> ## Documentation Index
> Fetch the complete documentation index at: https://docs.elementum.io/llms.txt
> Use this file to discover all available pages before exploring further.
# AI Models
> Comprehensive guide to available AI models, their capabilities, and when to use them
Elementum supports a wide range of AI models across multiple providers. Models are accessed through configured AI providers -- you'll select your provider first, then choose from the models available through that provider. To use a model in your workflows, create an [AI Service](/ai-agents/ai-services) that references the provider and model you want.
## Available Providers
* **Snowflake** -- Claude, OpenAI (Cortex), Llama, Mistral, DeepSeek, embedding models
* **OpenAI** -- GPT models
* **Anthropic** -- Direct access to Claude models. Primary provider supported on Studio Agents. See the [Anthropic setup guide](/ai-agents/anthropic-setup) for configuration details.
* **Gemini** -- Multimodal Gemini models
* **Bedrock** -- Claude models via AWS Bedrock. See the [AWS Bedrock Setup](/ai-agents/bedrock-setup) guide to configure the provider, or [AWS Bedrock Agents Setup](/ai-agents/bedrock-agents-setup) to connect a Bedrock Agent built in AWS.
* **Custom** -- Any OpenAI-compatible endpoint, including LLM gateways, proxies, and self-hosted models. See [Configure a Custom Provider](/ai-agents/ai-services#configure-a-custom-provider) for setup details.
**Important:** Models are accessed through providers you configure in Organization Settings. Claude models, for example, can be accessed directly through the Anthropic provider or through the Snowflake or Bedrock providers, depending on your data residency and compliance needs.
Model availability varies by provider configuration, account tier, and region. Not all models listed below may be available in your environment. Check your provider's [service creation screen](/ai-agents/ai-services#creating-ai-services) for the models currently accessible to your organization.
**Per-feature availability**: Even when a model is configured in your organization, it may not appear in every AI feature's dropdown. Elementum enables models per feature based on task suitability, so the model list inside **AI Classification** can differ from the list inside **Transform Data with AI**, **AI Summarization**, or **AI File Analysis**. See [Model Availability Across AI Actions](/ai-agents/ai-automations#model-availability-across-ai-actions) for details.
## Quick Reference: Model Capabilities
This table shows which capabilities each model category supports:
| Model | Provider | Multimodal | Structured Output | Reasoning | Agents | Best Use Case |
| ------------------------ | ----------------------------- | ---------- | ----------------- | ---------- | ------ | ---------------------------------------------------- |
| **GPT-5 Series** | OpenAI | No | Yes | Yes | Yes | Complex reasoning, demanding applications |
| **GPT-4.1 Series** | OpenAI | No | Yes | No | Yes | Production deployments, reliable automation |
| **GPT-4o Series** | OpenAI | Yes | Yes | No | Yes | General-purpose, document analysis |
| **o3-mini / o1-mini** | OpenAI | No | No | Yes (req.) | No | Mathematical reasoning, logic problems |
| **GPT-4 / GPT-3.5** | OpenAI | No | Partial | No | No | Legacy applications |
| **Claude 4.5 Sonnet** | Anthropic, Snowflake, Bedrock | No | No | No | Yes | Advanced reasoning, detailed analysis, Studio Agents |
| **Claude 4 Opus** | Anthropic, Snowflake, Bedrock | No | No | No | Yes | Highest-capability reasoning |
| **Claude 4 Sonnet** | Anthropic, Snowflake, Bedrock | No | No | No | Yes | Balanced performance |
| **Claude Haiku 4.5** | Anthropic, Snowflake, Bedrock | No | No | No | Yes | Fast, efficient processing |
| **Claude 3.7 / 3.5** | Anthropic, Snowflake, Bedrock | No | No | No | Yes | Cost-effective reasoning |
| **Cortex GPT-5** | Snowflake | No | Yes | Yes | Yes | OpenAI through Snowflake |
| **Cortex GPT-4.1** | Snowflake | No | Yes | No | Yes | Production OpenAI via Snowflake |
| **Cortex o4-mini** | Snowflake | No | No | Yes (req.) | Yes | Reasoning through Snowflake |
| **Gemini 3 Pro** | Gemini | Yes | No | No | Yes | Latest multimodal |
