Google introduced Deep Research Max on April 21, 2026 as the higher-comprehensiveness version of its new Deep Research agent. The product story is clear enough: it is built for long-horizon, background research work. The pricing story is not. Teams quickly run into three different layers of cost at once: model tokens, search/tool usage, and whichever Google seat plan or platform path they use to access it.
If you are trying to budget Deep Research Max for analysts, research teams, or agent-driven enterprise workflows, the right question is not just “what is the token price?” It is “what parts of a research run are billable, and which pricing path am I actually on?”
The short answer
For API users, Google says the Gemini Deep Research Agent is billed at standard Gemini list rates, including input tokens, output tokens, and intermediate reasoning/input tokens generated during agentic loops. Tool usage is billed separately under normal tool pricing rules.
Because Google says the new Deep Research agents are built with Gemini 3.1 Pro, the practical API benchmark to watch is the Gemini 3.1 Pro Preview pricing tier:
- Standard: $2.00 per 1M input tokens and $12.00 per 1M output tokens for prompts up to 200k tokens; $4.00 input and $18.00 output above 200k.
- Batch: $1.00 per 1M input tokens and $6.00 per 1M output tokens up to 200k; $2.00 input and $9.00 output above 200k.
- Flex: same list pricing as Batch.
- Priority: $3.60 per 1M input tokens and $21.60 per 1M output tokens up to 200k; $7.20 input and $32.40 output above 200k.
Search is a second charge. For Gemini 3.x pricing, Google lists 5,000 prompts per month free for Google Search, then $14 per 1,000 search queries. That matters because a single Deep Research Max job can fan out into multiple searches, not just one.
For enterprise app access, Google’s documentation says Deep Research requires Gemini Enterprise Plus or Agentspace Enterprise Plus. Public pricing pages show Google’s seat-based enterprise plans starting around the low tens of dollars per user per month, but Google does not publish a simple “Deep Research Max costs X per seat” line item. In practice, enterprise buyers usually end up in a bundled seat-plus-platform conversation rather than a clean per-report price.
How the API pricing really works
The most important detail in Google’s pricing language is that Deep Research is billed like an agent, not a one-shot model call. That means the bill is not limited to the prompt you typed and the final report you received.
A Deep Research Max run can generate cost from:
- your initial prompt, attachments, and any prior context,
- intermediate reasoning and planning steps,
- follow-up reads across URLs, files, or connected data sources,
- tool-generated content that gets pulled back into the model context, and
- the final output report.
That is why Deep Research Max can feel inexpensive on a per-token basis but still produce surprisingly large bills for real workloads. The agent is doing more than summarizing. It is planning, searching, reading, refining, and often looping.
A simple budgeting framework
| Cost layer | What gets billed | What teams miss |
|---|---|---|
| Model inference | Input, output, and intermediate reasoning/input tokens | The hidden loop tokens can be much larger than the initial prompt |
| Search | Google Search queries after the free monthly allotment | One research job can trigger multiple searches |
| Context size | Higher token costs when prompts exceed 200k tokens on Gemini 3.1 Pro | Large source packs or long retrieved context can move the run into the pricier tier |
| Other tools | URL context, file search, and related tokenized retrieval flows | Retrieved material still becomes billable model context in many cases |
The enterprise seat story is separate from the API story
Teams often mix up three different Google routes:
- Gemini API / Google AI Studio for developer-driven, token-metered usage.
- Gemini Enterprise for workplace and app-side access inside Google’s enterprise platform.
- Agentspace for broader agent deployment, search, connectors, and enterprise workflows.
That distinction matters because the buying motion changes. If you are embedding Deep Research Max into a product or internal workflow engine, the API path is usually the cleaner budget model. If you want employees to use Deep Research as a built-in enterprise agent, you are more likely dealing with seat licenses, storage/indexing entitlements, and Google sales packaging.
Google’s documentation says Deep Research requires a Plus-tier enterprise edition in the app environment. Public pricing pages show Business plans starting lower and Standard/Plus enterprise plans starting higher, but Google does not publicly break out a universal per-user list price specifically for the Deep Research agent. That means many enterprises should assume the app-side price is best treated as platform access cost, while report-level usage costs still need operational monitoring.
What makes Deep Research Max feel expensive fast
There are four common bill drivers.
1. Broad research scope
If you ask for a complete market map, competitive teardown, regulatory overview, and synthesis memo in one run, the agent may perform many more searches and retrieval passes than a lightweight research query.
2. Huge source packs
Uploading PDFs, spreadsheets, slides, transcripts, and large linked materials can push the run into the more expensive prompt tier on Gemini 3.1 Pro.
3. Background quality expectations
Max is explicitly positioned as the higher-comprehensiveness option. That is usually code for more compute, more searching, and more iteration. It is often the right tool for executive-grade reports, but it is the wrong cost profile for lightweight interactive Q&A.
4. Search-heavy workflows
Google charges search separately after the free allotment. If your team uses Deep Research Max for recurring due diligence, vendor monitoring, or sector tracking, search charges can become meaningful before token costs do.
How teams should budget it
The best way to budget Deep Research Max is to stop pretending every run is the same. Break usage into three buckets:
- Light: short internal research jobs with limited web/tool activity,
- Medium: multi-source reports that combine internal material with moderate search depth,
- Heavy: background, analyst-grade jobs with iterative search, synthesis, and large source packs.
Then model budget against reports per user per week, not just seats or token rates. A team that runs two high-value background reports per analyst each week can justify Max easily. A team that tries to use Max for every exploratory question will likely overspend.
A practical rule: use the faster standard Deep Research path for routine work, and reserve Deep Research Max for the jobs where comprehensiveness, synthesis quality, and defensibility matter enough to justify agent-loop costs.
The real takeaway
Deep Research Max is not expensive because Google put a flashy premium label on it. It is expensive when teams use a long-horizon research agent like a cheap chat completion. The underlying cost model reflects what the system actually does: plan, search, read, reason, and refine.
For product teams, the key budgeting question is API efficiency and query discipline. For enterprise buyers, the harder question is whether Google’s seat-based platform access plus report-level consumption creates a better operating model than a rival research agent. Either way, the headline model rates are only the beginning. The real bill is shaped by agent depth, search intensity, and context volume.