Cursor CLI `agent --model` ignores reasoning/thinking level suffix

Support Bug Reports
, , ,
1

Where does the bug appear (feature/product)?

Cursor CLI

Describe the Bug

Description

In the updated Cursor CLI, passing a model with an explicit reasoning or thinking level appears to preserve only the base model and drops the requested reasoning level.

For example:

agent --model claude-opus-4-7-thinking-high
agent --model gpt-5.5-extra-high

The CLI starts with the correct base model, such as claude-opus-4.7 or gpt-5.5, but the reasoning/thinking level is not applied. It defaults to medium instead of using thinking-high or extra-high.

System Message / Environment

  • OS: Linux 6.1.158-180.294.amzn2023.x86_64
  • Shell: zsh
  • Cursor CLI / agent version: 2026.05.04-08e5280
  • Cursor CLI command tested:
agent --model claude-opus-4-7-thinking-high
agent --model gpt-5.5-extra-high
  • Observed system behavior: the launched agent reports or behaves as the requested base model, but the reasoning level is medium rather than the requested high or extra-high setting.

Expected Behavior

The CLI should respect the full model identifier, including the requested thinking/reasoning level:

  • claude-opus-4-7-thinking-high should use Opus 4.7 with high thinking.
  • gpt-5.5-extra-high should use GPT-5.5 with extra-high reasoning.

Actual Behavior

The CLI only applies the base model type and ignores the reasoning level suffix, defaulting the session to default reasoning levels.

Impact

This makes it impossible to launch high-reasoning CLI agents directly via agent --model, even when the model string explicitly includes the desired thinking level.

Repro Steps

  1. Run:

    agent --model claude-opus-4-7-thinking-high

  2. Observe that the session uses Claude Opus 4.7, but the thinking level is the default one.

  3. Run:

    agent --model gpt-5.5-extra-high

  4. Observe that the session uses GPT-5.5, but the reasoning level is medium (the default one after the update).

Suspected Cause

The updated CLI model parser appears to parse only the base model family/version and drops the reasoning/thinking level component from the model identifier.

Steps to Reproduce

Repro Steps

  1. Run:

    agent --model claude-opus-4-7-thinking-high

  2. Observe that the session uses Claude Opus 4.7, but the thinking level is the default one.

  3. Run:

    agent --model gpt-5.5-extra-high

  4. Observe that the session uses GPT-5.5, but the reasoning level is medium (the default one after the update).

Operating System

Linux

Version Information

  • Cursor CLI / agent version: 2026.05.04-08e5280

Does this stop you from using Cursor

No - Cursor works, but with this issue

7

This is a known bug with the CLI’s model selection for parameterized variants. The --model flag and /model command don’t correctly handle reasoning level suffixes like thinking-high or extra-high, so the CLI silently falls back to the default level.

Our CLI team is tracking this and working on a fix. No ETA to share yet.

In the meantime, Auto mode in the CLI will route to a strong model configuration automatically. You can also use the full model picker with all reasoning levels in the Cursor desktop app (IDE), where these variants work as expected.

You may also find this related thread helpful: GPT-5.5 extra high not an option in Cursor CLI

8

Thanks Mohit. This bug becomes a hard blocker for me today. I have pipelines that use agent. While I can set the default ones to be my preferred ones, but EACH CLI update (e.g., 5-5’s update) resets the default one to medium with low context window. This breaks my pipelines entirely. Could you please share the priority of this bug fixing? Or is there a way we can roll back to previous versions?

9

We can’t share a specific ETA on the fix, but your feedback about the pipeline impact is noted and has been added to the internal tracking issue.

Two things that may help stabilize your setup in the meantime:

  1. Prevent CLI auto-updates from resetting your config: You can pass --disable-auto-update when launching the CLI to stop it from auto-updating mid-session. This should keep your selected defaults stable until you choose to update manually.

  2. For the underlying bug where --model drops the reasoning suffix, the workaround for now is to set your preferred model interactively via /model at the start of a session. Combined with --disable-auto-update, this should keep your pipelines consistent until the fix ships.