ChatGPT vs Gemini Deep Research: what's actually different in August 2026

"Deep Research" means slightly different things at each company, but the shape is the same: give it a research question, and instead of one chat reply it plans a multi-step investigation, browses dozens to hundreds of sources, and comes back minutes later with a long, cited report. OpenAI introduced the category first in February 2025; Google, Anthropic, and Perplexity have each shipped their own version since. This article compares the standalone chat-assistant versions of the feature. Gemini Notebook (formerly NotebookLM) also has its own Deep Research mode that works only against your notebook's sources rather than the open web; our Gemini Notebook Deep Research guide covers that version specifically.

Current tier limits, tool by tool

This table is the piece almost no comparison gets right, because the numbers keep moving and older articles simply don't get updated. Treat anything you read elsewhere with a publish date before mid-2026 as very likely stale.

Tool                Free tier          Mid tier                    Top tier
------------------------------------------------------------------------------------
ChatGPT              5 lightweight      Plus/Team/Edu: 25 full        Pro: 250 full
Deep Research         runs/month         runs/mo + unlimited           runs/month
                                          lightweight fallback

Gemini Deep          No published       No fixed run count. Google replaced flat
Research              number             daily limits with a rolling 5-hour compute
                                          quota on 17 May 2026, weighted by request
                                          complexity: Free/Standard baseline, AI Pro
                                          ~4x, Ultra $99.99 5x, Ultra $199.99 20x

Claude Research      Not available on   Pro/Max/Team/Enterprise:      Same as mid tier;
                      Free plan          no separate quota, uses       Max gets 5-20x more
                                          standard Claude usage         total usage than Pro
                                          limits

Perplexity Deep      5 runs/day         Pro ($20/mo): ~20 runs/       Max ($200/mo):
Research                                 month (cut from ~500-600/     unlimited
                                          day in an unannounced
                                          early-2026 change)

Sources: OpenAI's own Deep Research documentation; Google's Gemini
Deep Research page and support pages on the May 2026 compute-quota
restructure; Anthropic's support.claude.com Research article;
Perplexity usage-limit tracking via ailimit.watch, corroborated by
Reddit and Trustpilot complaints. Current as of 29 Aug 2026. Neither
Google nor Anthropic publishes an exact numeric Deep Research limit;
re-check in-app before assuming these figures hold indefinitely.

Gemini's row looks different from the others on purpose. Older comparisons, including some published this year, still list a flat "X Deep Research runs per day" number for Gemini, copied forward from before Google eliminated fixed daily prompt and feature counts on 17 May 2026 in favor of the rolling compute quota above, detailed further in our own Google AI Ultra pricing breakdown. If a Gemini comparison you're reading cites a specific daily Deep Research number without mentioning that restructure, it's describing a system Google no longer uses. ChatGPT's numbers trace back to OpenAI's own Deep Research rollout documentation, and it's worth knowing OpenAI split what used to be a single $200 Pro plan into separate $100 and $200 tiers in April 2026, both apparently sharing the same 250-run monthly allowance; check your own account's limit indicator if the exact figure matters for your workflow.

Why Google and Anthropic don't publish a number

This asymmetry is worth calling out on its own, because it shapes how predictable each tool is to actually use. OpenAI and Perplexity both publish an exact run count per billing period, so you always know where you stand. Google's official Gemini Deep Research page describes availability by subscription tier and says reports come back "in minutes," but states no maximum number of runs anywhere on the page itself. Anthropic's help center is explicit that Research "is subject to the same limits as standard Claude conversations," meaning there's no separate Deep Research quota to track at all, just your regular message allowance, which a research-heavy session burns through faster since Claude retrieves and reads multiple sources per turn. If predictability matters more to you than raw capability, that difference alone might decide which tool fits your workflow better than any quality comparison below.

The multimodal question, resolved

A specific claim keeps recirculating in older comparisons: that ChatGPT's Deep Research accepts your own files and images as input while Gemini's is web-only text. That was accurate when it was written, but Google closed the gap at its May 2025 I/O event, adding file and image upload support to Deep Research so it can blend your own documents with open-web research in a single run, the same basic capability ChatGPT already had. If you're reading a comparison that still frames this as a ChatGPT-only advantage, it's describing a gap that closed over a year ago. Both tools now let you upload source material directly; neither publishes an exact per-file size limit for Deep Research specifically, separate from their general chat upload limits.

Perplexity's quota collapse: the story other comparisons miss

Every existing "ChatGPT vs Gemini" piece we found ignores Perplexity's Deep Research entirely, or treats its old limits as still current. That's a real gap, because Perplexity's Pro-tier Deep Research allowance has been through the sharpest cut of any tool covered here. Through most of 2025, Perplexity Pro subscribers could run Deep Research roughly 500 to 600 times a day. In an unannounced change tracked to February 2026, that allowance dropped to around 20 runs a month, a reduction of more than 95%, described in Perplexity's own changelog only as usage limits being "adjusted to allocate more computing power per session." The change generated a visible backlash: "cancel perplexity" and "perplexity pro limits" both spiked in search interest that month, and complaints piled up on Reddit and Trustpilot from users who'd specifically bought Pro for its research capability. As of the most recent tracking, that lower ~20-per-month figure was still the current Pro allowance, with the $200/month Max tier remaining the only unlimited option.

If you're choosing Perplexity Pro today specifically for Deep Research, budget for roughly one run every day and a half to two days across a month, not the many-runs-per-day figure still floating around in coverage that hasn't caught up to the change. If your workflow genuinely needs more than that, Max's unlimited allowance or a competing tool's monthly quota may be the better fit before you hit a wall mid-project.

