OpenAI Responses API Web Search Tool Adds Image Results

OpenAI quietly shipped a meaningful capability gap between its two major text-generation endpoints on June 9: web search in the Responses API can now return image results alongside text, while the Chat Completions API cannot.

The change is listed simply in the OpenAI changelog: "Web search can now return image results alongside regular text results. Use image search when your application needs current or web-grounded visuals, such as product photos, landmarks, places, events, or visual references."

For AI engineering teams that run requests through a routing layer, the implication is more than a feature flag. Image search availability is now an endpoint-selection factor — not just a model-selection factor.

What Changed in the Responses API Web Search

The web_search tool in the Responses API now returns image result items in addition to the existing text citations and url_citation annotations. OpenAI's guidance is specific about when to use it:

• Product photos and e-commerce visuals
• Landmarks and place images
• Event coverage (conferences, sports, news)
• Visual references for multimodal downstream tasks

The web search tool itself supports three modes depending on the model used:

Image search results are available across these modes. The earlier web_search_preview tool does not receive new controls and does not support image results — migrate to { "type": "web_search" } for access.

Why Endpoint Choice Now Matters for Routing Teams

Before this change, the main routing reason to prefer Responses API over Chat Completions was stateful multi-turn conversation management and background task support. Now there is a second concrete capability gap: only Responses API web search returns image results. Chat Completions with gpt-5-search-api supports text web search for legacy integrations, but not image grounding.

If your agent pipeline needs to:

• Show users web-grounded images in responses
• Anchor downstream image-generation tasks to real-world visual references
• Run multimodal research that synthesizes current visual content

…then the correct endpoint is Responses API, and your routing policy should enforce that explicitly.

A minimal Responses API call with web search:

import OpenAI from "openai";
const client = new OpenAI();

const response = await client.responses.create({
  model: "gpt-5.5",
  tools: [{ type: "web_search" }],
  input: "Show me images of the new Anthropic London office",
});

console.log(response.output_text);
// citations available in response.output[n].content[0].annotations

What to Watch

Response shape changes. Image search results appear as additional items alongside the existing web_search_call action and url_citation annotations. Teams parsing response.output programmatically should handle the new image result item type gracefully — confirm your response-handling code does not throw on unknown output item types.

Tool cost accounting. Each search action in web_search_call incurs a tool call cost per OpenAI's pricing. Image search does not add a separate billing line — it is returned within the same search call — but longer responses with richer image metadata may affect token counts.

Legacy web_search_preview deprecation path. The original web_search_preview tool remains available but does not receive new features including image results, filters, external_web_access, or return_token_budget. OpenAI recommends migrating to { "type": "web_search" } for new integrations.

Routing Decision Framework

Use this matrix when selecting the right endpoint and tool configuration for web-search workloads:

If you route OpenAI requests through a unified gateway, this distinction means routing rules cannot treat all web_search workloads the same. A request that requires visual grounding must be directed to the Responses API path; otherwise the image results will simply not be returned.

Related

• AI Provider Routing — how TheRouter routes requests across providers and endpoint types
• OpenAI API Changelog — official source
• Web Search Guide — Responses API web search reference