Model & usage
—token cost / query
—search fees / query
—total / query
o4-mini vs o3 — same workload, both tiers
Your token and search profile priced on both deep research tiers. Search fees are identical on both — only the token rates change, so the gap widens the longer your reports get.
| Model | Tokens / query | Search / query | Total / query | Monthly |
|---|
Why deep research queries cost dollars, not cents
Deep research models are priced like ordinary models on paper — dollars per million tokens — but they consume tokens like nothing else in your stack. A single query autonomously runs a research loop: it issues its own web searches, reads the pages it finds into the context window as input tokens, reasons at length about what it found (billed as output tokens even though you never see most of the reasoning), and finally writes a long structured report. A realistic task lands around 50K input tokens, 20K output tokens, and 15 web searches — roughly $0.41 on o4-mini-deep-research and $1.45 on o3-deep-research. Compare that to a typical chat completion costing a fraction of a cent.
The line item most people miss is the search fee. Deep research models bill roughly $10 per 1,000 web search calls on top of tokens, and the searches cannot be turned off — autonomous searching is the product. Fifteen searches is $0.15 per query before a single token is counted, which means short queries have a hard price floor and per-token price comparison tables systematically understate what deep research actually costs. On brief queries the search fees can exceed the token bill entirely.
The practical lever is tier routing. o4-mini-deep-research handles routine research — competitive scans, structured summaries, fact-finding — at roughly 3.5x less per query than o3-deep-research. Route everything to the mini tier first and escalate only where the report quality falls short: at 100 queries a month that single routing decision is the difference between roughly $41 and $145, and it scales linearly from there. If you only need occasional single-source lookups rather than full research loops, a regular model with the web search tool is cheaper still — model that separately with the LLM web search tool cost calculator.
How this calculator works
Free Deep Research API cost calculator — full per-query cost for o3-deep-research and o4-mini-deep-research: input + reasoning output tokens plus the mandatory $10/1k web search fees, modeled together.
Frequently asked questions
How much does one Deep Research API query cost?
A realistic research task — competitive analysis, market research, a literature survey — typically consumes around 50K input tokens, 20K output tokens (including hidden reasoning), and roughly 15 web search calls. On o4-mini-deep-research ($2/M input, $8/M output) that lands near $0.41 per query: about $0.10 input, $0.16 output, and $0.15 in web search fees at $10 per 1,000 calls. The same task on o3-deep-research ($10/M input, $40/M output) costs roughly $1.45 — about 3.5x more. Short queries cost less, but the search fees put a floor under every query regardless of length.
Why can't I turn off the web search charges?
Web search is what makes a deep research model a deep research model — the API performs its own searches as part of the research loop, and those calls bill at roughly $10 per 1,000 searches on top of token costs. Unlike the optional web search tool on regular chat models, deep research models search autonomously and the calls cannot be disabled. On short queries the search fees can be the single largest line item: 15 searches is $0.15 before a single token is billed, which is why per-token price comparisons alone underestimate deep research costs.
Should I use o3-deep-research or o4-mini-deep-research?
o4-mini-deep-research costs roughly 3-4x less per query ($2/$8 per million tokens vs $10/$40) and handles most routine research tasks — summaries, competitive scans, structured fact-finding. o3-deep-research is the flagship tier for genuinely hard synthesis work where analysis depth justifies paying several dollars per query. A common production pattern is routing everything to o4-mini first and escalating only the queries where the cheaper report proves insufficient — at typical workloads that routing decision changes the monthly bill by 3.5x.