Cost and time

The last five lessons made the case that a process can do things a chat cannot. This one is the bill. Not to talk you out of it, but because the people who get the most out of these features are the ones who spend runs like money: freely on the tasks that deserve it, never on the ones that do not.

What you pay, as of August 2026

Prices and tiers change often. This is a snapshot; check the plan page in your app before you decide. The pattern, though, has held for a while: the research feature comes with the first paid tier, the browsing agent comes with the first paid tier or one above it, and everything comes with an allowance rather than unlimited use.

• ChatGPT. Free and the $8 Go tier do not include Deep Research or ChatGPT Work. Plus at $20 a month includes both. The Pro tier, split into $100 and $200 versions in April 2026, is the same features with far more allowance. Deep Research comes with a monthly cap of runs on each paid plan, and Work tasks draw from the same usage pool as OpenAI's Codex coding tool, with heavier tasks using more of it.
• Gemini. Deep Research is available without a paid plan, though Google notes it can be unavailable during busy periods. Paid tiers (AI Plus, AI Pro at $19.99, AI Ultra) get more usage, and as of mid-2026 Google measures that as a shared usage pool that Deep Research draws from faster, not a fixed number of reports. Auto browse in Chrome needs AI Pro or Ultra. Gemini Spark, the always-on agent, is Ultra only.
• Claude. Research and Claude in Chrome come with every paid plan, starting at Pro, $20 a month ($17 on annual billing), with Max at $100 or $200 for much higher limits. There is no separate research quota. A run uses your normal allowance, just faster, so a heavy research day can leave you rate-limited for chat.

If you are on a free plan and curious, the cheapest way in is the first paid tier that lists the research feature in whichever app you already use: Plus for ChatGPT, AI Plus for Gemini, Pro for Claude. Buy the browsing agent tier only after you have a weekly task for it.

The allowance is the real limit

An allowance is the number of runs, or the amount of model work, your plan includes per day or month before you are cut off or dropped to a lighter version. On a $20 plan it is easy to burn through a month of research runs in a weekend of curiosity. Three habits keep you inside it:

• Chat first, research second. Ask the plain chat with web search on. If the answer is good enough, stop. If it is thin or you need citations you can act on, spend the run.
• One brief, written well. A sloppy brief costs a rerun, which is two runs for one answer. Lesson 2's brief format exists to make the first run the only run.
• Batch related questions into one run. "Compare these three" is one run. Three separate "tell me about" runs is three, and the comparison is worse.

How long things take

Rough expectations, so the wait does not worry you:

• Plain chat: seconds. With web search on, ten to thirty seconds.
• Research run: a few minutes for a narrow question, and somewhere between five and thirty minutes for a real comparison. ChatGPT's Deep Research quotes that range, Gemini reports usually land in minutes, and Claude's Research runs typically finish in minutes with hard questions going much longer. Start it, then go do something else.
• Browsing agent: a few minutes for a bounded task, plus however long you take at each stop-and-ask moment. It is rarely faster than you on a single site. It wins on six sites at once, or on tasks you would otherwise put off.
• Long agent jobs (ChatGPT Work, Gemini Spark): designed to run for an hour or more on a job that ends in a document or a scheduled routine. Treat these like assigning work to a person: check in at the plan step, then at the end.

Add your review time to every one of those. Lesson 5's five-minute review is part of the cost of a research run, not an optional extra. A ten-minute run plus a five-minute review is fifteen minutes. Ask whether the question is worth fifteen minutes.

When a plain chat is better

A surprising amount of the time. A chat, with web search on if you need it, wins when:

• You already have the inputs. Rewriting, summarizing, reformatting, translating, tone changes. Nothing to go find.
• The topic is well covered. How a gutter guard works, what a business license costs in Utah in general terms, the difference between two kinds of tint film. The model knows this, and one search confirms it.
• You want a conversation. Thinking out loud, exploring options, asking follow-ups. A research run is a one-shot. A chat is a back-and-forth.
• Speed matters more than depth. You need something in the next two minutes, not the next twenty.

The research run wins when the answer lives on many pages you have not read, when the numbers must be current, and when you need citations you will act on. The browsing agent wins when the job is clicks inside real websites, especially several at once. That is the whole decision, and the explainer below drills it.

A decision table you can keep

Screenshot this, or write it on an index card. Ask the questions in order and stop at the first yes.

• Is everything the AI needs already in my message? Yes: plain chat.
• Would one web search settle it? Yes: plain chat with search on.
• Does the task require clicking inside a website, or a login? Yes: browsing agent, with stops at money and submit.
• Will I act on specific numbers I do not yet have, from sources I have not read? Yes: research run, then the five-minute review.
• Is this a long job that ends in a document, and will I run it again? Yes: a long agent job, and then the Agents module, which is next.
• None of the above? Plain chat. Upgrade only if the answer feels thin.

Try this yourself

For one week, keep a note on your phone with two columns: tasks you sent to a research or agent feature, and tasks you handled in plain chat. Next to each, one word: worth it, or not. Before you start, paste this into your everyday chat and keep the answer handy.

I run a small business in Southern Utah and I am learning when to use AI research and agent features versus plain chat. Here are five tasks I do most weeks:

1. [task]
2. [task]
3. [task]
4. [task]
5. [task]

For each one, tell me: plain chat, plain chat with web search, a research run, or a browsing agent, and one sentence on why. Then tell me which single task is the best first use of a research feature, and write me the research brief for it.

At the end of the week, count the "worth it" marks in each column. Most people find the research column is shorter than they expected and better than they expected. That is the right place to land. You have finished Agentic Processes. You know what the loop is, where each feature lives as of today, how to brief it, and how to check it. The next module, Agents, is about what happens when the loop gets tools, permissions, and a place in your business.