Skip to main content
Date: March 4, 2026

​
What’s New

  • Greatly improved performance for Exa Deep workloads.
  • Latency ranges:
    • deep: 4 to 12 seconds
    • deep-reasoning: 12 to 50 seconds
  • New search type: deep-reasoning for higher-effort reasoning tasks.
  • 20% lower price for regular deep search.
  • Structured outputs for Deep using output schemas. Built to provide web grounded fields perfect for enrichments.
  • New request schema controls: type, description, and properties on outputSchema (output_schema in Python SDK).
  • New synthesized response fields: output.content and output.grounding.

​
Structured Outputs

For deep search variants (deep, deep-reasoning), you can define the output shape using:
  • outputSchema in JavaScript/TypeScript and REST
  • output_schema in Python
Supported schema controls include:
  • type
  • description
  • properties
{
  "query": "top aerospace companies",
  "type": "deep-reasoning",
  "outputSchema": {
    "type": "object",
    "properties": {
      "companies": {
        "type": "array"
      }
    }
  }
}

​
Deep Output Response Shape

Deep responses now include:
  • output.content for synthesized text or structured JSON
  • output.grounding for field-level citations and confidence
{
  "output": {
    "content": {
      "companies": ["SpaceX", "Boeing"]
    },
    "grounding": [
      {
        "field": "companies[0]",
        "citations": [{ "url": "https://example.com", "title": "Example Source" }],
        "confidence": "high"
      }
    ]
  }
}

​
Pricing Update

Regular deep search is now 20% cheaper:
  • deep: $12 per 1,000 requests
  • deep-reasoning: additional $3 per 1,000 requests (total $15 per 1,000 requests)
See full pricing details in Exa Pricing Update.