| **Gemini 2.5 Pro/Flash** | Gemini | Yes | No | No | Yes | Production multimodal |
| **Gemini 2.0 Flash** | Gemini | Yes | No | No | No | Cost-effective multimodal |
| **Gemini 1.5 Pro** | Gemini | Yes | No | No | No | Established multimodal |
| **DeepSeek R1** | Snowflake | No | No | No | No | Open-source reasoning |
| **Llama 3.3 70B** | Snowflake | No | No | No | No | Open-source, balanced |
| **Llama 3.1 Series** | Snowflake | No | Partial | No | No | Open-source, structured output |
| **Llama 3 Series** | Snowflake | No | No | No | No | Open-source, function calling |
| **Mistral Large 2** | Snowflake | No | No | No | No | Multilingual, European focus |
| **Mistral 7B / Mixtral** | Snowflake | No | No | No | No | Efficient small models |
| **Snowflake Arctic** | Snowflake | No | No | No | No | Data cloud native |
| **Arctic Embeddings** | Snowflake | N/A | N/A | N/A | N/A | Semantic search |
**Legend:**
* **Multimodal**: Processes text and images together
* **Structured Output**: Guaranteed JSON/XML format responses
* **Reasoning**: Advanced reasoning mode (req. = required, always on)
* **Agents**: Supports Elementum agent workflows
## Models by Provider
### Snowflake Cortex
Snowflake Cortex provides access to multiple AI model families through your Snowflake data cloud.
**Data Residency**: All Snowflake Cortex models run within your Snowflake environment, keeping data in your cloud.
#### Anthropic Claude (via Snowflake)
**claude-sonnet-4-5**
* **Best for:** Complex reasoning and analysis tasks
* **Capabilities:** Strong reasoning, nuanced understanding, extensive context windows
* **Use cases:** Complex research, detailed analysis, advanced automation, agent workflows
* **Temperature range:** 0.0 - 1.0 (default: 0.7)
* **When to use:** Demanding applications requiring deep understanding and analysis
**Claude 4 Opus** (claude-4-opus)
* **Best for:** Highest-capability reasoning tasks
* **Use cases:** Strategic decisions, complex research, mission-critical analysis
* **When to use:** Tasks where quality matters more than cost
**Claude 4 Sonnet** (claude-4-sonnet)
* **Best for:** Balanced performance and cost for demanding tasks
* **Use cases:** Business automation, production workflows, detailed analysis
* **When to use:** Production workloads needing strong reasoning
**Claude Haiku 4.5** (claude-haiku-4-5)
* **Best for:** Fast, efficient processing
* **Use cases:** High-volume operations, real-time interactions, simple automation
* **When to use:** Speed and cost-efficiency are priorities
**Claude 3.7 Sonnet** (claude-3-7-sonnet)
* **Best for:** Daily tasks requiring strong reasoning at lower cost
* **Use cases:** Standard business automation, customer support, content generation
* **When to use:** Cost-effective production deployments
**Claude 3.5 Sonnet** (claude-3-5-sonnet)
* **Best for:** Reliable production performance
* **Use cases:** Established workflows, production automation
* **When to use:** Stability and consistent performance matter
#### OpenAI via Cortex
Access OpenAI models through your Snowflake environment:
**openai-gpt-5** - Advanced reasoning
**openai-gpt-5-mini** - Efficient reasoning
**openai-gpt-5-nano** - Maximum efficiency
**openai-gpt-5-chat** - Optimized for conversations
* **Capabilities:** Structured output, reasoning mode, through Snowflake
* **Use cases:** Complex analysis, conversations, classification, intelligence features, agents
* **When to use:** Need OpenAI capabilities with Snowflake data residency
**openai-gpt-4.1**
* Production-ready OpenAI through Snowflake
* Structured output, reliable reasoning
**openai-o4-mini**
* Reasoning model through Snowflake
* Required temperature: 1.0 (not adjustable)
* System role not supported
**When to use:** Production OpenAI workloads within Snowflake environment
#### Open Source Models (via Snowflake)
**deepseek-r1**
* **Best for:** Advanced reasoning with open-source flexibility
* **Capabilities:** Strong reasoning, open-source architecture
* **Use cases:** Research, academic applications, cost-conscious deployments
* **When to use:** Open-source requirements or research projects
