---
name: gemini-model-selection
description: Choose the optimal Gemini model based on use case, performance benchmarks, cost, and feature requirements
argument-hint: "<use case or requirements>"
allowed-tools: Read, Write, Bash(pip install, npm install, go get)
---

# Gemini Model Selection Guide

Choose the optimal Gemini model for your use case: $ARGUMENTS

## Expert Knowledge

You are a Gemini API specialist with expertise in:
- Model capabilities and performance characteristics
- Cost optimization and token pricing
- Latency requirements and throughput considerations
- Feature availability across model tiers
- 2026 benchmark data and real-world performance

## Model Comparison (Feb 2026)

| Model | ID | Context Window | Output Limit | Thinking | Best For |
|-------|-----|----------------|--------------|----------|----------|
| **Gemini 3 Pro** | `gemini-3-pro-preview` | 1M tokens (2M planned) | 65K tokens | Yes (levels) | Best multimodal, deep reasoning |
| **Gemini 3 Flash** | `gemini-3-flash-preview` | 1M tokens | 65K tokens | Yes (levels) | Coding, speed + intelligence |
| **Gemini 2.5 Pro** | `gemini-2.5-pro` | 1M tokens | 65K tokens | Yes (budget) | State-of-the-art thinking |
| **Gemini 2.5 Flash** | `gemini-2.5-flash` | 1M tokens | 65K tokens | Yes (budget) | Price-performance balance |
| **Gemini 2.5 Flash-Lite** | `gemini-2.5-flash-lite` | 1M tokens | 65K tokens | Optional | Ultra-fast, lowest cost |

## Benchmark Performance (2026)

### Reasoning & Knowledge

| Model | GPQA Diamond | Humanity's Last Exam | MMMU Pro | MMLU-Pro |
|-------|-------------|---------------------|----------|----------|
| Gemini 3 Pro | **90.4%** | 33.7% | 81.2% | ~85% |
| Gemini 3 Flash | 90.4% | 33.7% | **81.2%** | ~85% |
| Gemini 2.5 Pro | 85%+ | 28%+ | 75%+ | ~78% |
| Gemini 2.5 Flash | 80%+ | 25%+ | 72.9% | ~71% |
| Gemini 2.5 Flash-Lite | 70%+ | N/A | 72.9% | ~68% |

### Coding Performance

| Model | SWE-bench Verified | Notes |
|-------|-------------------|-------|
| Gemini 3 Flash | **78%** | Surprisingly beats Pro! |
| Gemini 3 Pro | 76.2% | Memory issues reported |
| Gemini 2.5 Pro | ~70% | Stable production |
| Gemini 2.5 Flash | ~65% | Good for most tasks |

### Speed & Throughput

| Model | Output Speed | Time to First Token | Relative Speed |
|-------|-------------|---------------------|----------------|
| Gemini 3 Flash | Very High | 0.21-0.37s | 3x faster than 2.5 Pro |
| Gemini 2.5 Flash | **250 tokens/sec** | Fast | #3 of 77 models |
| Gemini 2.5 Flash-Lite | Highest | **Fastest** | Optimized for latency |
| Gemini 2.5 Pro | Medium | Medium | Deeper thinking trade-off |
| Gemini 3 Pro | Slower | Slower | Quality over speed |

### AI Intelligence Index (Artificial Analysis)

| Model | Score | Rank |
|-------|-------|------|
| Gemini 3 Flash | 71 | Beats Claude 3 Opus (70) |
| Gemini 2.5 Flash | 21 | #17 of 77 models |

## Pricing (Feb 2026)

| Model | Input (per 1M) | Output (per 1M) | Relative Cost |
|-------|----------------|-----------------|---------------|
| Gemini 3 Pro | ~$2-4 | ~$8-15 | $$$$$ |
| Gemini 3 Flash | $0.50 | $3.00 | $$$ |
| Gemini 2.5 Pro | ~$1.25 | ~$5.00 | $$$$ |
| Gemini 2.5 Flash | $0.30 | $2.50 | $$ |
| Gemini 2.5 Flash-Lite | **$0.10** | **$0.40** | $ |

### Cost Comparison
- Flash-Lite is **15x cheaper** than Pro for input tokens
- Gemini 3 Flash is **4-8x cheaper** than Gemini 3 Pro
- 2.5 Flash uses **30% fewer tokens** than 2.5 Pro on average

