[BUG] Unexpected session model tokens usage for opus

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Seems that when you use a session with opus-A and running the commands using opus-B they have some complicated relations. Context copying or something. But you are paying for it.
My case was:

When running session on opus47, command on sonnet and subagents on haiku:

Total cost:            $0.96
  Total duration (API):  6m 48s
  Total duration (wall): 4m 4s
  Total code changes:    118 lines added, 0 lines removed
  Usage by model:
     claude-sonnet-4-6:  366 input, 14.0k output, 638.6k cache read, 48.4k cache write ($0.58)
      claude-haiku-4-5:  265 input, 18.1k output, 741.4k cache read, 165.2k cache write ($0.3712)

And this is what I expect to see ^

When running session on opus47, command on opus46, and subagents on haiku:

Total cost:            $2.08
  Total duration (API):  7m 51s
  Total duration (wall): 3m 34s
  Total code changes:    120 lines added, 0 lines removed
  Usage by model:
       claude-opus-4-6:  1.5k input, 6.7k output, 295.6k cache read, 35.5k cache write ($0.54)
      claude-haiku-4-5:  266 input, 19.6k output, 746.1k cache read, 168.3k cache write ($0.3830)
       claude-opus-4-7:  532 input, 8.9k output, 1.0m cache read, 65.9k cache write ($1.15)

basically command run is the most expensive thing here.

When running session on opus46, command on opus47, and subagents on haiku:

Total cost:            $1.58
  Total duration (API):  5m 17s
  Total duration (wall): 2m 27s
  Total code changes:    121 lines added, 0 lines removed
  Usage by model:
       claude-opus-4-7:  528 input, 2.7k output, 108.1k cache read, 72.0k cache write ($0.57)
      claude-haiku-4-5:  253 input, 18.9k output, 715.9k cache read, 160.5k cache write ($0.3671)
       claude-opus-4-6:  356 input, 4.9k output, 447.8k cache read, 46.9k cache write ($0.64)

Same thing here, just cheaper.

This should be fixed or properly explained in the documentation.

What Should Happen?

I should see usages by command and subagents models, not by session model.

Error Messages/Logs

no

Steps to Reproduce

• make a command running subagents and specify opus-A
• make subagent using haiku
• start a session, switch to opus-B
• run the command
• cost shows strange numbers for the session model

Claude Model

Not sure / Multiple models

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.143

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

iTerm2

Additional Information

_No response_