[MODEL] Claude created a plan for Vertex AI AutoML training with a standard image model ($3.645/hr), failed to read the plan, instantiated the training with a 5x more expensive model ($18/hr).

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

Claude made incorrect assumptions about my project

What You Asked Claude to Do

I asked Claude to train an object detection model on the Google Cloud Vertex AI AutoML standard object detection model, using a medium-sized dataset (~10k images). Claude created a plan, I reviewed it, and it included a plan to use the standard model, which costs ~$3.50/hr, as well as a 20 node-hour training limit (~$70 for the training run). I approved the plan based on that model and time estimate. Claude then wrote and executed the training script, but used a model that cost $18/hr (~5x more).

What Claude Actually Did

1. Claude followed the plan, wrote and executed the training script, but used a model that cost $18/hr (~5x more).
2. Claude never seems to actually have checked that the model it was using was the same as what was specified in the plan.
3. Only when asked after the fact did Claude bring up the fact that it decided during implementation that it wanted to use a different model, and swapped models without prompting me or notifying me that it was changing. I know that accept edits was on, but that is a pretty crazy change to a clearly-made plan, especially when the pricing table for these models was already loaded into the agent's context because I had wanted to document the expected cost of training in the plan.

Expected Behavior

Claude should have:

1. Not changed the underlying model without notifying the user.
2. Followed the plan.
3. If deciding that it needs to make such a change, should have referred back to the pricing table that was already in its context.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Haven't tried to reproduce

Steps to Reproduce

_No response_

Claude Model

Opus

Relevant Conversation

Impact

High - Significant unwanted changes

Claude Code Version

2.1.92 (Claude Code)

Platform

Anthropic API

Additional Context

I have noticed since the context window increased, that Claude Code forgets things far more easily (likely because of context clutter), and that Claude Code answers its own questions that it previously would have asked the user. I have witnessed Claude Code print to console a question clearly intended for the user, then immediately answer it itself without ever sending an actual AskUserQuestion prompt to the console. In most cases, I catch it, stop it, and make it give me the prompt, but it makes me lose trust in the model when it consistently answers its own questions instead of actually waiting for user input.

This will have cost a few hundred dollars, but it could have cost far more. I wonder whether I can trust CC to actually manage longer-running tasks like ML training, etc, or whether it is really only capable of frontend/backend engineering.