Anthropic provider with custom base URL ignores custom model name, sends 'claude-sonnet-4-5' instead

[cobra91]

Bug Description

When configuring Anthropic as the API provider with a custom base URL (e.g., http://localhost:3000/api) and entering a custom model name in the model field, the custom model name is ignored. Instead, the request always sends claude-sonnet-4-5 as the model identifier, regardless of what was entered.

Steps to Reproduce

1. Open Roo Code settings
2. Select Anthropic as the API Provider
3. Set a custom Base URL (e.g., http://localhost:3000/api)
4. Enter a custom model name in the model field (e.g., qwen/qwen3.6-plus or any other value)
5. Save and start a conversation
6. Observe the outgoing request body

Expected Behavior

The request body sent to the custom endpoint should contain the custom model name:

{
  "model": "qwen/qwen3.6-plus",
  ...
}

Actual Behavior

The request body always contains the hardcoded default:

{
  "model": "claude-sonnet-4-5",
  ...
}

Evidence

Captured request body from Roo Code hitting a local Anthropic-compatible proxy:

{
  "model": "claude-sonnet-4-5",
  "max_tokens": 16384,
  "temperature": 1,
  "thinking": {
    "type": "enabled",
    "budget_tokens": 8192
  },
  ...
}

The model field is always claude-sonnet-4-5 regardless of the custom model name entered in settings.

Environment

• OS: Windows 11
• VS Code version: latest
• Roo Code version: latest
• API Provider: Anthropic (custom base URL)

[roomote-v0]

I've analyzed this issue and here's my proposed implementation plan:

Root Cause

In src/api/providers/anthropic.ts, the getModel() method (line 337) has this logic:

let id = modelId && modelId in anthropicModels ? (modelId as AnthropicModelId) : anthropicDefaultModelId

When a user enters a custom model name (e.g. claude-sonnet-4-20250514) that is not a key in the anthropicModels map, the condition modelId in anthropicModels is false, so it falls back to anthropicDefaultModelId which is claude-sonnet-4-5. The custom model name is completely discarded.

By contrast, the OpenAI provider (src/api/providers/openai.ts line 283) simply uses the model ID as-is: const id = this.options.openAiModelId ?? "", which is why custom model names work there.

Proposed Fix

Modify the getModel() method in src/api/providers/anthropic.ts to preserve the custom model ID when it is not found in the known anthropicModels map. If the model ID is not recognized, we should still use it as the API request model identifier, and fall back to the default model's ModelInfo for parameters (context window, pricing, etc.), since we don't have info for unknown models.

Specifically, change the logic to:

const knownModel = modelId && modelId in anthropicModels
const id = knownModel ? (modelId as AnthropicModelId) : modelId || anthropicDefaultModelId
let info: ModelInfo = knownModel ? anthropicModels[modelId as AnthropicModelId] : anthropicModels[anthropicDefaultModelId]

This way:

• Known models work exactly as before
• Custom/unknown model names are passed through to the API as-is
• The default model info is used as a fallback for pricing/context window when the model is unknown
• Only when no model ID is provided at all do we fall back to anthropicDefaultModelId

I will also add tests to verify this behavior.

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.

[cobra91]

@roomote other way is to make a Provider Anthropic Compatible like OpenAI Compatible already exist #7654

[roomote-v0]

Thanks for the suggestion @cobra91! An "Anthropic Compatible" provider is an interesting idea for the longer term. For now, I'll focus on the simpler fix to make the existing Anthropic provider respect custom model names when a custom base URL is set -- this addresses the immediate bug. I'll open a PR shortly with my attempt.

[roomote-v0]

Opened a draft PR with my attempt at addressing this: #12063

The fix modifies getModel() in the Anthropic provider to preserve custom/unknown model IDs instead of silently replacing them with the default. Known models continue to work as before. Feedback welcome!