Reasoning Models
Reasoning Models

Reasoning Models

Reasoning models represent a significant advancement in how AI agents handle complex, multi-step tasks. When combined with Sema4's agent tuning, these models deliver dramatically improved accuracy and reliability for enterprise workflows.

Available Models

Available Reasoning ModelsProvider
OpenAI o3, o4-mini, GPT-5OpenAI, Microsoft Azure OpenAI
Claude Sonnet 4/4.5 and Opus 4.0/4.1AWS Bedrock, Snowflake Cortex

For full details on models available from each provider, see the full model compatibility matrix below.

Reasoning Levels

You can configure reasoning levels for each agent to balance accuracy and response time. Higher reasoning levels deliver more accurate results but take longer to respond.

LevelUse Case
LowFor faster replies on straightforward tasks with clear logic paths.
Similar to ChatGPT's extended reasoning setting.
MediumFor most enterprise workflows (recommended default)
HighFor complex analytical tasks and worker agents where response latency doesn't matter

Improved Accuracy, Reduced Hallucination

Reasoning models combined with Sema4's agent tuning deliver dramatically better accuracy on complex, multi-step tasks - reducing hallucinations and improving the reliability of your agent outcomes.

These improvements have been validated through τ2-telecom benchmark, a widely-recognized evaluation for measuring agent performance on long-running, multi-step tasks. OpenAI and Anthropic use this benchmark to validate their tool-calling accuracy because it tests an agent's ability to maintain context, follow complex instructions, and complete sequential operations without degradation.

Sema4.AI agent accuracy with τ2-telecom benchmark

Requirements

For OpenAI's GPT-5 and Anthropic's Claude 4 model families:

  • Building agents: Studio 1.6.0 or above
  • Deploying agents: Agent Compute 1.6.0 for Sema4.AI Enteprise Edition, App Version 1.6 for Sema4.AI Team Edition

Using older versions? OpenAI's o3 reasoning model is available with Studio 1.4.6 and Agent Compute 1.4.1. See the full model compatibility matrix below.

Two steps to enable reasoning-powered agents

What to expect with reasoning models: While reasoning models significantly improve accuracy and reduce errors, they also change how your agents approach problems. Agents will provide more thorough, step-by-step responses and may surface edge cases or considerations that standard models miss. This means:

  • Response style changes: Answers become more structured and comprehensive.
  • Timing variations: Complex tasks may take longer but with higher success rates.

Always test reasoning-enabled agents with your actual workflows before deploying to production.

Step 1: Connect your models

Add your reasoning model credentials based on your access level:

  • Control Room access: Add model credentials in Control Room settings
  • Studio access only: Add model credentials in Studio > Settings > Models

Verify your setup by checking Studio → Settings → Models. Your reasoning-enabled models will show available reasoning levels.

If your organization manages AI model access centrally, contact your administrator for model credentials.

Step 2: Enable Reasoning for Your Agents

Create a New Agent with Reasoning Models

  1. Launch Sema4.AI Studio
  2. Click Build AI Agent
  3. Select Conversational or Worker (or import an existing agent)
  4. In the Large Language Model dropdown, select your model and reasoning level (Low/Medium/High)
  5. Assign a name and description
  6. Click Open in Runbook Editor to add instructions
  7. Click Create
  8. Test your agent by sending a message

Update an Existing Agent

  1. Launch Sema4.AI Studio
  2. Go to AgentsMy Agents and select your agent
  3. In the right panel, click Edit Agent
  4. In the Large Language Model dropdown, select your model and reasoning level
  5. Save and test the updated agent

Architecture note: Enabling reasoning models updates your agent to our latest "v2.1" architecture. To revert, go to Advanced Options → select v2.0 Architecture → then reselect your original model.

Full Model Compatibility Matrix

Use this table to determine which version of Sema4.AI products you need for a specific model. The table shows when each model was first supported - newer versions continue to support all previously introduced models unless otherwise noted.

Model
Benchmark Score
τ2-telecom
Build
Deploy
Minimum Studio Version
Minimum Agent Compute
Enterprise Edition
Minimum Snowflake App
Team Edition
OpenAI
GPT-5
Reasoning
95% (high)1
93% (medium)
90% (low)
1.61.4.11.6*
o3
Reasoning
58%21.4.61.4.1N/A
o4-mini
Reasoning
41%31.4.61.4.1N/A
gpt 4.1
Non-reasoning
34%31.31.334%
gpt-4o
Non-reasoning
24%21.21.6N/A
Anthropic
Anthropic calls reasoning "extended thinking"
Claude Sonnet 4.5
Reasoning
98%41.61.61.6*
Claude Opus 4.1
Reasoning
71%41.61.61.6*
Claude Opus 4
Reasoning
57%51.61.61.6*
Claude Sonnet 4
Reasoning
49.641.61.61.6*
Claude Sonnet 3.7
Non-reasoning
4931.4.61.4.11.4
Claude Sonnet 3.5
Non-reasoning
N/A1.21.21.2

For Team Edition users

  • Reasoning models with an asterisk (*) are fully supported via AWS Bedrock, OpenAI, or Azure OpenAI. These models are not yet supported in the Sema4.AI platform through the Cortex API.
  • A timeline update for reasoning models with an asterisk (*) is expected in early November. For current model availability through Snowflake Cortex, see Snowflake's Cortex REST API documentation (opens in a new tab).
Footnotes