**Llama 3.3 Series**
* **Llama 3.3 70B** (llama3.3-70b): Latest generation, balanced performance
**Llama 3.2 Series**
* **Llama 3.2 3B** (llama3.2-3b): Efficient, compact
* **Llama 3.2 1B** (llama3.2-1b): Maximum efficiency for simple tasks
**Llama 3.1 Series**
* **Llama 3.1 405B** (llama3.1-405b): Largest, most capable
* **Llama 3.1 70B** (llama3.1-70b): Production-ready, structured output support
* **Llama 3.1 8B** (llama3.1-8b): Cost-effective, structured output support
**Llama 3 Series**
* **Llama 3 70B** (llama3-70b): Function calling, reliable
* **Llama 3 8B** (llama3-8b): Efficient operation
**Llama 2 Series**
* **Llama 2 70B Chat** (llama2-70b-chat): Conversational focus
**When to use:** Open-source requirements, cost optimization, specific model sizes
**Mistral Large 2** (mistral-large2)
* Advanced capabilities, multilingual support, European markets
**Mistral Large** (mistral-large)
* Previous generation, reliable performance
**Mistral 7B** (mistral-7b)
* Compact, efficient, cost-effective
**Mixtral 8x7B** (mixtral-8x7b)
* Mixture-of-experts architecture, balanced performance
**When to use:** International applications, multilingual needs, small model requirements
**Snowflake Arctic** (snowflake-arctic)
* Data cloud native processing, integrated with Snowflake infrastructure
* **When to use:** Data-intensive workflows within Snowflake
**Gemma 7B** (gemma-7b)
* Lightweight Google-developed model
* **When to use:** Efficient processing on smaller tasks
**Jamba Instruct** (jamba-instruct)
* Instruction-following optimization
* **Note:** Not recommended for JSON/YAML parsing
**Reka Core / Flash** (reka-core, reka-flash)
* Advanced processing or fast operation
#### Snowflake Embedding Models
* **Arctic L V2.0** (`snowflake-arctic-embed-l-v2.0`) -- High-quality embeddings for semantic search. Use for new AI search implementations.
* **Arctic M V1.5** (`snowflake-arctic-embed-m-v1.5`) -- Balanced performance and quality. Use for production search systems.
### Anthropic Direct
Access Claude models directly through the Anthropic API. The Anthropic provider is the primary provider supported on Studio Agents and gives organizations a direct path to Claude without routing requests through Snowflake or AWS Bedrock.
**Studio Agents**: Anthropic is the primary provider supported on [Studio Agents](/ai-agents/studio-agents). Configure an Anthropic provider before using Studio Agents to build automations, agents, and flows.
**claude-sonnet-4-5**
* **Best for:** Complex reasoning, detailed analysis, and Studio Agents
* **Capabilities:** Strong reasoning, nuanced understanding, extensive context windows, agent support
* **Use cases:** Studio Agents, advanced automation, conversational agents, complex research
* **When to use:** Demanding applications that benefit from direct Anthropic access
**Claude 4 Opus** (claude-4-opus)
* **Best for:** Highest-capability reasoning tasks
* **Use cases:** Strategic decisions, complex research, mission-critical analysis
* **When to use:** Tasks where quality matters more than cost
**Claude 4 Sonnet** (claude-4-sonnet)
* **Best for:** Balanced performance and cost for demanding tasks
* **Use cases:** Business automation, production workflows, detailed analysis
* **When to use:** Production workloads needing strong reasoning
**Claude Haiku 4.5** (claude-haiku-4-5)
* **Best for:** Fast, efficient processing
* **Use cases:** High-volume operations, real-time interactions, simple automation
* **When to use:** Speed and cost-efficiency are priorities
**Claude 3.7 Sonnet** (claude-3-7-sonnet)
* **Best for:** Daily tasks requiring strong reasoning at lower cost
* **Use cases:** Standard business automation, customer support, content generation
* **When to use:** Cost-effective production deployments
**Claude 3.5 Sonnet** (claude-3-5-sonnet)
* **Best for:** Reliable production performance
* **Use cases:** Established workflows, production automation
* **When to use:** Stability and consistent performance matter
**Note**: Embeddings for AI Search are handled exclusively through Snowflake Cortex. Anthropic models are used for LLM services only. Specific model availability depends on your Anthropic account access.
