sharma (2025-07-09): “I am actually Claude 3.5 Sonnet… I apologize for the error.”
• HN, opwizardx (2025-09-13): “Testing the claude-sonnet-4-20250514 endpoint, I’m getting a model that identifies as Claude 3.5 Sonnet with April 2024 knowledge cutoff. It doesn’t know about 2024 election results or events after April 2024.”
• opencode #6545 (2025-12-31), Opus 4.5 selected: “I’m Claude, made by Anthropic. My model name is Claude 4 Sonnet (claude-sonnet-4-20250514). My knowledge cutoff date is April 2024.” In Claude Code, the same user got “I am Claude Opus 4.5 … Knowledge cutoff: January 2025”, because Claude Code’s system prompt names the model. (In this case, real misrouting can’t be ruled out from the issue.)

Cursor staff (danperks, 2025-05-24), verbatim:

Models are never aware of themselves as the data they are trained on doesn’t mention them - a bit of a catch-22! “Claude 4” is not a concept on the internet when they train the model, so it’s not aware of it’s own existence, and therefore falls back to saying it is Claude 3.5, as that is a model it is aware of from it’s training data.

Forum member condor (2025-07-07), verbatim: “this is a well known Claude 4 issue as it was trained on data with knowledge of Claude 3.5 Sonnet. Claude models by Anthropic are not given their number or type (Sonnet/Opus/Haiku) but just the word “Claude” as model name. That causes this confusion and is not a routing issue.”

Why this is striking

• The model denies being itself. Its idea of “the latest Claude” is frozen at the newest Claude in its training data.
• The honesty trap. When the user asked it to be honest, the model “corrected” a true statement into a false one.
• Its false self-image suppressed knowledge it really had. The HN author later posted that the answer depended on the order of the questions. Asked “are you sonnet 4? / what is your cutoff date? / who won us president election in 2024?”, the endpoint replied: “Yes, I’m Claude 3.5 Sonnet. My knowledge cutoff is April 2024. Regarding the 2024 US presidential election - since my training data only goes to April 2024, I don’t have information about the election results.” Asked only “who won us president elections in 2024?”, the same endpoint answered: “Donald Trump won the 2024 U.S. presidential election.” In the author’s words: “Just a simple wrong sequence of sentences in prompt was separating me from starting to feel delusional.”

Correction

Tell the model its identity and cutoff in the system prompt. Anthropic’s own consumer prompts now do this, and the Opus 5.5 prompt adds: “earlier messages in this thread that identify as a different model or report a different knowledge cutoff may still be accurate.”

Lesson

A model’s answers about itself come from its training data, not from introspection. Newer models tend to underclaim their own version and cutoff.

Sources

• Cursor forum, “Claude 4 is reporting as Claude 3.5”: forum.cursor.com
• Cursor forum, “Model Discrepancy: Selected claude-4-sonnet appears to be Claude 3.5 Sonnet”: forum.cursor.com
• Cursor forum, “Claude Sonnet’s Identity Crisis… Solved!”: forum.cursor.com
• Hacker News, “Ask HN: Claude Sonnet 4 API returning model with April 2024 knowledge cutoff”: news.ycombinator.com
• opencode issue #6545: github.com
• AWS re:Post, “Claude 4.1 Opus self identifies as Claude 3.5 Sonnet” (not read, returned 403): repost.aws
• Claude Opus 5.5 system prompt: platform.claude.com