PR #23742: fix(anthropic): strip vector_store_ids from request payload to prevent 400 error

• Repository: BerriAI/litellm
• Author: xykong
• State: open | merged: False
• Link: https://github.com/BerriAI/litellm/pull/23742

Description (problem / solution / changelog)

Summary

• Strip vector_store_ids and vector_store_id from optional_params in AnthropicConfig.transform_request() before building the final request payload
• These are OpenAI Assistants API parameters injected by LiteLLM proxy for internal RBAC/vector-store hooks — Anthropic rejects them with 400: Extra inputs are not permitted
• Adds 2 unit tests covering both the stripping behavior and the no-op case

Problem

AnthropicConfig.transform_request() builds the Anthropic payload via **optional_params. LiteLLM proxy injects vector_store_ids into the request body for internal pre-call hooks (proxy/auth/auth_checks.py, http_parsing_utils.py). This parameter leaks through to Anthropic, causing:

AnthropicException - {"type":"error","error":{"type":"invalid_request_error","message":"vector_store_ids: Extra inputs are not permitted"}}

is_vertex_request was already being stripped at this same location — this PR extends that pattern to cover vector_store_ids and vector_store_id.

Changes

• litellm/llms/anthropic/chat/transformation.py: pop vector_store_ids and vector_store_id before data = {..., **optional_params}
• tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py: 2 new mock-only tests

Related

Fixes #23741

Pre-Submission checklist

• I have Added testing in the tests/litellm/ directory
• My PR's scope is as isolated as possible, it only solves 1 specific problem

Changed files

• litellm/llms/anthropic/chat/transformation.py (modified, +2/-0)
• tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py (modified, +40/-0)

PR #23974: fix(ui): stop injecting empty vector_store_ids into model litellm_params on save

• Repository: BerriAI/litellm
• Author: dkindlund
• State: open | merged: False
• Link: https://github.com/BerriAI/litellm/pull/23974

Description (problem / solution / changelog)

Summary

• Admin UI model edit form unconditionally injects empty/zero values for vector_store_ids, guardrails, input_cost_per_token, and output_cost_per_token into the update payload, even when the user hasn't touched those fields
• Fix: apply conditional guards — only include these fields when explicitly set, otherwise delete them from the payload

Problem

The handleSaveChanges function in model_info_view.tsx always includes these fields in the update payload:

// Before fix:
input_cost_per_token: values.input_cost / 1_000_000,   // 0/1M = 0.0 when empty
output_cost_per_token: values.output_cost / 1_000_000,  // 0/1M = 0.0 when empty

When a user edits a model to change only the credential (or any other field), the empty cost fields produce input_cost_per_token: 0.0 and output_cost_per_token: 0.0, which:

1. Overrides the built-in pricing table with custom $0 pricing
2. Silently disables spend tracking for the model
3. Bypasses budget enforcement (spend never increments)
4. Causes LiteLLM's reported spend to diverge from cloud provider billing

On one production deployment, this caused ~$6,500 of untracked Claude spend over 14 days before the root cause was identified.

The same pattern applies to vector_store_ids: [] (causes Anthropic 400 errors) and guardrails: [] (clears configured guardrails).

Fix

Apply the same guard pattern already used for litellm_credential_name:

// After fix:
if (values.input_cost != null && values.input_cost !== "") {
  updatedLitellmParams.input_cost_per_token = values.input_cost / 1_000_000;
} else {
  delete updatedLitellmParams.input_cost_per_token;
}

Test plan

• should not inject empty vector_store_ids into litellm_params on save
• should preserve non-empty vector_store_ids in litellm_params on save
• should not inject empty vector_store_ids when model previously had empty array
• should remove vector_store_ids when user clears all vector stores
• should not inject empty guardrails into litellm_params on save
• should preserve non-empty guardrails in litellm_params on save
• should remove guardrails when user clears all guardrails
• should not inject input_cost_per_token: 0 when cost fields are empty on save
• should preserve explicitly set cost values on save

Generated with Claude Code

Changed files

• ui/litellm-dashboard/src/components/model_info_view.test.tsx (modified, +423/-1)
• ui/litellm-dashboard/src/components/model_info_view.tsx (modified, +21/-11)