Available LLMs in Genesis¶

Genesis supports a wide range of Large Language Models from multiple providers with automatic cycling and dynamic selection capabilities.

Supported Providers¶

Genesis integrates with 4 major LLM providers via unified API:

Anthropic Claude (via Anthropic API)
OpenAI GPT (via OpenAI API)
Google Gemini (via OpenAI-compatible API)
DeepSeek (via OpenAI-compatible API)
AWS Bedrock (via AnthropicBedrock API)
Azure OpenAI (via Azure API)

Available Models¶

Anthropic Claude Models¶

Model	Input Price	Output Price	Best For
`claude-3-5-haiku-20241022`	$0.80/M	$4.00/M	Fast, lightweight tasks
`claude-3-5-sonnet-20241022`	$3.00/M	$15.00/M	Balanced performance
`claude-3-5-sonnet-latest`	$3.00/M	$15.00/M	Latest Sonnet version
`claude-3-opus-20240229`	$15.00/M	$75.00/M	Most capable
`claude-3-7-sonnet-20250219`	$3.00/M	$15.00/M	Extended thinking (reasoning)
`claude-4-sonnet-20250514`	$3.00/M	$15.00/M	Claude 4 with reasoning
`claude-sonnet-4-5-20250929`	$3.00/M	$15.00/M	Latest reasoning model

Features: - Extended thinking mode for reasoning models - Native structured output (not yet implemented) - High context windows (200K+ tokens)

OpenAI GPT Models¶

Model	Input Price	Output Price	Best For
`gpt-4o-mini`	$0.15/M	$0.60/M	Fastest, cheapest
`gpt-4o-mini-2024-07-18`	$0.15/M	$0.60/M	Specific mini version
`gpt-4o-2024-08-06`	$2.50/M	$10.00/M	General purpose
`gpt-4.1`	$2.00/M	$8.00/M	Latest GPT-4 series
`gpt-4.1-2025-04-14`	$2.00/M	$8.00/M	Specific 4.1 version
`gpt-4.1-mini`	$0.40/M	$1.60/M	Lightweight 4.1
`gpt-4.1-nano`	$0.10/M	$1.40/M	Ultra lightweight
`gpt-4.5-preview-2025-02-27`	$75.00/M	$150.00/M	Most advanced (preview)
Reasoning Models:
`o1-2024-12-17`	$15.00/M	$60.00/M	Deep reasoning
`o3-mini`	$1.10/M	$4.40/M	Efficient reasoning
`o3-mini-2025-01-31`	$1.10/M	$4.40/M	Reasoning mini
`o3-2025-04-16`	$10.00/M	$40.00/M	Advanced reasoning
`o4-mini`	$1.10/M	$4.40/M	Latest mini reasoning
GPT-5 Series (Hypothetical):
`gpt-5`	$1.25/M	$10.00/M	Next generation
`gpt-5-mini`	$0.25/M	$2.00/M	Next gen mini
`gpt-5-nano`	$0.05/M	$0.40/M	Next gen nano
`gpt-5.1`	$1.25/M	$10.00/M	Future version

Features: - Reasoning models with chain-of-thought - Structured output support - Function calling - Vision capabilities (not used in Genesis)

DeepSeek Models¶

Model	Input Price	Output Price	Best For
`deepseek-chat`	$0.27/M	$1.10/M	General chat, very cheap
`deepseek-reasoner`	$0.55/M	$2.19/M	Reasoning tasks, cost-effective

Features: - Most cost-effective models - Native reasoning support - Great for code generation - Chinese & English bilingual

Google Gemini Models¶

Model	Input Price	Output Price	Best For
`gemini-2.5-pro`	$1.25/M	$10.00/M	Most capable Gemini
`gemini-2.5-flash`	$0.30/M	$2.50/M	Fast, balanced
`gemini-2.5-flash-lite-preview-06-17`	$0.10/M	$0.40/M	Ultra-fast preview
`gemini-3-pro-preview`	$2.00/M	$12.00/M	Next generation (preview)

Features: - Long context (up to 1M tokens for some models) - Multimodal (not used in Genesis) - Native thinking mode - Very competitive pricing

AWS Bedrock Models¶

Access Anthropic models via AWS Bedrock:

Model	Notes
`bedrock/anthropic.claude-3-5-sonnet-20241022-v2:0`	Claude Sonnet via Bedrock
`bedrock/anthropic.claude-3-5-haiku-20241022-v1:0`	Claude Haiku via Bedrock
`bedrock/anthropic.claude-3-opus-20240229-v1:0`	Claude Opus via Bedrock
`bedrock/us.anthropic.claude-3-7-sonnet-20250219-v1:0`	Claude 3.7 Sonnet
`bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0`	Claude 4 Sonnet

Requirements: - AWS credentials in environment: AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_REGION_NAME - Same pricing as direct Anthropic API (may vary by region)

Azure OpenAI Models¶

Access OpenAI models via Azure:

Model Prefix	Example	Notes
`azure-*`	`azure-gpt-4.1`, `azure-o3-mini`	Prefix any OpenAI model with `azure-`

Requirements: - AZURE_OPENAI_API_KEY - AZURE_API_VERSION - AZURE_API_ENDPOINT

Model Selection System¶

Genesis supports 3 modes for model selection:

1. Static Single Model (Default)¶

Use one model for all mutations:

evo_config:
  llm_models:
    - gpt-4.1

2. Static Multi-Model (Random Selection)¶

Randomly select from a list:

evo_config:
  llm_models:
    - gpt-4.1
    - claude-3-5-sonnet-20241022
    - gemini-2.5-flash

Models are selected randomly with equal probability.

