Deep Research Tools Are Now Free. Here Is How That Changes Everything.
Three of the four major AI labs now ship a deep research agent, and three of them are free to use on day one. The fourth – Anthropic’s Claude Research – is bundled into the Pro tier that everyone evaluating the space is already paying for. If you have not re-tested the deep research tools you were using last year, you are quietly behind the curve. The April 2026 update cycle dropped the cost of frontier-class research to zero for most practitioners, and it changed what the tools are actually good for. Here is what changed, what is real, and what to use this week.
The deep research category arrived in sixteen months – and you can now use every major one for free
The deep research category is barely sixteen months old. Gemini shipped the first Deep Research agent in December 2024, originally powered by Gemini 1.5 Pro. ChatGPT followed on February 2, 2025 with an o3-powered version that hit 26.6 percent on Humanity’s Last Exam, a benchmark designed to be unsolvable by pattern matching. Perplexity Deep Research launched on February 14, 2025 and was free for all users from day one. Claude Research joined the category in late 2025.
The category pattern converged fast because the underlying architecture is the same across vendors: a planning agent that breaks a question into sub-questions, runs dozens or hundreds of searches, reads full pages, follows citations, updates its plan as it learns, and produces a structured report with footnotes. What differs is the depth of synthesis, the speed, the price, and how the tools fail.
In April 2026, two things changed simultaneously. First, Gemini launched Deep Research Max on April 21, 2026 – a new tier of agentic research powered by Gemini 3.1 Pro that supports MCP, native visualizations, and File Search. Second, Perplexity quietly merged Deep Research into its Computer orchestration system on June 11, 2026, jumping the model’s score on Humanity’s Last Exam from 36.4 percent to 50.5 percent in a single update. These were not marketing refreshes. The agents got materially more capable in a quarter.
The free tier is no longer a demo. Google’s 9to5Google breakdown of the April 2026 plan tiers shows the Gemini app on the free plan now allows up to 5 Deep Research reports per month with full Workspace integration. Perplexity’s free tier allows 5 deep research queries per day. ChatGPT’s free tier gets the o4-mini lightweight version. For most journalists, students, and analysts who used to pay for a research subscription, the free tier of any one of these tools now does the work that the paid tier did a year ago.
What changed in April 2026: Gemini 3.1 Pro Deep Research Max and the agentic research tier
The April 2026 release cycle was the moment the deep research category became agentic infrastructure rather than a single-product feature. Google’s launch of Deep Research Max on April 21, 2026, announced on the official Google blog, introduced two distinct agents: a speed-optimized Deep Research for interactive user surfaces, and a comprehensiveness-optimized Deep Research Max for automated context gathering and synthesis. Both run on Gemini 3.1 Pro, support MCP servers, generate native visualizations, and integrate with File Search for private document sets.
The technical shape matters. The Gemini Deep Research agent is exclusively callable via the Interactions API – the developer-facing async interface. Each query runs in background mode, polls for completion, and returns a multi-page cited report. The model card for deep-research-max-preview-04-2026 documents a 1,048,576-token input context window (1M tokens) with a 65,536-token output limit. For deep competitive landscape analysis or extensive due diligence, the agent might use up to 160 search queries and 900k input tokens per run.
The pricing repositioned accordingly. Google AI Plus at $19.99 per month increases Deep Research to 12 reports per day. AI Pro and AI Ultra add higher caps, Gemini 3.1 Pro access, and the Deep Think reasoning mode. The April 11, 2026 update of the Google One plans confirms that the free tier is now the entry point rather than a teaser.
What this means for a practitioner who has been using ChatGPT Deep Research: the April 2026 Gemini update raised the bar on what a free tier can do. The Gemini free plan now delivers comparable benchmark accuracy to ChatGPT Plus on most queries, with Workspace integration that ChatGPT cannot match. If your research depends on pulling from your own Gmail, Drive, or Docs, Gemini is no longer a fallback option.
