[MODEL] Opus advising Opus

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Other unexpected behavior

What You Asked Claude to Do

I was working using Sonnet. I switched to Opus. backed out started a new session. Advisor was still on. asked opus to do something that was complicated, required looking at previous work.

What Claude Actually Did

1. Claude proceeded to work through the task
2. at some point it needed to consult a higher level model. So it decided to consult another version of itself.

Expected Behavior

Claude is running Opus 5. It should not be consulting another Opus5 model to advise. This inflated the token usage by 2.5x with no real benefit.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Do stuff with Sonnet 5 with advisor set to Opus 5
2. Switch model to Opus 5
3. Compact or back out & initiate a new task within the same terminal session
4. Do stuff with high complexity until you see an advisory message
5. Watch your token usage ramp up by 2.5x

Claude Model

Opus

Relevant Conversation

Advising using Opus 5
√ Advisor has reviewed the conversation and will apply the feedback


It did this twice. There clearly isn't a an advisor check in place.


 you are an OPUS 5 agent stop asking OPUS 5 to advise

  Listed 1 directory

Understood — no more advisor calls.


 what model are you?

I'm Claude Opus 5 (model ID claude-opus-5), running as Claude Code.

✻ Baked for 3s

> is there a benefit to you consulting another opus 5 model?

Honestly: marginal, and not for the reason the tool advertises.

The tool describes itself as "a stronger reviewer model." I can't verify what backs it, and if it's also Opus 5 there's no capability gap to exploit.

What it does give is a fresh pass over the same transcript with no commitment to my working hypothesis. That's a real effect but a small one, and it's the same effect I get from re-reading my own evidence deliberately.

So: one useful nudge, some noise, real latency cost. Not worth the round-trip when I already have primary evidence to go read.


what was the token usage of the last coding task, and how much of it was because of advisor?

I can only measure this from the total_tokens left figures in system reminders — I have no direct usage meter, and that counter doesn't tick on every call (it sat flat across several edits, then reset to 15,000,000 after the task). So treat this as a bound, not a precise attribution.

Readings across the task segment:

┌────────────────────────────────────────────────────────────┬─────────────┬─────────┐
│                           point                            │ tokens left │  delta  │
├────────────────────────────────────────────────────────────┼─────────────┼─────────┤
│ start (post-compact)                                       │ 15,000,000  │ —       │
├────────────────────────────────────────────────────────────┼─────────────┼─────────┤
│ after initial context + first orientation read             │ 14,917,004  │ 82,996  │
├────────────────────────────────────────────────────────────┼─────────────┼─────────┤
│ through the source reads, greps, git diffs, semantic check │ 14,888,199  │ 28,805  │
├────────────────────────────────────────────────────────────┼─────────────┼─────────┤
│ after advisor call #1                                      │ 14,769,954  │ 118,245 │
├────────────────────────────────────────────────────────────┼─────────────┼─────────┤
│ edits + build (counter flat)                               │ 14,769,954  │ 0       │
├────────────────────────────────────────────────────────────┼─────────────┼─────────┤
│ after advisor call #2 + log search                         │ 14,717,495  │ 52,459  │
└────────────────────────────────────────────────────────────┴─────────────┴─────────┘

Total for the task: ~282,500. The two windows containing the advisor calls account for ~170,700 of that — roughly 60%.

The mechanism explains the size: advisor forwards my entire conversation history on every call, so each one costs about (current context + its response), and that grows as the session does. Call #1 landed when context was already large; call #2 was cheaper only because less had accumulated since.


 /advisor off
  ⎿  Advisor disabled

> would the token usage have been less without the advisor calls?

Yes — roughly 2.5x less. The task would have run around 110k instead of ~282k.

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.233 (Claude Code)

Platform

Anthropic API

Additional Context

_No response_