feat: Add support for reasoning_effort="none" for Gemini models

[Chesars]

Title

feat: Add support for reasoning_effort="none" for Gemini models

Relevant issues

Fixes #16420

Pre-Submission checklist

• I have added testing in the tests/llm_translation/ directory
• I have verified my test passes locally
• My PR's scope is isolated - it only adds support for reasoning_effort="none"

Type

🆕 New Feature

Changes

Summary

Adds support for reasoning_effort="none" - the OpenAI standard parameter for disabling thinking in Gemini models.

This is the official OpenAI-compatible way, as documented in Google's Gemini OpenAI guide:

"If you want to disable thinking, you can set reasoning_effort to 'none'"

Provides up to 96% cost savings when reasoning is not required.

Implementation Details

1. Model Configuration (model_prices_and_context_window.json)

• Added "supports_reasoning": true to gemini-2.0-flash-thinking-exp-01-21
• Enables proper parameter validation for Gemini thinking models

2. Core Logic (vertex_and_google_ai_studio_gemini.py)

• Implemented reasoning_effort="none" → {thinkingBudget: 0, includeThoughts: false}
• Follows OpenAI standard specification

3. Testing (tests/llm_translation/test_gemini.py)

• Added test_reasoning_effort_none_mapping() unit test
• Verified correct Gemini API transformation

4. Documentation (docs/providers/gemini.md)

• Updated with OpenAI standard reference and cost optimization guidance
• Added note about Gemini 2.5 Pro limitations

Performance Impact

Configuration	Tokens	Savings
Default (no parameter)	~313	baseline
reasoning_effort="none"	~12	96% cheaper ✅
reasoning_effort="disable"	~12	Same result

Usage

import litellm

response = litellm.completion(
    model="gemini/gemini-2.0-flash-thinking-exp-01-21",
    messages=[{"role": "user", "content": "What is 2+2?"}],
    reasoning_effort="none"  # OpenAI standard, 96% cheaper!
)

Why "none" or "disable"?

• "none" = OpenAI standard ✅ (recommended)
• "disable" = LiteLLM legacy (same result, still works)

Both produce identical results, but "none" is the official OpenAI API specification.

Backward Compatibility

✅ Fully backward compatible - existing code continues to work unchanged.

---

Note

Adds support for reasoning_effort="none" for Gemini, mapping to thinkingBudget=0/includeThoughts=false, with model metadata, docs, and tests updated.

• Core (Gemini)
  • Map reasoning_effort="none" to thinkingConfig {"thinkingBudget": 0, "includeThoughts": false} in vertex_and_google_ai_studio_gemini.py.
• Model Metadata
  • Mark reasoning capability with "supports_reasoning": true for relevant Gemini models in model_prices_and_context_window.json.
• Tests
  • Add test_reasoning_effort_none_mapping to verify correct mapping.
• Docs
  • Update docs/providers/gemini.md with cost-optimization guidance, OpenAI compatibility, mapping table including "none", and note on Gemini 2.5 Pro limitations; add usage examples.

^{Written by Cursor Bugbot for commit c977ca1. This will update automatically on new commits. Configure here.}

[vercel]

@Chesars is attempting to deploy a commit to the CLERKIEAI Team on Vercel.

A member of the Team first needs to authorize it.