### OpenAI Direct
Access OpenAI models directly through OpenAI API.
**GPT-5** (gpt-5)
* **Best for:** Complex reasoning and demanding applications
* **Capabilities:** Structured output, reasoning mode, advanced problem-solving
* **Use cases:** Complex analysis, strategic planning, research tasks
**GPT-5 Mini** (gpt-5-mini)
* **Best for:** Daily reasoning at lower cost
* **Capabilities:** Structured output, reasoning mode, balanced performance
* **Use cases:** Standard business logic, moderate analysis, automation
**GPT-5 Nano** (gpt-5-nano)
* **Best for:** Simple reasoning requiring efficiency
* **Capabilities:** Structured output, reasoning mode, cost-effective
* **Use cases:** Basic classification, simple analysis, high-volume operations
**GPT-5.1** (gpt-5.1)
* **Best for:** Enhanced reasoning with improved accuracy
* **Use cases:** Business intelligence, detailed analysis, critical decisions
**GPT-5.2** (gpt-5.2)
* **Best for:** Highest-tier reasoning
* **Use cases:** Complex problem-solving, research, mission-critical applications
**All GPT-5 models support:** Prompts, conversations, classification, intelligence features, agents
**GPT-4.1** (gpt-4.1)
* **Best for:** Production applications requiring consistent performance
* **Capabilities:** Structured output, reliable reasoning
* **Use cases:** Customer-facing applications, production workflows
**GPT-4.1 Mini** (gpt-4.1-mini)
* **Best for:** Cost-effective production deployments
* **Use cases:** High-volume automation, chatbots, content generation
**GPT-4.1 Nano** (gpt-4.1-nano)
* **Best for:** Maximum efficiency for simple tasks
* **Use cases:** Real-time interactions, simple classification, quick responses
**All GPT-4.1 models support:** Prompts, conversations, classification, intelligence features, agents
**GPT-4o** (gpt-4o)
* **Best for:** Balanced performance and capability
* **Capabilities:** Structured output, multimodal support (text + images)
* **Use cases:** General-purpose applications, document analysis, versatile automation
* **Supports:** Prompts, conversations, email analysis, classification, intelligence, agents
**GPT-4o Mini** (gpt-4o-mini-2024-07-18)
* **Best for:** Cost-effective general-purpose tasks
* **Capabilities:** Structured output, multimodal support, efficient operation
* **Use cases:** Standard automation, customer support, content processing
* **Supports:** Prompts, conversations, email analysis, translation, classification, intelligence, agents
**o3-mini** (o3-mini)
* **Best for:** Latest reasoning-focused tasks requiring deep analysis
* **Capabilities:** Advanced reasoning mode (required temperature: 1.0)
* **Use cases:** Mathematical problems, logical analysis, complex problem-solving
* **Note:** System role not supported; fixed temperature requirement
**o1-mini** (o1-mini)
* **Best for:** Previous-generation reasoning tasks
* **Capabilities:** Reasoning mode (required temperature: 1.0)
* **Use cases:** Logic puzzles, analytical tasks, structured problem-solving
* **Note:** System role not supported; fixed temperature requirement
**When to use:** Tasks requiring explicit step-by-step reasoning, math, logic
**GPT-4 Turbo Preview** (gpt-4-turbo-preview)
* Function calling, extended context
* **Recommendation:** Consider upgrading to GPT-4.1 or GPT-5 series
**GPT-4** (gpt-4)
* Structured output, reliable performance
* **Supports:** Prompts, conversations, summarization, email analysis, classification
**GPT-3.5 Turbo** (gpt-3.5-turbo, gpt-3.5-turbo-1106)
* Function calling, basic capabilities
* **Recommendation:** Upgrade to GPT-4.1 Mini for better performance
### Google Gemini
Access Google's multimodal Gemini models directly.
