ChatGPT Deep Research Now Shares Your Work Budget

Deep Research became available in ChatGPT Work and Codex on September 9, 2026, but the useful change isn't another research button. A research brief in Work now spends from the same Work/Codex allowance or credit pool as the agent work your team uses to build, analyze, and ship.

The change in one sentence

Deep Research now has two different meters inside ChatGPT.

Run it in ordinary Chat and it uses Chat's separate, plan-dependent research-task allowance. Run it in Work or Codex and it uses the allowance or credits already shared by Work and Codex. A Work research task does not take a task from your Chat research allowance, and the reverse is also true.

That distinction is the whole budget story. The same research capability can sit on a different meter depending on where you start it.

OpenAI's September 9 release note says the feature is available to Plus, Pro, Business, Enterprise, and Edu users who have Work access. It runs across web, desktop, iOS, and Android, subject to account and workspace availability.

What Deep Research in Work actually does

Deep Research is for a question that needs several sources, comparison, and a documented result. You give it the outcome, audience, constraints, files, and allowed sources. It can ask clarifying questions, gather evidence, and produce a structured result with citations or source links. You can steer or interrupt the task while it runs.

The Work version matters because the result can become part of a deliverable. You can ask for an editable document, presentation, spreadsheet, or Site when the available tools support that output. In Codex, the same research can sit beside technical work, which makes it useful for a migration brief, dependency review, or implementation decision before code changes begin.

If Work itself is new to you, the ChatGPT Work review explains the difference between Chat, Work, and Codex. The short version is simple: Chat gives you an answer, Work owns a longer deliverable, and Codex works against software and technical systems.

The source boundary does not disappear. Deep Research can use the public web, files you provide, accessible workspace files, web search when available, and supported apps that are enabled and authorized for the current account. Starting a research task does not grant access to another folder, another person's Drive, or an app your workspace has blocked.

That also means an app connection is not a blanket promise. Not every app or plugin supports Deep Research. Even when one does, provider permissions and workspace controls still decide what the task can read or write. The Deep Research guide is explicit about those limits.

The shared budget is the real change

Research and execution now compete for the same agentic capacity in Work and Codex.

For Plus and Pro accounts, supported agentic features use included allowance first, then purchased credits when that option is available. For ChatGPT Business standard seats, included plan limits also come first, and eligible work can continue from purchased workspace credits. Enterprise and Edu workspaces on flexible pricing scale with credits; workspaces without flexible pricing remain subject to their plan limits.

There is no honest universal number of Work research tasks per month. Usage depends on the model, source volume, cached context, output length, reasoning, tools, and any other work running from the pool. Your usage page is the account-specific answer.

The current Business, Enterprise, and Edu ChatGPT rate card makes one easily missed distinction. It lists Deep Research in Chat at approximately 50 credits per task for covered credit-based plans. That fixed Chat row is not the price of Deep Research in Work or Codex. Work and Codex are charged from actual token use.

For GPT-5.6 Sol, the current Work and Codex rates are 100 credits per 1 million input tokens, 10 credits per 1 million cached input tokens, and 500 credits per 1 million output tokens. The formula is:

Credits used = input share + cached-input share + output share.

Consider a clearly hypothetical team budget on a workspace governed by that rate card. A team reserves 1,000 credits for one month and plans 20 research briefs. Assume each brief uses 100,000 input tokens, 200,000 cached input tokens, and 20,000 output tokens on GPT-5.6 Sol.

The input costs 10 credits, the cached input costs 2, and the output costs 10. That is 22 credits per brief. Twenty briefs use 440 credits, leaving 560 credits for every other Work and Codex task in that planning reserve.

Those token counts are an example, not a forecast. A short research memo can use less. A broad brief with many sources and a long editable deck can use more. The useful habit is to budget the research beside execution, then replace the hypothetical inputs with usage from your own first few tasks.

Purchased usage credits are separate from API credits. Buying more for Work or Codex does not fund an API project, and an API balance does not enlarge this pool. The personal credits guide also makes clear that purchase options vary by account, region, and plan.

Who can use this tomorrow

A research lead at a small agency

The lead can ask Work to compare a client's category using approved research sites, uploaded interview notes, and a connected document store. The deliverable is an editable market brief with citations, not another chat transcript.

