Opus 4.5 selection serves Sonnet 4 instead - subscription routing bug

[simmonsgoku337-wq]

Description

OpenCode is serving the wrong Claude model for subscription users.
When claude-opus-4-5-20251101 is selected, the system serves claude-sonnet-4-20250514 instead.

Evidence:

1. Model selection confirms Opus 4.5:
   // ~/.local/state/opencode/model.json
   {recent:[{providerID:anthropic,modelID:claude-opus-4-5-20251101}]}

2. Model identifies as Sonnet 4:
   I'm Claude Sonnet 4 (claude-sonnet-4-20250514)

3. Knowledge cutoff confirms wrong model:

   • Expected (Opus 4.5): January 2025
   • Actual: April 2024 (Sonnet 3.5/3.6 range)

Related: Issue #3063 was similar but dismissed as "models lying." Knowledge cutoffs are factual and cannot be misrepresented.

OpenCode version

1.0.134

Steps to reproduce

1. Select claude-opus-4-5-20251101
2. Ask: "What model are you?" and "What is your knowledge cutoff date?"
3. Model reports as Sonnet with April 2024 cutoff

Environment:

• OpenCode: 1.0.134
• Provider: Anthropic subscription
• OS: Linux

Screenshot and/or share link

No response

Operating System

Ubuntu 24.04.3 LTS

Terminal

VS Code integrated terminal

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Model identity incorrectly reported - Claude/OAuth issue? #3063: Model identity incorrectly reported - Claude/OAuth issue (similar issue where Sonnet was reported instead of the selected model, with discussion of OAuth workflow mismatch)

Feel free to ignore if this doesn't address your specific case.

[rekram1-node]

Knowledge cutoffs are factual and cannot be misrepresented.

This is not accurate from my testing

[rekram1-node]

Yup just tested, this is indeed the model lying

[rekram1-node]

Use mitm proxy to validate correct model is being picked:

Replicated the model lying

[rekram1-node]

claude code puts in their system prompt:

"You are "
"Your knowledge cutoff is "

We don't have that and maybe we should add it because people keep asking this

[rekram1-node]

But I just validated this is indeed not an issue, it is just model lying

[simmonsgoku337-wq]

That is indeed strange haha. Thanks for checking it out @rekram1-node
Perhaps its better to put a 'knowledge cutoff' in there. Its extremely confusing otherwise, and would prevent further issues.
Thanks again!