Gemini Deep Research: old agent shuts down, cost and migration

Gemini Deep Research is switching agents in the Gemini API: on October 8, 2026 Google deprecated the December 2025 agent and it shuts down on October 23. The notice is about the API, the one developers use, not the Gemini app. The replacements have existed since April 21: Deep Research and Deep Research Max.

What shuts down and what does not

Google's changelog deprecated the December 2025 agent, and it shuts down on October 23. For the Flash models announced that day, Google did promise automatic redirection; for this agent, it did not. If your code uses the old name, assume it fails after the 23rd. If you only use Gemini in the browser, this notice does not affect you.

How it works: Interactions API and background runs

These agents only live in the Interactions API, not in the normal generate-content call. You launch them in the background, get an ID back, and poll until the status moves from in progress to completed or failed. Background requires that the interaction be stored. A report can take up to 60 minutes; most should finish in 20. It is not a chat: it is a job you come back for. On the same thread you can keep asking by citing the previous ID.

The full walkthrough, step by step

1. Create the interaction: in the Interactions API you set the agent name: Max for depth, the standard one for speed. Add your question and turn on background and store.
2. Turn on collaborative planning: instead of the report, the agent returns a research plan. You adjust it with new messages and, once you approve it, you turn the flag off and it starts.
3. Pick tools: by default it uses Google search, page reading and code execution. You can add MCP servers or your own documents, or turn the web off.
4. Ask for automatic visualization: the report arrives with charts.

Where it snags: streaming drops, so save the ID and the last event to resume. How to know it came out right: open the citations behind the figures that carry the conclusion.

The math: what each report costs

According to the BLS, the US Bureau of Labor Statistics, there were almost 953,000 market research analysts in 2025, with median pay of $78,760 a year ($37.87 an hour). Google estimates $1 to $3 per standard report and $3 to $7 with Max. Those are preview estimates. Seven divided by 37.87 is about 11 minutes of an analyst.

WAiBOT assumption, not Google's: 20 reports a month, 4 hours each by hand. That is 80 hours, about $3,029. With Max, up to $140, plus half an hour of review per report (10 hours, $379): $519 in total. Hypothetical saving: about $2,500 a month, as long as a report saves you more than 11 minutes.

Three business cases (WAiBOT analysis)

• Competitor study: a one-person consultancy sells studies to local businesses. Input: the business type, the city and five competitors. With collaborative planning, three minutes fixing the plan and the standard agent, out comes a cited report with a chart. What was half a day becomes half an hour of review, at an estimated $1 to $3. You charge for judgment, not for searches.
• Prospect briefs overnight: a B2B sales team leaves 20 accounts in the queue. Google pitches that use for Max (overnight due diligence reports). Worst case for that batch: $140. A brief is worth what its question is worth: ask what changed in the last six months, not everything about the company.
• Reports from your own documents only: an accounting firm turns the web off and connects the agent to its clients' documents to see which clauses changed between two versions. Google says it works with FactSet, S&P Global and PitchBook on MCP servers, but those are ongoing collaborations, not launched integrations.

How to apply it in your business? Five-step migration checklist

1. Search your code for the December agent's name, before the 23rd.
2. Switch to the standard one if your app shows results live, or Max if the report runs overnight.
3. Turn on background and store, and keep every interaction ID.
4. Test five real questions with collaborative planning and work out your cost from your bill, not the range.
5. Ask it to flag missing data and check the five key figures in every report.

Do not scale to a hundred reports a day on a $1 to $7 estimate: those are estimates. Measure your first ten bills. And for another Google work announcement, read Gemini agent, Google's AI coworker with its own email and Drive.

Who it is not for

• Anyone who needs an answer in seconds: it takes minutes, maybe an hour.
• Anyone who needs structured output, like clean JSON: the documentation says it is not supported.
• Anyone who depends on custom functions: also unsupported.
• Anyone who only uses Gemini in the browser: this notice is about the API.

The objection: a well-dressed hallucination?

A $3 report with citations can look like a well-dressed hallucination, and in one part that is right: a citation does not prove the agent read the source correctly. But before, the draft cost hours and verification was blind; now the draft costs minutes and verifying fits. Check the five figures that carry the conclusion and open their sources. Google also recommends telling the agent to mark what it cannot find as unavailable, instead of estimating. The objection wins when a legal or financial decision rests on a report nobody read: then it is a draft, never a deliverable.

Where the claim that Max is better comes from

Google shows it with win-rate charts against the December version, from an internal expert evaluation, with no exact numbers. The only figure I found, 93.3% on DeepSearchQA, comes from a third-party site, not Google. To believe it, run your own test: 20 real questions, the same prompt, compare the old agent against the new one blind. And the $1 to $7 range is a preview estimate: the pricing page talks about tokens plus tools and does not even mention Max.

What breaks in practice

• Long connections: streaming drops and you have to resume.
• One developer reported on Google's forum that Max sat in progress for minutes and failed with a 500 error while the standard one finished. It is one report with no official confirmation.
• Multimedia inputs raise the cost and can overflow the context.
• The limit of 60 minutes per task.
• If you do not migrate, the first error is the old name, after the 23rd.

Frequently asked questions

When does the old Gemini Deep Research agent shut down? October 23, 2026; Google deprecated it on October 8.

Does it affect the Gemini app? No. The notice is about the Gemini API.

How much does a report cost? According to Google's preview estimates, $1 to $3 with the standard agent and $3 to $7 with Max. The pricing page does not mention Max.

How do I migrate? Change the agent name in the Interactions API, turn on background and store, and test with real questions.

Conclusion

The news is not that an agent shuts down, but that research now has a price per report and it fits on a receipt. The cost of researching dropped to minutes and dollars; what did not drop is the cost of being wrong with a report nobody read. Tell us: which report would you order every Monday, and which figure would you let a human check? Sources: Gemini API release notes, Deep Research documentation and Gemini API pricing.