search
Query live web information with a single AI model to get current answers and sources. Includes query tips for accurate results.
Instructions
Web search with a single model. Use for real-time lookups. For multi-model parallel research, use the research tool.
Query tips (major accuracy improvement):
NEVER include specific years (e.g. "2024", "2025") — use "latest", "current", or "newest" instead. The search engine returns recent results automatically; hardcoded years risk missing newer content or retrieving outdated info
One topic per query — split broad questions into multiple calls
Prefer English for technical topics (official docs are mostly English). Use Japanese when searching for Japan-specific services or local information
Include exact terms: version numbers, API names, config keys
Yes/No form for fact checks: "Does X support Y?" prevents false assertions
Add site: prefix to target official docs
Model guide ($ = cheapest, $$$ = most expensive):
perplexity/sonar ($): Extremely cheap (~1/20 cost of others). Fastest. Best for single-fact lookups and GitHub Issue searches
openai/gpt-5.4-nano ($$): Latest gen nano with web search, cheaper than gpt-5-mini ($0.20/$1.25 vs $0.25/$2). Search quality comparable to gpt-5-mini for simple lookups, concise output
openai/gpt-5-mini ($$): More thorough/comprehensive search results. Best for deep technical docs — migration guides, changelogs, CVE details
google/gemini-3-flash ($$ default): General queries, broad coverage, good all-around
openai/gpt-5.4-mini ($$$): Strong reasoning + search ($0.75/$4.50). Concise, high-confidence answers. Better cost-performance than claude-haiku-4.5
anthropic/claude-haiku-4.5 ($$$): Last resort when other models lack needed info. Most expensive search model ($1/$5) — only use for a different perspective
For parallel multi-model results or diverse perspectives, use the research tool.
Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| model | No | Search-capable model. Default: google/gemini-3-flash ($$ general queries, broad coverage). For quick single-fact lookups: perplexity/sonar ($ extremely cheap, fastest). For cheap search: openai/gpt-5.4-nano ($$ cheaper than gpt-5-mini, latest generation). For official doc precision (migration guides, changelogs): openai/gpt-5-mini ($$). For stronger reasoning in search: openai/gpt-5.4-mini ($$$). Last resort for a different perspective: anthropic/claude-haiku-4.5 ($$$, similar quality to $$ models but significantly more expensive). Also supports other anthropic and openai models with native search. | google/gemini-3-flash |
| query | Yes | Search query in natural language with detailed context. Use English for technical topics (most docs are in English), Japanese for Japan-specific info. One topic per query, include specific technical terms (version numbers, API names, config keys) for best accuracy. IMPORTANT: Never include specific years (e.g. "2024", "2025") in queries — use "latest", "current", or "newest" instead. Year numbers become outdated quickly and may miss newer results. See tool description for full query tips and model guide. | |
| max_tokens | No | Max output tokens. If set, output is hard-truncated at this limit (may cut off mid-response). Omit to let the model decide output length naturally. Only set when you need strict cost control. Reasoning models consume tokens internally, so set 2x-3x higher than expected visible output | |
| max_retries | No | Max number of retries when search result is poor (empty or unhelpful response). Default: 1. Set to 0 to disable retries. | |
| include_sources | No | Include source URLs in the response. When true, appends a Sources section with links at the end |