**Gemini 3 Pro Preview** (gemini-3-pro-preview)
* **Best for:** Latest multimodal capabilities
* **Capabilities:** Multimodal processing (text, images, audio), advanced reasoning
* **Use cases:** Document analysis with images, multimedia processing, complex automation
* **Temperature range:** 0.0 - 1.0 (default: 0.7)
* **Supports:** Prompts, translation, classification, file analysis, agents
**Gemini 2.5 Pro** (gemini-2.5-pro)
* **Best for:** Complex multimodal tasks requiring high performance
* **Capabilities:** Multimodal, large context windows, detailed analysis
* **Use cases:** Document understanding, comprehensive analysis, advanced automation
**Gemini 2.5 Flash** (gemini-2.5-flash)
* **Best for:** Fast multimodal processing
* **Capabilities:** Multimodal, efficient operation, quick responses
* **Use cases:** Real-time document analysis, responsive automation
**Both support:** Prompts, translation, classification, file analysis, agents
**Gemini 2.0 Flash** (gemini-2.0-flash)
* **Best for:** Cost-effective multimodal processing
* **Use cases:** Standard document processing, general automation
**Gemini 2.0 Flash Lite** (gemini-2.0-flash-lite)
* **Best for:** Lightweight multimodal tasks
* **Use cases:** Simple document analysis, high-volume operations
**Both support:** Prompts, translation, classification, file analysis
**Gemini 1.5 Pro** (gemini-1.5-pro)
* **Best for:** Established multimodal performance
* **Capabilities:** Multimodal processing, reliable operation
* **Use cases:** Production workloads, established workflows
* **Supports:** Prompts, translation, classification, file analysis
### AWS Bedrock
Access Claude models through your own AWS Bedrock account. Bedrock-hosted models run within your AWS infrastructure, keeping AI workloads inside your cloud compliance boundaries.
Claude models available through Bedrock mirror the Anthropic Claude family. Availability depends on your AWS region and account access.
* **Best for:** Organizations that need Claude capabilities within their own AWS environment
* **Use cases:** Automations, agents, content generation, classification, reasoning
* **When to use:** AWS compliance requirements, existing AWS infrastructure, data residency needs
**Model Availability**: Available Claude models depend on your AWS region and Bedrock model access grants. Check the Amazon Bedrock console for current availability.
### Custom Providers
Connect any OpenAI-compatible endpoint as an AI provider in Elementum. This includes LLM gateways, proxies, internally hosted inference servers, and self-hosted models. Once configured, custom-provider models appear in the model dropdown alongside built-in providers and can be used in the same agents and automations.
* **Best for:** Standardizing on internal infrastructure, routing through an enterprise LLM gateway, or adopting models that aren't yet supported natively
* **Use cases:** Self-hosted open-source models, proxied access to multiple upstream providers, region- or compliance-specific deployments
* **When to use:** You need a model or routing path that isn't covered by the built-in providers, or your organization requires all model traffic to flow through a managed endpoint
**Capability availability**: Multimodal input, structured output, reasoning mode, and other advanced features are only available if the underlying custom endpoint supports them. Test each model against your use case before relying on a specific capability. See [Configure a Custom Provider](/ai-agents/ai-services#configure-a-custom-provider) for connection details.