The payoff is cleaner handoff to strategy and creative. The budget trade is that every deep brief leaves less shared capacity for the decks, spreadsheets, and production tasks the same team runs in Work.

A product manager at a SaaS company

The product manager can combine customer notes, policy pages, competitor documentation, and an internal requirements file into a decision memo. They can steer the task when it overweights one source, then ask for the result as a document the product and legal teams can edit.

The payoff is one inspectable evidence trail. The manager still owns the decision, and every consequential claim still needs a source check before it enters a roadmap or customer promise.

An engineering lead planning a migration

The lead can start Deep Research in Codex to compare current vendor documentation, compatibility notes, and repository constraints before asking Codex to change the code. Research and implementation stay in one working context.

The payoff is fewer context handoffs. The cost consequence is direct: the investigation and the later implementation draw from the same Work/Codex budget, so an oversized research scope can reduce the capacity left to ship the migration.

A workspace admin setting team policy

The admin can give members one rule for where research belongs. Use Chat when the output is a researched answer and the separate Chat task allowance is appropriate. Use Work or Codex when the result must become an editable team artifact or feed execution.

The payoff is a budget people can explain. Enterprise and Edu admins also need the Deep Research permission and web search enabled. Without those controls, eligible members may not see the feature at all.

A safe operating loop

1. 1

   1. Pick the meter on purpose

   Start in Chat for a researched answer. Start in Work for an editable business deliverable. Start in Codex when the research belongs beside technical execution. Do not treat the three entry points as financially identical.

2. 2

   2. Bound the evidence

   Name the sites, files, date range, audience, and decision the report must support. Exclude sources that should not be used. A smaller evidence set is easier to verify and usually consumes less context.

3. 3

   3. Name the output

   Ask for one supported format and say where it should land. An editable output is available only when the task has the right tools and permissions, so include a fallback such as a structured report in the conversation.

4. 4

   4. Steer before it drifts

   Watch the task's progress. If it follows weak sources or widens the question, interrupt and narrow it. More research is not automatically better research.

5. 5

   5. Record the usage delta

   Check the Work/Codex usage view before and after the task. Log the model, scope, output type, credits used, and review time. After a few repeated briefs, those receipts are more useful than a generic task estimate.

6. 6

   6. Check the artifact

   Open the delivered file or link. Sample the citations behind the decision-carrying claims, confirm that each source says what the report claims, and ask for a revision when the evidence or requested output is missing.

The honest limits

Deep Research in Work is not unlimited research. It consumes the existing Work/Codex allowance or credits, and availability still depends on the account, workspace, and surface. A task can stop being a research win if it crowds out the implementation work that follows.

An editable output is not a verified output. The file can be well structured and still cite a weak page, miss a contradiction, or fail to answer the brief. OpenAI tells users to check that citations support the claims and that a requested file or link was actually created and opens.

Teams that only use Deep Research in ordinary Chat are not affected by the new shared Work/Codex meter. Their separate Chat allowance is unchanged. People without Work access, or members whose admins have disabled Deep Research or web search, also gain no new workflow from this release.

What to do now

Act this week if your team already produces a repeated, source-heavy deliverable and can see its Work/Codex usage. Pick one owner, one output, and one approval gate.

Wait if you cannot inspect the sources, cannot see which account or workspace will pay, or need an app that does not support Deep Research. Fix those boundaries before adding a recurring task.

Stay with Chat if you only need a researched answer and do not need the result to become a Work or Codex deliverable. The separate Chat allowance is exactly what that mode is for.

The Monday move

Take one brief your team already writes and run it once, not on a schedule. Record the Work/Codex usage before you start, then use this scoped request:

@Deep Research. Prepare an editable decision memo for our product lead on whether to renew Vendor A. Use only the attached contract, our last two quarterly reviews, Vendor A's current pricing and security pages, and three named alternatives. Separate verified facts from analysis, cite every price and security claim, list unresolved gaps, and recommend renew, renegotiate, or replace. Do not contact any vendor or change any file outside this report.

When it finishes, open the delivered memo and click every citation that carries the recommendation. Confirm the source supports the sentence, check that the requested file exists and opens, then record the usage delta beside the review time. If the citations or the artifact fail that check, the task is not done.

For more plain-English operating guides when AI product math changes, subscribe to the newsletter.