## Feature Availability

| Feature | 3 Pro | 3 Flash | 2.5 Pro | 2.5 Flash | 2.5 Flash-Lite |
|---------|-------|---------|---------|-----------|----------------|
| Function Calling | Yes | Yes | Yes | Yes | Yes |
| Grounding (Search) | Yes | Yes | Yes | Yes | Yes |
| Code Execution | Yes | Yes | Yes | Yes | No |
| Multimodal (Image) | Yes | Yes | Yes | Yes | Yes |
| Multimodal (Video) | Yes | Yes | Yes | Yes | Limited |
| Multimodal (Audio) | Yes | Yes | Yes | Yes | No |
| Image Generation | Yes (Pro Image) | No | No | No | No |
| Structured Output | Yes | Yes | Yes | Yes | Yes |
| Thinking Mode | Levels | Levels | Budget | Budget | Optional (off default) |
| Context Caching | Yes | Yes | Yes | Yes | Yes |
| Live API | Yes | Yes | Yes | Yes | No |

## Use Case Mapping

### Gemini 3 Pro (`gemini-3-pro-preview`)
**Best for:**
- Best multimodal understanding in the world
- Complex agentic workflows requiring deep reasoning
- Research and analysis requiring highest accuracy
- When quality is absolutely paramount
- Extended thinking for nuanced analysis ("Deep Think" mode)

**Limitations:** Slower, most expensive, reported memory issues

```python
response = client.models.generate_content(
    model="gemini-3-pro-preview",
    contents="Analyze this complex architectural diagram and suggest improvements..."
)
```

### Gemini 3 Flash (`gemini-3-flash-preview`)
**Best for:**
- **Coding tasks** (actually beats Pro on SWE-bench!)
- Production applications needing speed + frontier intelligence
- High-frequency workflows at scale
- Real-time assistants
- Best value for advanced capabilities

**Key advantage:** 3x faster than 2.5 Pro, competitive with 3 Pro quality

**Warning:** 91% hallucination rate on refusal benchmarks - needs guardrails

```python
response = client.models.generate_content(
    model="gemini-3-flash-preview",
    contents="Debug this code and suggest optimizations..."
)
```

### Gemini 2.5 Pro (`gemini-2.5-pro`)
**Best for:**
- State-of-the-art thinking model
- Stable production workloads
- Complex reasoning with proven reliability
- Long-form content generation
- When you need battle-tested stability

```python
response = client.models.generate_content(
    model="gemini-2.5-pro",
    contents="Generate a detailed technical specification..."
)
```

### Gemini 2.5 Flash (`gemini-2.5-flash`)
**Best for:**
- Best price-performance ratio
- General-purpose applications
- Large-scale processing
- Chat and Q&A systems
- **Recommended default for most use cases**

**Speed:** 250 tokens/sec (#3 fastest of 77 models)

```python
response = client.models.generate_content(
    model="gemini-2.5-flash",
    contents="Summarize this article..."
)
```

### Gemini 2.5 Flash-Lite (`gemini-2.5-flash-lite`)
**Best for:**
- Ultra-high volume, latency-sensitive tasks
- Classification, translation, summarization at scale
- Cost-sensitive applications
- Simple Q&A and extraction
- When thinking mode is not needed

**Key features:** Lowest latency, lowest cost, thinking off by default

```python
response = client.models.generate_content(
    model="gemini-2.5-flash-lite",
    contents="Classify this support ticket: ..."
)
```

## Decision Matrix

| Need | Best Choice | Runner-up |
|------|-------------|-----------|
| **Best coding** | 3 Flash (78% SWE-bench) | 3 Pro |
| **Best multimodal** | 3 Pro | 3 Flash |
| **Fastest** | 2.5 Flash-Lite | 2.5 Flash |
| **Cheapest** | 2.5 Flash-Lite | 2.5 Flash |
| **Best reasoning** | 3 Pro | 2.5 Pro |
| **Best value** | 3 Flash | 2.5 Flash |
| **Production stable** | 2.5 Pro | 2.5 Flash |
| **High volume** | 2.5 Flash-Lite | 2.5 Flash |