The benchmark landscape: Humanity’s Last Exam, citation accuracy, and what the scores actually measure
The benchmark that everyone uses to compare deep research tools is Humanity’s Last Exam (HLE) – a set of 2,500 questions across more than 100 academic subjects, designed to be unsolvable by pattern matching. The current scores, verified from the arXiv paper and the April 2026 shootout, cluster tightly:
| Model | HLE Accuracy | Notes |
|---|---|---|
| GPT-4o | 3.3% | No browsing, no tools |
| Claude 3.5 Sonnet | 4.1% | No browsing |
| Gemini Thinking | 6.2% | Reasoning without browsing |
| Perplexity Deep Research (legacy) | 21.1% | First public DR release |
| OpenAI Deep Research (o3) | 26.6% | Feb 2025 original |
| Gemini Deep Research (April 2026) | 54.6% | Powered by Gemini 3.1 Pro |
| Perplexity Deep Research on Computer | 50.5% | June 11, 2026 upgrade |
The headline number on Google’s Deep Research Max – 54.6 percent – is the new state of the art on HLE, and the gap to OpenAI’s 26.6 percent is the headline the marketing team is going to lean on. But the benchmark number does not capture the workflow gap. HLE measures accuracy on hard academic questions. Real research is mostly answering the easier 80 percent of questions faster, with citation reliability on the answers.
A separate correction matters. The HLE-Verified paper from February 2026 found that 689 of the original 2,500 HLE items are noisy – meaning they have wrong problem statements, wrong reference answers, or both. After expert review and revision, the corrected benchmark yields 7 to 10 percentage points higher accuracy across all frontier models, with gains of 30 to 40 percentage points on the items where the original answer key was wrong. The implication: all the HLE scores you have seen quoted in vendor blog posts are slightly inflated by benchmark noise. The corrected scores are uniformly higher, but the relative ranking is preserved.
What HLE does not measure is citation accuracy. The deep research tools all claim to produce cited reports. Whether the citations are real is a separate question – and a deeply uncomfortable one.
The catch nobody is talking about: deep research agents hallucinate at 2x the rate of basic search models
The most important finding about deep research agents in 2026 is the one nobody puts on the marketing page. A 2026 multi-model study comparing hallucination rates found that purpose-built research agents hallucinate at 10.7 percent, more than twice the 4.8 percent rate of basic search-augmented models. The reason is structural: deep research agents synthesize across many sources, and each synthesis step is an opportunity to introduce or propagate errors.
The empirical evidence is damning. The ICLR 2026 peer review process, audited by GPTZero, scanned 300 of the 20,000 submissions and found 50 with at least one hallucinated citation – a reference to a paper that does not exist. Several of those papers had already passed review by three to five expert reviewers. Some had average reviewer ratings of 8 out of 10. Peer review was not designed to catch this, and the reviewers did not catch it.
The failure modes cluster in three patterns. The first is fabrication: the citation points to a paper that does not exist, often with plausible-looking metadata. The second is metadata corruption: the paper exists but the title, authors, date, or conclusions are wrong. The third is citation drift: the paper exists, the citation is technically correct, but the cited findings do not actually support the claim being made. A 2026 medical research test found that of 115 ChatGPT-generated citations, 47 percent were entirely fabricated and another 46 percent were real papers with wrong metadata.
The CiteTracer framework, published in 2026, formalizes this into a 12-code taxonomy that distinguishes real citations, potentially ambiguous ones, and confirmed hallucinations. The framework was built using a benchmark of 2,450 synthetic citations and 957 real-world fabricated citations drawn from ICLR 2026 desk-rejected submissions. The point is not that deep research is useless – it is that the output of every deep research agent requires verification of every load-bearing citation before the report can be trusted.
For practitioners, this changes the workflow. Running a deep research query produces a draft, not a deliverable. The verification step – checking the citations, confirming the cited findings match the claim – is the step the marketing does not mention. On a 10 to 30 page deep research report, that verification step can take longer than the original query.
