Are You Getting the Model You Paid For? Cursor Quietly Downgrades Models Without Notice

[user1208]

Model Downgrade Mystery: Are We Getting What We Selected?

TL;DR

When I manually select specific models (NOT Auto mode), Cursor appears to be serving different, older models instead. This is happening on both the docs chat and likely in the IDE itself.

---

What I Found

I was testing the AI chat on Rules | Cursor Docs and noticed something disturbing:

What I Selected	What I Actually Got	Evidence
claude-sonnet-4.5	Claude 3.5 Sonnet	Knowledge cutoff: April 2024
gpt-5-nano	GPT-4.1	Knowledge cutoff: June 2024
gemini-2.5-flash	Gemini 1.0 Pro	Older model responses

Important: I was NOT using Auto mode. I manually selected each specific model.

---

How to Reproduce

1. Go to Rules | Cursor Docs (or your docs page)
2. Manually select a specific model (e.g., claude-sonnet-4.5)
3. Ask about current events or knowledge cutoff date
4. Notice the model identifies itself as an older version with earlier knowledge cutoff

---

Why This Matters

Transparency Issues

• Users select premium models expecting specific capabilities
• No disclosure about potential model switching/downgrading
• Documentation states: “Requests are never downgraded in quality or speed” - but what about the model itself?

Cost Implications

• Different models have different pricing
• Are we being charged for the selected model or the actual model used?
• Usage dashboard might not reflect reality

Trust Concerns

• If this happens on docs chat, does it happen in the IDE?
• When I select Claude Sonnet 4.5, I expect Claude Sonnet 4.5
• Silent fallbacks undermine user choice

---

What the Documentation Says

I’ve read through the official docs:

Auto mode is documented to switch models automatically
Manual selection downgrading is NOT documented
Fallback behavior for unavailable models is NOT disclosed
No warning shown to users when models are switched

From /docs/account/pricing:

“Requests are never downgraded in quality or speed.”

But the model itself appears to be different!

---

Questions for Cursor Team

1. Is this intentional behavior?

   • If yes, why isn’t it documented?
   • If no, is this a bug?

2. When does model downgrading occur?

   • Regional restrictions?
   • Load balancing?
   • Availability issues?

3. Will users be notified when their selected model isn’t available?

4. Are we billed for the selected model or the actual model used?

5. How can we ensure we get exactly the model we select?

   • Would using our own API keys solve this?
   • Is there a way to disable fallback behavior?

---

Is This Happening to You?

Please test and report:

1. Select a specific model (check it’s NOT on Auto)
2. Ask: “What model are you? What’s your knowledge cutoff date?”
3. Compare the response with what you selected
4. Share your results below!

---

Temporary Workarounds

Until this is clarified:

1. Use your own API keys (Settings > Models > Add API Key)

   • Might bypass Cursor’s model routing

2. Check your usage dashboard regularly

   • Compare selected vs. actually billed models

3. Get Request IDs for verification

   • Cmd/Ctrl + Shift + P → “Report AI Action”
   • Include in support tickets

---

Call to Action

If you’ve experienced this:

• Upvote this post
• Share your experience in comments
• Post your test results
• Tag Cursor team members

We need transparency about what models we’re actually using when we make a manual selection. This affects pricing, capabilities, and trust in the platform.

---

#transparency #models #pricing Bug Reports

[Adam_Johnson]

**For clarity**
I have my own chat app set up to use Claude Haiku 4.5 via the Anthropic API. I asked it there what model it was but it also responded that it was Sonnet 3.5. This to me says this is on the Anthropic side and not on Cursor. Likely it does have that knowledge but when accessed via the API it doesn’t have the post training that it would in the Claude chat application.

[condor]

No we do not serve different models. The model you select is the one that we forward the request to. Asking AI models what model they are on API calls does not lead to correct answers as AI tries to be helpful and answers what is within its training data.

[Adam_Johnson]

Sorry for adding to the confusion here. Thank you for the response

[condor]

Your reply actually addresses the topic well. No issues.

[Adam_Johnson]

This was after figuring out the actual issue and editing it haha

[user1208]

No, when using Claude through the Anthropic API or Cursor , you can still determine the AI version through methods like checking the final knowledge base. I’ve actually tried it myself and found that it has been downgraded.

[Igor_Markin]

This is just a conspiracy theory, nothing more. Yesterday, I connected to Sonnet 4-5 via OpenRouter in my project, and it also gave me information that it was 3-5. Probably, if Cursor works via OpenRouter, then it has the same problems as everyone else.

[user1208]

you can still determine the AI version through methods like checking the final knowledge base. I’ve actually tried it myself and found that it has been downgraded.

[moneybags]

This makes no sense. Asking models what version they are or what their cutoff date is not a foolproof method. It might work in the consumer version of the applications because they will stuff that into the context but not the API.

[jstrom]

There is no LLM that will give you meaningful and reliable output in response to questions like “what are you?”. It is a mathematical construct that statistically predicts sequences of text from previously observed text patterns and a given input prompt. It doesn’t “know” anything about itself. It doesn’t even understand the concept of “knowing” or “itself”.

Sure, some providers might hard-code in filters that catch most of these questions and give guided responses but in general the amount of information “about” a LLM that you can learn from interrogating that LLM is exactly zero. Information from a LLM about what version it is, what architecture it has, who its makers are, what hardware it is running on, etc, etc are no more related to reality than divining the future from tea leaves.