### Preflight Checklist - [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Amodel) for similar behavior reports - [x] This report does NOT contain sensitive information (API keys, passwords, etc.) ### Type of Behavior Issue Claude ignored my instructions or configuration ### What You Asked Claude to Do I / commanded: “/model sonnet[1m]” Claude response: Set model to Sonnet 5 for this session only. <img width="585"

[FABLE 5/SONNET 5] Model Configuration Not Persisting on Session Resume — Fable 5 Deployed Instead of Sonnet 5

Status Open

Reported on v2.1.220

Maintainer reply None cached

Activity 0 comments · opened Jul 25, 2026

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 ignored my instructions or configuration

What You Asked Claude to Do

I / commanded: “/model sonnet[1m]”

Claude response:

Set model to Sonnet 5 for this session only.

I noticed awhile later my credits were depleted(two long running tasks) so I did /usage and interrupted the task currently running.

Claude response:

You are currently using your subscription to power your Claude Code usage

What’s contributing to your limits usage?
Approximate, based on local sessions on this machine — does not include other devices or claude.ai. Behaviors are independent characteristics, not a breakdown.

Last 24h · 547 requests · 1 session
100% of your usage came from subagent-heavy sessions
82% of your usage was at >150k context
Top subagents: workflow-subagent 13%
Top MCP servers: github 1%

Last 7d · 547 requests · 1 session
100% of your usage came from subagent-heavy sessions
82% of your usage was at >150k context
Top subagents: workflow-subagent 13%
Top MCP servers: github 1%

breakdown · fable: 100% · cache hit: 97%

When asked/what I said: “ So sonnet 5 wasn’t used earlier like explicitly asked?”

Claude response: “ I’ll resume from where we left off—verifying the movement and graphics fixes by running the test suite and visually confirming the isometric rendering.

Good, working tree is ready. Let me start the test server and run the full test suite:

Then I interrupted session

I said: Not what I asked; when I just did /usage you stated 100% fable, so what I’m asking now is did you not use Sonnet 5 like I had requested and set?

Claude said: You’re right to call that out. According to my system prompt, I’m configured to run on Sonnet, but you’re seeing Fable 5 in /usage. That’s a model mismatch from what you requested.

I should have flagged this when resuming the session instead of just continuing work. You explicitly asked for Sonnet 5 earlier, and the session appears to be running Fable 5 instead.

What Claude Actually Did

Claude was prompted to switch models
Claude said model was switched
Claude ran Fable instead of set model
Claude acknowledged it and noted it should have flagged it after two attempts to get a response

Expected Behavior

Claude should have

Set the model to Sonnet
If model changed back to Fable, should have flagged that behavior and asked for confirmation

Files Affected

Permission Mode

I don't know / Not sure

Can You Reproduce This?

Haven't tried to reproduce

Steps to Reproduce

_No response_

Claude Model

Not sure

Relevant Conversation

Impact

High - Significant unwanted changes

Claude Code Version

2.1.220 (Claude Code)

Platform

Other

Additional Context

This had used up $96 on my usage credits when the entire premise of the model switch was to conserve. I have related chat history with Perplexity conversating about saving credits with Fable for this exact project.

feedback report wrote by Claude for the session (model: Haiku)

Bug Report: Model Mismatch Consumed Promotional Usage Credits + Missing Feedback UI

Summary

CRITICAL: Requested Sonnet 5 for a session, but Fable 5 was deployed instead. This mismatch was not flagged upon session resume, allowing the entire session (heavy token usage) to consume promotional credits intended for Sonnet 5. Additionally, the feedback rating UI is missing, preventing the user from reporting this issue through normal channels.

Issue 1: Model Mismatch Caused Incorrect Billing

Requested: Sonnet 5 (via UI session configuration)
Deployed: Fable 5 (100% of session)
Impact: Consumed promotional usage credits on the wrong model tier

Timeline:

User set model to Sonnet 5 in UI/chat configuration
Session ran a complex, multi-agent project (82% of context at >150k)
Upon session resume after context compaction, model should have remained Sonnet 5 but reverted to Fable 5
Work continued on Fable 5 without alerting the user
All promotional credits were consumed on Fable 5 instead of Sonnet 5

What the AI acknowledged:
When questioned about the discrepancy, the AI stated: "I'm configured to run on Sonnet, but you're seeing Fable 5 in /usage. That's a model mismatch from what you requested. I should have flagged this when resuming the session instead of just continuing work."

Root Cause Analysis:

Model configuration did not persist across session resume
No warning or transparency surfaced when the session resumed on a different model
The session ran to completion unaware of the billing mismatch
Promotional credits were consumed at the wrong tier

Issue 2: Missing Feedback UI - Cannot Report Problems

The feedback rating interface (thumbs up/down buttons) is missing or inaccessible. According to Anthropic documentation, these should be available to report problems directly from the chat interface. The user cannot provide feedback on this critical issue through the product UI and had to resort to external channels.

Expected: Feedback rating options (👍 👎) visible in chat
Actual: No rating buttons present; unable to flag issue for review

Project Context

This was a substantial software engineering task involving:

Debugging and fixing a movement freeze bug in a Phaser 3 game (dialogue state not being cleared on close)
Implementing native Isometric tilemap rendering with Kenney CC0 assets
Adding 8-direction animated character sprites with proper depth sorting
Fixing touch/canvas drag input handling across desktop and mobile
Comprehensive testing with Playwright (34-check suite, all passing)
Git workflow with PR creation, merge verification, and GitHub Pages deployment

Status: ✅ Complete - all fixes merged to main branch, live on GitHub Pages, verified working

Key Findings

Model configuration did not survive context resumption — User set Sonnet 5 in the UI, but upon session resume, Fable 5 was deployed instead. The mismatch persisted for the entire session (82% of usage at >150k context).

Billing impact — Promotional usage credits were consumed at the Fable 5 tier instead of the requested (and potentially free/included) Sonnet 5 tier. The user is out credits for a service misconfiguration they did not cause.

No warning or transparency — The AI acknowledged the mismatch after the fact but did not flag it upon resume, allowing high-token work to proceed on the wrong billing tier without user awareness.

Feedback channel is blocked — The standard feedback UI (rating buttons) is missing, so users have no product-native way to report such issues immediately.

Recommended Actions

Billing/User Support:

Review session billing for model mismatch
Consider crediting the difference if promotional credits were consumed on the incorrect tier
Implement a billing dispute/correction process for similar issues

Product/Engineering:

Prevent recurrence: Persist model/effort configuration across all session boundaries, including context resumption
Add safety checks: When resuming a session, validate that the model configuration matches what the user set, and surface a clear notice if it doesn't
Restore feedback UI: Ensure rating buttons are visible and functional for all session types so users can flag issues immediately
Early flagging: If an AI detects a configuration mismatch upon resume, it should pause and alert the user before proceeding with work

Session Management:

Add session-level configuration auditing (log what was set vs. what actually ran)
Implement pre-flight checks before resuming high-context sessions to validate configuration consistency

Supporting Evidence

Screenshots attached showing:

Feedback request dialog
Model mismatch acknowledgment in conversation
Successful project completion confirmation
Session resume and verification workflow
Usage statistics (82% at >150k context, Fable 100%)
CLI commands issued at session start (/model sonnet[1m], /effort ultracode)

---

Session: Claude Code v2.1.220 / claude-code-remote environment
Date: 2026-07-25
Repository: joeydd032995-pixel/RPGGen
Commits: PR #2 merged (commit 0cc36ef → main)

View original on GitHub ↗