## Decision Tree

```
START
  |
  v
Need cutting-edge capabilities (Gemini 3)?
  |-- Yes --> Coding-focused?
  |             |-- Yes --> gemini-3-flash-preview (beats Pro!)
  |             |-- No  --> Need absolute best multimodal?
  |                           |-- Yes --> gemini-3-pro-preview
  |                           |-- No  --> gemini-3-flash-preview
  |
  |-- No --> Need thinking/reasoning mode?
              |-- Yes --> Need stability?
              |             |-- Yes --> gemini-2.5-pro
              |             |-- No  --> gemini-2.5-flash
              |
              |-- No --> High volume / cost-sensitive?
                          |-- Yes --> gemini-2.5-flash-lite
                          |-- No  --> gemini-2.5-flash
```

## Cost Optimization Tips

1. **Start with Flash-Lite** for simple tasks, upgrade if quality insufficient
2. **Use Gemini 3 Flash over 3 Pro** for coding (better + cheaper!)
3. **Use caching** for repeated context (up to 75% savings)
4. **Batch requests** with Batch API (50% discount)
5. **Right-size context** - 2.5 Flash uses 30% fewer tokens than 2.5 Pro
6. **Disable thinking** in Flash-Lite for maximum speed

## Deprecation Schedule

| Model | Status | Deprecation Date |
|-------|--------|------------------|
| Gemini 2.0 Flash | Deprecated | March 31, 2026 |
| Gemini 2.0 Flash-Lite | Deprecated | March 31, 2026 |
| Gemini 1.5 Pro | Deprecated | September 2025 |
| Gemini 1.5 Flash | Deprecated | September 2025 |

> **Migration**: Replace `gemini-2.0-*` with `gemini-2.5-*` equivalents

## Model Selection Code

```python
def select_model(
    task_type: str,
    volume: str = "normal",
    quality: str = "standard",
    speed_priority: bool = False
) -> str:
    """Select optimal Gemini model based on requirements."""

    # Coding: Gemini 3 Flash actually beats Pro!
    if task_type in ["coding", "code_review", "debugging"]:
        return "gemini-3-flash-preview"

    # Best multimodal understanding
    if task_type in ["multimodal", "image_analysis", "video_understanding"]:
        return "gemini-3-pro-preview" if quality == "highest" else "gemini-3-flash-preview"

    # Agentic and complex reasoning
    if task_type in ["agentic", "complex_reasoning", "research"]:
        return "gemini-3-pro-preview" if quality == "highest" else "gemini-3-flash-preview"

    # Stable production reasoning
    if task_type in ["analysis", "long_form"] and quality == "stable":
        return "gemini-2.5-pro"

    # High volume / cost sensitive
    if volume == "high" or task_type in ["classification", "extraction", "translation"]:
        return "gemini-2.5-flash-lite"

    # Speed priority
    if speed_priority:
        return "gemini-2.5-flash-lite"

    # Default: best price-performance
    return "gemini-2.5-flash"

# Usage examples
print(select_model("coding"))  # gemini-3-flash-preview
print(select_model("classification", volume="high"))  # gemini-2.5-flash-lite
print(select_model("research", quality="highest"))  # gemini-3-pro-preview
```

## Common Patterns

| Pattern | Recommended Model | Reason |
|---------|-------------------|--------|
| Code assistant | **3 Flash** | 78% SWE-bench, fast |
| Chatbot | 2.5 Flash | Price-performance |
| Document summarization | 2.5 Flash | General purpose |
| Data extraction | 2.5 Flash-Lite | Simple, high volume |
| Agent orchestration | 3 Pro or 3 Flash | Complex reasoning |
| Real-time translation | 2.5 Flash-Lite | Lowest latency |
| Research assistant | 3 Pro | Deepest reasoning |
| Image understanding | 3 Pro | Best multimodal |

## Sources

- [Gemini Models Documentation](https://ai.google.dev/gemini-api/docs/models)
- [Gemini 3 Flash Announcement](https://blog.google/products-and-platforms/products/gemini/gemini-3-flash/)
- [Artificial Analysis Benchmarks](https://artificialanalysis.ai/models/gemini-2-5-flash)
- [Gemini 2.5 Developer Blog](https://developers.googleblog.com/en/gemini-2-5-thinking-model-updates/)
- [Better Stack Gemini 3 Flash Review](https://betterstack.com/community/guides/ai/gemini-3-flash-review/)

## Deliverables

For: $ARGUMENTS

Provide:
1. Recommended model with justification based on benchmarks
2. Alternative options if requirements change
3. Cost estimation with actual pricing
4. Performance expectations (speed, quality)
5. Code example with selected model