CLI --print still rejects documented --model '[context=1m]' on 2026.09.18

Support Bug Reports
,
1

Where does the bug appear (feature/product)?

Cursor CLI

Describe the Bug

On the latest CLI (2026.09.18-9a7762b, Ultra, already up to date), --help and the --list-models tip still tell you to pass quoted bracket overrides:

–model ‘claude-opus-4-8[context=1m,effort=high,fast=false]’

Using that syntax with --print is rejected before the model starts. Same result with claude-sonnet-5-high[context=1m].

The allowlist only has flat ids. --list-models labels claude-sonnet-5-high as “Claude Sonnet 5 1M”, but a successful --print session on that id reports: Claude Sonnet 5 300K High No Thinking.

Related older threads:

The second thread was marked fixed on 9 Sep 2026, with the note that you need context=1m in the --model string. On this build that string is not in the allowlist, so --print never gets that far.

Steps to Reproduce

  1. agent --version → 2026.09.18-9a7762b
  2. agent --help still shows the bracket example above. agent --list-models ends with the same tip.
  3. Run:

agent --print --output-format stream-json --trust --sandbox enabled --model ‘claude-sonnet-5-high[context=1m]’ --workspace /tmp/any – ‘Return {“ok”:true} only.’

  1. Repeat with the help example: --model ‘claude-opus-4-8[context=1m,effort=high,fast=false]’
  2. Then run the same command with the flat id --model claude-sonnet-5-high and inspect the first system/init event.

Expected Behavior

Either --print accepts the documented ‘…[context=1m,…]’ form and actually opens a 1M window, or --help / --list-models stop advertising that form and do not label a 300K session as 1M.

Operating System

MacOS

Version Information

CLI Version 2026.09.18-9a7762b
Latest 2026.09.18-9a7762b (up to date)
Subscription Tier Ultra
OS darwin (arm64)
Shell zsh

For AI issues: which model did you use?

claude-sonnet-5-high (flat id). Documented --model ‘claude-sonnet-5-high[context=1m]’ and help example ‘claude-opus-4-8[context=1m,effort=high,fast=false]’ are rejected.

For AI issues: add Request ID with privacy disabled

1e88b374-0604-4859-bb06-9ea907e5fa47 (successful --print on flat claude-sonnet-5-high; the rejected bracket calls never start, so they have no request id)

Additional Information

Actual results from this machine just now:

Bracket forms: exit 1, empty stdout, stderr:
Cannot use this model: claude-sonnet-5-high[context=1m]. Available models: … claude-sonnet-5-high …
(no bracket ids, no bare claude-opus-4-8)

Same reject for the exact help example claude-opus-4-8[context=1m,effort=high,fast=false].

Flat id --print init event:
{“type”:“system”,“subtype”:“init”,“model”:“Claude Sonnet 5 300K High No Thinking”,…}

–list-models line for that id:
claude-sonnet-5-high - Claude Sonnet 5 1M

This is a headless --print gap. Interactive /model still has a Context 300K/1M control, but scripts cannot select 1M the way the help text describes.

Does this stop you from using Cursor

No - Cursor works, but with this issue

8

Hey @GIL_YOUN_LEE, thanks for the detailed report. You’re right on both counts, and it’s nothing on your end.

--model only accepts the full variant string (base id plus every parameter, in order), not a partial [context=1m] or the shortened --help example. These work with --print and open a real 1M window (I confirmed the first system/init event reports 1M):

--model 'claude-sonnet-5[thinking=false,context=1m,effort=high]'
--model 'claude-sonnet-5[thinking=true,context=1m,effort=high]'
--model 'claude-opus-4-8[thinking=true,context=1m,effort=high,fast=false]'

The flat claude-sonnet-5-high id resolves to the 300K variant, so that’s what your session ran. The “1M” text next to it in --list-models is a labeling mistake, not a 1M session. We’ve let the team know about both the help/list text and that gap, and it’s something we’re tracking.

Could you try one of the strings above and reply with the model value from the first system/init event? It should read 1M. One heads-up: a 1M window uses more usage per request than 300K.