Does anyone know what the “Compass” model is?

Support Help
1

Ever since Grok stopped being discounted, I’ve been using Cursor’s Auto mode for my daily work.

Today, I checked my Usage page and noticed that 7% of my API quota had been used by a model called accounts/anysphere/models/compass-v4-step601#anysphere/global-big-char. I’ve never heard of this model, and I couldn’t find any information about it online.

Does anyone know what Compass is and why it showed up in my usage? Is it a model that Cursor’s Auto mode selects behind the scenes?

image
image1106×431 10.4 KB

4

Same here. I’ve been using auto but getting charged to my API usage with this model.

I want a refund on the API usage because I confirmed in my billing it says I had auto selected even though it’s using this compass-v4 model and charging me for API

5

Same here, I’m using “auto” all day, but this guy just showed up and used my api quota instead

image
image695×245 21.6 KB

image
image1317×343 11.4 KB

6

Same here… I have also been stuck with this problem..

image
image1080×668 21.5 KB

Here is what i got when i asked Cursor to do the analysis by itself…
“”"
That string is not a model you picked. It’s an internal Cursor/Anysphere model that Auto routes to behind the scenes.

What is compass-v4-step601#anysphere/global-big-char?

Breaking down the name:

Part Likely meaning
anysphere Cursor’s company (Anysphere)
compass-v4 Internal codename — “Compass” model, version 4
step601 Training checkpoint (step 601)
global-big-char Internal variant/config (likely large-context)

Compass is not in the public model picker. Cursor hasn’t documented it. It only shows up in usage logs when Auto (or the agent harness) routes to it internally.

You’re not alone — a forum thread from July 30 has multiple users reporting the same thing: using Auto all day, then seeing Compass eat Other Models quota.

What your numbers actually show

Your Auto isn’t just using one backend. It’s fanning out to several:

Cursor Models pool:

  • auto — 580M tokens (main Auto usage)
  • compass-v4 — 58M tokens

Other Models pool (the problem):

  • compass-v4 — 47M tokens (18% of your Other Models!)
  • claude-sonnet-5-thinking-medium — 63M (17%)
  • claude-4.6-opus-medium-thinking — 6M
  • gpt-5.5-low — 4M

So even with Auto selected, Cursor is routing to Compass + Claude + GPT — and billing those from Other Models. That explains the 24% → 46% jump.

Compass alone is ~105M tokens total across both pools, and nearly half of its usage hit Other Models — which you never intended to use.

Why this is frustrating (and arguably wrong)

You selected Auto, which is supposed to use bundled/Cursor Models pricing. Instead:

  1. Auto silently picks Compass (undocumented internal model)
  2. Auto also picks Claude Sonnet 5, Opus, GPT-5.5
  3. Those hit your $70 Other Models pool
  4. You get no UI control (no Optimize For on Pro+)
  5. The model name in usage is an opaque internal ID

Forum users are already asking for refunds on API usage charged while Auto was selected.

What to do

Immediately:

  1. Turn off Auto — use Composer 2.5 explicitly for daily work
  2. Disable on-demand spending (or set limit to $0)

Contact support — you have a strong case. Email [email protected]… They will look into it.
“”"