# PRD #294: Add Gemini 3 Model Variants Support
## Overview
**Status**: Complete
**Priority**: Medium
**Created**: 2025-12-21
**GitHub Issue**: [#294](https://github.com/vfarcic/dot-ai/issues/294)
## Problem Statement
Users currently only have access to Gemini 3 Pro (`google` provider) but lack access to other Gemini 3 variants that offer different trade-offs:
- **Gemini 3 Flash**: Faster response times, lower cost, same 1M context
- **Thinking configurations**: Control reasoning depth vs speed/cost
## Solution
Add Gemini 3 variants iteratively: implement one variant, run all tests, document based on results, then proceed to the next.
## Gemini 3 Model Specifications
| Variant | Model ID | Input Context | Output Limit | Thinking Levels |
|---------|----------|---------------|--------------|-----------------|
| Pro (existing) | `gemini-3-pro-preview` | 1,048,576 | 65,536 | low, high |
| Flash | `gemini-3-flash-preview` | 1,048,576 | 65,536 | minimal, low, medium, high |
**Pricing** (Gemini 3 Flash): $0.50/1M input, $3/1M output
## Proposed Provider Names
| Provider Name | Model | Thinking Level | Use Case |
|---------------|-------|----------------|----------|
| `google` | gemini-3-pro-preview | high (default) | Existing - complex reasoning |
| `google_flash` | gemini-3-flash-preview | high (default) | Fast + smart - balanced |
| `google_flash_fast` | gemini-3-flash-preview | minimal | Maximum speed, simple tasks |
## Implementation Approach
**Iterative process for each variant:**
1. Add to `model-config.ts`
2. Add case to `vercel-provider.ts`
3. Run full integration test suite
4. Document results (pass/fail, any issues)
5. Update `docs/setup/mcp-setup.md` if tests pass
6. Proceed to next variant
## Success Criteria
- [x] Each new provider works with existing functionality
- [x] Integration tests pass for each variant (98.7% pass rate for `google_flash`)
- [x] Documentation accurately describes each option
- [~] Thinking level configuration works correctly (deferred - not needed for current variants)
## Milestones
### Milestone 1: Gemini 3 Flash (google_flash) ✅
- [x] Add `google_flash: 'gemini-3-flash-preview'` to `CURRENT_MODELS`
- [x] Add `google_flash` case to `vercel-provider.ts`
- [x] Run integration tests (74/75 passed - 98.7%)
- [x] Document results and update docs if successful
### Milestone 2: Gemini 3 Flash Fast (google_flash_fast) - DEFERRED
- [x] Research Vercel AI SDK support for `thinking_level` parameter
- [~] Add `google_flash_fast` with minimal thinking configuration
- [~] Run integration tests
- [~] Document results and update docs if successful
**Deferral Reason**: Gemini 3 uses dynamic thinking by default (`high` level). The model automatically adjusts reasoning depth based on prompt complexity - simple tasks already use minimal thinking. The `google_flash_fast` variant (explicit `minimal` level) provides marginal benefit over this dynamic behavior. Will implement if user demand emerges.
### Milestone 3: Documentation & Cleanup ✅
- [x] Final documentation review
- [x] Update CLAUDE.md if needed
- [x] Verify all variants work end-to-end
## Technical Notes
### Thinking Level Configuration
Gemini 3 uses `thinking_config` in generation config:
```python
thinking_config=ThinkingConfig(thinking_level="minimal")
```
Need to verify Vercel AI SDK (`@ai-sdk/google`) support for this parameter.
### API Key
All Gemini 3 variants use the same `GOOGLE_GENERATIVE_AI_API_KEY`.
## Out of Scope
- Gemini 2.5 models
- Gemini 3 Pro Image (`gemini-3-pro-image-preview`)
- Custom Gemini endpoint support
- Per-request thinking level configuration
## Dependencies
- Existing `@ai-sdk/google` package
- `GOOGLE_GENERATIVE_AI_API_KEY` environment variable
## Risks & Mitigations
| Risk | Likelihood | Impact | Mitigation |
|------|------------|--------|------------|
| Models still in preview | Medium | Low | Document preview status; monitor for GA |
| Vercel SDK lacks thinking_level support | Medium | Medium | Skip fast variant or implement workaround |
| API behavior differences between variants | Low | Medium | Test each variant thoroughly |
## Progress Log
| Date | Update |
|------|--------|
| 2025-12-21 | PRD created with iterative implementation approach |
| 2025-12-23 | **Milestone 1 complete**: Added `google_flash` provider. 98.7% test pass rate (74/75). Fixed JSON parsing for multi-model support, standardized `GOOGLE_GENERATIVE_AI_API_KEY` env var, updated docs. |
| 2025-12-23 | **Milestone 2 deferred**: Research confirmed Vercel AI SDK supports `thinkingLevel` parameter. However, Gemini 3's dynamic thinking (default `high`) already minimizes reasoning for simple tasks automatically. `google_flash_fast` deferred until user demand emerges. |
| 2025-12-23 | **PRD complete**: `google_flash` provider shipped. Milestone 2 deferred. |
## Decisions
| Decision | Date | Rationale |
|----------|------|-----------|
| Iterative implementation | 2025-12-21 | Test each variant before proceeding; reduces risk |
| Start with `google_flash` | 2025-12-21 | Most straightforward - same config as existing `google` |
| Share API key across variants | 2025-12-21 | Same Google AI API; reduces configuration burden |
| Defer `google_flash_fast` | 2025-12-23 | Dynamic thinking at `high` level already optimizes for simple tasks; explicit `minimal` level provides marginal benefit |