## Model Selection Guide
### By Use Case
**Recommended Models:**
1. **GPT-4o Mini** - Best balance of cost and performance
2. **Claude 3.7 Sonnet** - Strong reasoning at reasonable cost
3. **Gemini 2.5 Flash** - Fast multimodal conversations
4. **GPT-5 Mini** - Advanced reasoning for complex interactions
**Why these models:**
* Support structured output for reliable responses
* Handle context well for conversation continuity
* Cost-effective for high-volume interactions
* Consistent reliability in production
**Recommended Models:**
1. **Gemini 2.5 Pro** - Complex multimodal analysis
2. **Gemini 3 Pro Preview** - Latest document understanding
3. **GPT-4o** - Strong multimodal processing
4. **Gemini 2.5 Flash** - Fast multimodal analysis
**Why these models:**
* Multimodal support for images and text together
* Large context windows for lengthy documents
* Strong reasoning for extracting insights
* Handle charts, diagrams, and visual elements
**Recommended Models:**
1. **GPT-4.1 Nano** - Fast, cost-effective
2. **GPT-4o Mini** - Structured output for consistency
3. **Claude Haiku 4.5** - Quick, efficient
4. **GPT-4.1 Mini** - Production-ready reliability
**Why these models:**
* Structured output ensures consistent categorization
* Cost-effective for high-volume operations
* Fast response times for real-time classification
* Reliable accuracy for business logic
**Recommended Models:**
1. **o3-mini** - Specialized reasoning mode for logic
2. **Claude 4 Opus** - Highest-capability reasoning
3. **GPT-5** - Advanced problem-solving
4. **Claude Sonnet 4.5** - Detailed analysis
**Why these models:**
* Advanced reasoning capabilities
* Handle multi-step logic effectively
* Understand complex relationships
* Provide detailed explanations
**Recommended Models:**
1. **Claude Sonnet 4.5** - High-quality writing
2. **GPT-5** - Creative and coherent content
3. **Gemini 2.5 Pro** - Long-form content
4. **Claude 4 Sonnet** - High-quality balanced output
**Why these models:**
* Natural, fluent writing style
* Good creativity control via temperature
* Handle various content types well
* Consistent quality and tone
**Required Models:**
1. **Snowflake Arctic L V2.0** - Latest, highest quality
2. **Snowflake Arctic M V1.5** - Reliable production
**Why these models:**
* Optimized for semantic similarity
* Consistent vector representations
* Efficient processing at scale
* **Note:** Must use Snowflake provider for embedding models
### By Budget
**Lowest Cost Models:**
* **GPT-4.1 Nano**: Minimal cost, simple tasks
* **Claude Haiku 4.5**: Fast and efficient
* **Gemini 2.0 Flash Lite**: Lightweight multimodal
* **Llama 3.2 1B/3B**: Maximum efficiency
* **Mistral 7B**: Small but capable
**Best for:** High-volume operations, simple automation, basic classification
**Best Value Models:**
* **GPT-4o Mini**: Strong cost-performance ratio
* **GPT-4.1 Mini**: Production-ready efficiency
* **Claude 3.7 Sonnet**: Strong reasoning, moderate cost
* **Gemini 2.5 Flash**: Fast multimodal, reasonable cost
* **GPT-5 Mini**: Advanced capabilities, acceptable cost
**Best for:** Production applications, customer-facing features, standard automation
**High-Performance Models:**
* **GPT-5.2**: Highest OpenAI performance
* **Claude 4 Opus**: Highest-capability reasoning
* **Claude Sonnet 4.5**: Latest advanced capabilities
* **GPT-5**: Advanced reasoning and quality
* **Gemini 3 Pro Preview**: Latest multimodal
**Best for:** Mission-critical applications, complex analysis, high-performance features
### By Provider Strengths
**Strengths:**
* Data residency in your cloud
* Wide model selection
* Claude and OpenAI access
* Native data processing
**Choose when:** Data security, Snowflake integration, diverse model needs
**Strengths:**
* Latest GPT models first
* Structured output
* Mature ecosystem
* Reliable performance
**Choose when:** Latest OpenAI features, production performance
**Strengths:**
* Direct access to the full Claude lineup
* Primary provider supported on Studio Agents
* Strong reasoning and natural writing quality
* Simple API-key setup without a data warehouse
**Choose when:** Building Studio Agents, expanding model choice beyond OpenAI/Gemini, or you want direct Anthropic access without Snowflake or Bedrock
**Strengths:**
* Multimodal capabilities
* Large context windows
* Fast processing
* Latest multimodal features
**Choose when:** Document analysis with images, large contexts, latest AI
**Strengths:**
* Claude models within your AWS account