3. Dynamic Model Selection (UCB Bandit)¶

Automatically learns which models perform best and increases their usage:

evo_config:
  llm_models:
    - gpt-4.1
    - gpt-4.1-mini
    - claude-3-5-sonnet-20241022
    - claude-3-5-haiku-20241022
    - gemini-2.5-flash
    - deepseek-chat
  llm_dynamic_selection: "ucb"  # Enable adaptive selection
  llm_dynamic_selection_kwargs:
    exploration_coef: 1.0       # Exploration vs exploitation
    epsilon: 0.2                 # Random exploration rate
    auto_decay: 0.95             # Decay old observations

How it works: 1. Initial Phase: All models tried equally 2. Learning Phase: Models producing better offspring get higher selection probability 3. Exploitation Phase: Best-performing models used more frequently 4. Continuous Adaptation: Keeps exploring to find better models

Algorithm: Asymmetric UCB (Upper Confidence Bound) - Tracks mean reward per model - Adds confidence bonus for exploration - Shifts rewards relative to baseline/parent fitness - Adapts to changing distributions

Benefits: - Automatically finds best models for your task - Balances cost vs performance - Adapts as evolution progresses - No manual tuning needed

Configuration Examples¶

Cost-Optimized Setup¶

evo_config:
  llm_models:
    - gpt-4o-mini        # $0.15/M input
    - deepseek-chat      # $0.27/M input
    - gemini-2.5-flash-lite-preview-06-17  # $0.10/M input
  llm_dynamic_selection: "ucb"

Performance-Optimized Setup¶

evo_config:
  llm_models:
    - claude-3-5-sonnet-20241022
    - gpt-4.1
    - gemini-2.5-pro
  llm_dynamic_selection: "ucb"

Reasoning-Focused Setup¶

evo_config:
  llm_models:
    - o3-mini-2025-01-31
    - claude-3-7-sonnet-20250219
    - deepseek-reasoner
    - gemini-2.5-pro
  llm_dynamic_selection: "ucb"

Cost-Performance Balanced¶

evo_config:
  llm_models:
    - gpt-4.1-mini       # Good balance
    - claude-3-5-haiku-20241022
    - gemini-2.5-flash
    - deepseek-chat
  llm_dynamic_selection: "ucb"

Environment Variables¶

Set these in your .env file:

# OpenAI
OPENAI_API_KEY=sk-...

# Anthropic
ANTHROPIC_API_KEY=sk-ant-...

# DeepSeek
DEEPSEEK_API_KEY=sk-...

# Gemini
GEMINI_API_KEY=AIzaSy...

# Azure OpenAI
AZURE_OPENAI_API_KEY=...
AZURE_API_VERSION=2024-02-15-preview
AZURE_API_ENDPOINT=https://your-resource.openai.azure.com/

# AWS Bedrock
AWS_ACCESS_KEY_ID=AKIA...
AWS_SECRET_ACCESS_KEY=...
AWS_REGION_NAME=us-east-1

Model Selection Monitoring¶

When using dynamic selection, Genesis logs model performance:

AsymmetricUCB Summary:
arm                       n    mean  exploit  explore  score    post
gpt-4.1                   5   0.850   0.850    0.120   0.970   0.45
claude-3-5-sonnet         3   0.820   0.820    0.156   0.976   0.35
gemini-2.5-flash          2   0.750   0.750   0.189   0.939   0.15
deepseek-chat             1   0.700   0.700   0.268   0.968   0.05

n: Number of times used
mean: Average normalized reward
exploit: Exploitation component (mean fitness)
explore: Exploration bonus (uncertainty)
score: Combined UCB score (exploit + explore)
post: Selection probability for next iteration

Cost Tracking¶

Genesis tracks costs per model in ClickHouse:

SELECT 
    model,
    count() as calls,
    sum(cost) as total_cost,
    avg(cost) as avg_cost_per_call
FROM llm_logs
GROUP BY model
ORDER BY total_cost DESC;

Recommendations¶

For Research/Experiments¶

Use dynamic selection with diverse models
Include both fast and capable models
Monitor costs via ClickHouse

For Production¶

Start with dynamic selection to find best model
Once identified, switch to single best model
Use cost-optimized models for large populations

For Code Evolution¶

Reasoning models (o-series, claude-3-7, deepseek-reasoner) often perform best
Balance with faster models (mini/flash variants) for speed
DeepSeek offers best cost-performance for code tasks

Adding New Models¶

To add a new model:

Add pricing to genesis/llm/models/pricing.py:

OPENAI_MODELS = {
    "new-model-name": {
        "input_price": X / M,
        "output_price": Y / M,
    }
}

Add to appropriate provider function in genesis/llm/models/
Use in config:
```
llm_models:
  - new-model-name
```

Summary¶

✅ 60+ models across 6 providers ✅ Dynamic model selection with UCB bandit algorithm ✅ Cost tracking and monitoring ✅ Reasoning models for complex tasks ✅ Automatic adaptation to find best models

Genesis provides unparalleled flexibility in LLM selection, allowing you to optimize for cost, performance, or automatically discover the best model for your specific evolutionary task.