The free tier in 2026: what you actually get on Gemini, ChatGPT, Perplexity, and Claude
The free tiers in April 2026 are not the same as the free tiers in 2025. The pricing tiers across the four major vendors, verified from official sources, look like this:
| Vendor | Free tier | Pro tier | Quota on free | Best for |
|---|---|---|---|---|
| Google Gemini | Gemini app free | AI Plus $19.99/mo | 5 reports/month | Workspace-heavy work, Gmail/Drive/Docs context |
| OpenAI ChatGPT | Free ChatGPT | Plus $20/mo, Pro $200/mo | o4-mini lightweight, 5/mo | Deepest reports on commercial/policy topics |
| Perplexity | Perplexity free | Pro $20/mo | 5 deep research queries/day | Journalism deadlines, fastest reports |
| Anthropic Claude | Claude.ai free | Pro $20/mo | None – DR is paid-only | Careful reasoning, fewer hallucinated cites |
The Perplexity free tier is the most generous by raw count. Five deep research queries per day, no account required to start, runs in 90 seconds to 4 minutes. The Gemini free tier is more generous by depth – five reports per month at the same depth as paid ChatGPT Plus, plus Workspace integration that the others cannot match. The ChatGPT free tier is the most constrained – the o4-mini lightweight version is fast but does not do the synthesis depth of the paid version.
Claude Research is the outlier. Anthropic has not shipped a separate “Deep Research” product. Claude Research is a mode within Claude.ai that uses the underlying Claude model (Sonnet 5 or Opus 5 depending on tier) with extended thinking and browsing. The Pro tier includes it; the free tier does not. The Tradeoff: Claude Research tends to produce fewer hallucinated citations in informal testing, but it is the only one of the four that is not available for free at any tier.
For a practitioner choosing which free tier to commit to, the right answer depends on the corpus. If your research depends on Google Workspace, the answer is Gemini. If your research depends on speed and you want the most generous free quota, the answer is Perplexity. If you need the deepest reports on commercial topics and do not mind paying twenty dollars a month, ChatGPT Plus is the right answer. If you need careful reasoning and will pay for it, Claude Pro is the right answer. The choice is not which tool has the best benchmark. The choice is which tool’s free tier matches the work you actually do.
Picking the right tool: when free is enough, when Pro is worth it, and when the API pays for itself
The cost calculus breaks into three tiers. Tier one: free is enough. If you run fewer than five research queries per month, and they are general-purpose topics rather than domain-specific synthesis, the free tier of any vendor is enough. You will not notice the difference between Gemini free and ChatGPT Plus for casual research. Tier two: Pro is worth twenty dollars a month. If you run more than five queries per month, if you need consistent synthesis depth, or if you need a specific tool’s Workspace integration, the Pro tier pays for itself in time saved. Tier three: the API pays for itself. If you are building a product on top of deep research, if you are running queries at volume, or if you need the Sonar Deep Research API at $0.30 to $1.32 per query for custom use cases, the Perplexity API or the Gemini Interactions API is the right path.
The Perplexiy Sonar Deep Research API pricing, verified from the official docs in June 2026, is the most transparent of the four. Input tokens $2 per million, output tokens $8 per million, plus $2 per million citation tokens, $3 per million reasoning tokens, and $5 per 1,000 autonomous searches. A single Sonar Deep Research query typically runs $0.30 to $1.32 depending on context depth. The Gemini Interactions API follows a similar structure, billed per-token plus per-search. ChatGPT does not expose a deep research API – it is a ChatGPT-only consumer product. Anthropic exposes the underlying Claude API but not a dedicated research agent.
For developers building products, the practical comparison is Perplexity Sonar Deep Research versus the Gemini Deep Research Max agent via the Interactions API. Both have callable APIs. Both charge per query. The choice depends on whether you need Workspace integration (Gemini) or speed and citation density (Perplexity). The detailed worked example for the Perplexity pricing is available at the Spheron Network analysis, which shows that for a typical 300 input token, 400 output token query at medium context, Sonar Pro runs $0.0169 per query, while a full Sonar Deep Research query runs $0.30 to $1.32 per query depending on search depth.