* Data stays in your cloud infrastructure
* Leverages existing AWS security and compliance
* Bedrock Agent orchestration
**Choose when:** AWS compliance requirements, existing AWS infrastructure, data residency within your own cloud
**Strengths:**
* Any OpenAI-compatible endpoint (gateways, proxies, self-hosted)
* Bring models that aren't natively supported
* Route all traffic through internal infrastructure
* Flexible authentication via API key or OAuth Client Credentials
**Choose when:** Standardizing on an internal LLM gateway, hosting your own models, or adopting new models ahead of native support
## Key Model Capabilities
### Multimodal Processing
**What it is:** Process text and images together in the same request
**Supported Models:**
* All Gemini models (2.0+)
* GPT-4o, GPT-4o Mini
**Use cases:**
* Document analysis with charts/diagrams
* Image-based data extraction
* Visual content understanding
* OCR and form processing
### Structured Output
**What it is:** Guaranteed JSON/XML format responses for reliable automation
**Supported Models:**
* All GPT-4o, GPT-4.1, GPT-5 series
* Cortex GPT models
* GPT-4 (partial)
* Llama 3.1 8B, 70B
**Use cases:**
* Data extraction to databases
* Automated classification
* API integrations
* Workflow automation
### Reasoning Mode
**What it is:** Extended thinking for complex problems with step-by-step reasoning
**Supported Models:**
* o1-mini, o3-mini (dedicated reasoning, always on)
* GPT-5 series (configurable)
* Cortex o4-mini (dedicated reasoning)
**Use cases:**
* Mathematical problems
* Logic puzzles
* Complex analysis
* Multi-step problem-solving
**Note:** Dedicated reasoning models (o1/o3/o4-mini) require temperature = 1.0 and don't support system roles
### Agent Support
**What it is:** Optimized for Elementum agent workflows and multi-step tasks
**Supported Models:**
* GPT-4o, GPT-4.1, GPT-5 series
* All Claude models (via Anthropic, Snowflake, or Bedrock)
* Cortex OpenAI models
* Gemini 2.5+, Gemini 3 Pro
**Use cases:**
* Conversational agents
* Multi-turn interactions
* Complex workflows
* Autonomous task execution
## Temperature Settings
All models except dedicated reasoning models support customizable temperature. Temperature is configured at the **LLM Service level** in Organization Settings.
* **0.0 - 0.3**: Deterministic, consistent (classification, data extraction)
* **0.4 - 0.7**: Balanced creativity (conversation, general tasks)
* **0.8 - 1.0**: Creative, diverse (content generation, brainstorming)
**Default:** 0.7 for most models
**Special cases:** o1-mini, o3-mini, Cortex o4-mini require temperature 1.0 (not adjustable)
## Best Practices
### Model Selection
1. **Identify your use case** -- Determine if you need conversation, classification, analysis, generation, or search.
2. **Check required capabilities** -- Verify if you need multimodal, structured output, or reasoning capabilities.
3. **Consider your provider** -- Choose based on data residency, integration, and model access requirements.
4. **Balance cost and performance** -- Select the smallest model that meets your quality requirements.
5. **Test before committing** -- Compare 2-3 models with your actual use cases.
6. **Monitor and optimize** -- Track quality, cost, and speed metrics to refine your selection.
### Cost Optimization
**Choose right-sized models:**
* Use Nano/Mini for simple tasks
* Reserve Pro/Opus for complex analysis
* Test if smaller models meet needs
**Optimize prompts:**
* Write concise, clear instructions
* Remove unnecessary context
* Set appropriate max tokens
* Use structured output formats
**Consider provider costs:**
* Snowflake Cortex models cost \~4.5x base rate (includes infrastructure and data residency)
* Direct provider access may be more cost-effective for high-volume, simple tasks
* Snowflake provides value through data residency and unified platform
### Performance Optimization
**For speed:**
* Use Mini/Nano/Haiku models
* Lower max tokens
* Choose geographically close providers
**For quality:**
* Use Pro/Opus/Sonnet tier models
* Provide detailed context
* Test with real examples
**For consistency:**
* Use low temperature (0.0-0.2)
* Enable structured output
* Choose models with structured output support
## Next Steps
Set up your AI provider connections
Configure specific model instances for your workflows
Create conversational AI assistants using these models
Use AI models in automation workflows