Which one actually hallucinates less: a real benchmark

Nearly every existing comparison relies on one person's subjective read of a couple of test reports. A May 2026 research paper, "Cited but Not Verified: Parsing and Evaluating Source Attribution in LLM Deep Research Agents", is a more rigorous alternative. It tested 14 closed- and open-source models used in deep research agents against three separate dimensions: whether a cited link actually works, whether the linked content is topically relevant, and whether the cited content actually supports the claim it's attached to, which the paper calls its Fact Check score. That last, strictest dimension complicates the assumption that more citations means a more trustworthy report.

Model family (frontier models tested)    Link works   Relevant content   Fact Check
--------------------------------------------------------------------------------------
OpenAI (GPT-5 Mini, 5.2, 5.4, Codex)       97-100%      87-94%             39-59%
Claude (Sonnet 4.5/4.6, Opus 4.5/4.6,      97-99%       84-96%             52-77%
  Haiku 4.5)                                                                (Opus 4.5
                                                                             highest: 77%)
Gemini (3 Flash, 3.1 Pro)                  ~94%         81-83%             45-49%

Source: "Cited but Not Verified" (arxiv.org/abs/2605.06635), Table 1,
submitted May 2026. Figures are the paper's own model-level testing of
each family's frontier entries, not a live re-run performed for this
article, and reflect the underlying models' research-agent behavior
rather than a test of the exact commercial product wrapper (ChatGPT,
Gemini, or Claude's own app) each company ships. Treat as directional
evidence of a real accuracy gap, not a guarantee about any single
Deep Research run today.

Nearly every model tested gets the mechanics right: links resolve and the content is topically on-target well over 90% of the time across all three families. The gap opens on the strictest test, whether the cited passage actually backs up what the report claims. There, OpenAI's tested models cluster lowest (39-59%), Gemini's sit in the middle (45-49%), and Claude's models score highest, with Opus 4.5 the single best performer at roughly 77%. None of this means one company's product is simply "better"; it means a report with confident-looking citations isn't automatically one you can trust without independently checking a few of those sources, a distinction worth remembering before you cite an AI-generated Deep Research report in something that matters.

Speed and report depth: treat every specific number skeptically

Older comparisons love precise-sounding numbers: one widely cited 2025 test measured Gemini producing a 7,500-word, 55-source report in 12 minutes against ChatGPT's 1,700 words and 38 sources in 17 minutes, and concluded Gemini was both deeper and about 40% faster. Take numbers like that as a snapshot of one run, not a stable property of either tool: Gemini Deep Research users have reported source counts swinging from roughly 10 to over 600 for essentially the same kind of query, with no change to the tool itself. Source count and word count depend heavily on how narrow or broad the question is and how much the model decides a given topic warrants, so they aren't stable, comparable metrics between tools or even between two runs of the same tool on the same day. Anchor your expectations on capability and cost, covered above, rather than on any single number describing report length or run time.

Which tool fits which job

• Need an exact, predictable quota you can plan around → ChatGPT (Plus/Pro) or Perplexity (Free/Max), both of which publish a specific number. Skip Perplexity Pro specifically if Deep Research volume matters, given the 2026 quota cut.
• Already paying for Google Workspace or Gemini, and want Gmail/Drive/Docs context blended into the research → Gemini Deep Research, which pulls in your own Workspace content alongside the open web without a separate subscription.
• Citation accuracy matters more than raw report length, for example anything you'll cite onward in your own work → Claude Research, based on the Fact Check results above, with the tradeoff that Anthropic doesn't publish exact usage numbers or a maximum report length to plan around.
• You're already inside a Gemini Notebook research project and want Deep Research to work specifically against the sources you've collected, not the open web → Gemini Notebook's own Deep Research mode, a different, narrower feature covered in our dedicated guide.

People also ask

Which AI has the best Deep Research feature: ChatGPT, Gemini, Claude, or Perplexity?

There isn't a single winner. ChatGPT and Perplexity publish exact run limits, making them the most predictable to plan around; Gemini blends in your own Workspace content; Claude's model family scored highest for citation fact-check accuracy in an independent May 2026 study, though Anthropic doesn't publish a numeric usage limit. The right choice depends on whether you value predictable quota, Workspace integration, or citation trustworthiness most.

Can Gemini Deep Research use my own uploaded files, or only the open web?

Both. Google added file and image upload support to Gemini Deep Research in May 2025, so it can combine your own documents with open-web research in a single run. Older articles claiming Gemini is web-only are describing a gap that closed over a year ago.

Why did Perplexity's Deep Research suddenly get worse in 2026?

Perplexity cut its Pro-tier Deep Research allowance from roughly 500-600 runs a day to about 20 runs a month in an unannounced change tracked to February 2026, described in its own changelog only as reallocating compute per session. The cut triggered visible backlash on Reddit and Trustpilot from Pro subscribers who'd specifically paid for research volume.

Does more citations in a Deep Research report mean it's more accurate?

Not necessarily. A May 2026 study testing 14 models used in deep research agents found link validity and topical relevance were consistently high (over 90%) across OpenAI, Gemini, and Claude model families, but factual accuracy of the cited content varied widely: 39-59% for OpenAI's tested models, 45-49% for Gemini's, and 52-77% for Claude's, with Claude Opus 4.5 highest at roughly 77%. A confident-looking citation isn't automatically an accurate one.

Is Claude Research available on the free plan?

No. Anthropic's Research feature requires a paid plan: Pro, Max, Team, or Enterprise. It has no separate published usage quota; it draws from the same usage limits as regular Claude conversations, which a research session consumes faster than a normal chat.