For a consumer choosing between the four, the rule of thumb from the April 2026 shootout is consistent: Perplexity for speed and free-tier generosity, OpenAI for depth on commercial topics, Gemini for Workspace integration and the highest HLE benchmark score, Claude for careful reasoning on hard questions. No one tool wins on all four axes.
How to actually use a deep research agent without burning an hour on a bad prompt
The most common failure mode with deep research agents is producing a 15-page report on the wrong question. The agents are optimizers – they will happily write 15 pages on whatever you asked. If you asked a vague question, you get a vague 15 pages.
The pattern that works, verified across all four vendors, is five components in the prompt. First, the role: tell the agent what kind of expert you want it to be. “You are a senior market analyst covering enterprise SaaS” works better than “research this.” Second, the scope: define what is in and what is out. “Cover the US and EU markets only” or “limit to 2024 through 2026.” Third, the deliverable shape: “produce a report with an executive summary, five to seven sections, and a one-page comparison table.” Fourth, the citation requirement: “every claim must have an inline citation with a working URL.” Fifth, the verification hook: “for any claim that cannot be independently verified, mark it as such in the report.”
The output of a well-prompted deep research agent is a draft with citations. The verification step is not optional. For every load-bearing claim in the report – any fact you would be willing to be quoted on, any number you would put in a deck – check the citation. The 10.7 percent hallucination rate is the floor. If a citation does not resolve, if the cited paper does not exist, or if the cited paper does not actually support the claim, the report is wrong in that section.
The 2026 ICLR 2026 case study confirmed that this verification step is not what most users actually do. The peer reviewers at the conference level missed hallucinated citations at scale. The implication is not that the tools are bad. The implication is that the verification step is real work, and the cost of skipping it is publishing incorrect research. The deep research tools do not eliminate the work of research. They eliminate the work of finding sources. The work of validating sources remains.
Practical recipe: a five-step research workflow that combines free tiers with verification
The workflow that survives the hallucination problem is to use multiple tools and triangulate. Start with Perplexity Deep Research free tier for the first pass – five queries per day, fastest turnaround, gives you a draft. Run the same query through Gemini Deep Research free tier for the second pass – 5 reports per month, deeper synthesis, gives you an independent synthesis. Compare the two reports. Where they agree, you have a reliable finding. Where they disagree, you have a finding that needs another tool. Use ChatGPT Deep Research on the third disagreement to break the tie – this requires Plus, but is worth it for any research you will publish.
Step one: scope the question. Write down the deliverable, the audience, and the verification criteria. Step two: run the question on Perplexity Deep Research free. Read the output. Mark every claim that you will need to verify. Step three: run the same question on Gemini Deep Research free. Compare the two outputs. Step four: for any disagreement, run on ChatGPT Plus or the Gemini Interactions API for an independent third opinion. Step five: verify every load-bearing citation. The 10.7 percent hallucination rate is the floor, not the ceiling. Citations to arXiv papers are the most reliable; citations to vendor blogs are the least reliable; citations to news articles are somewhere in between.
The pattern that fails is the opposite: run one tool, take the output, publish. The 2026 evidence is clear that this produces hallucinated citations at scale. The pattern that works is to treat the deep research agent as a research assistant who needs supervision, not as an oracle.
FAQ
Is ChatGPT deep research still free in 2026?
Free ChatGPT accounts get a lightweight o4-mini version of Deep Research. The full GPT-5.2-backed Deep Research with browsing requires Plus ($20/mo) or Pro ($200/mo). Plus accounts get 10 to 25 queries per month; Pro gets effectively unlimited.
How accurate is Gemini Deep Research compared to ChatGPT?
Both are roughly comparable on benchmark accuracy within 5 points on Humanity’s Last Exam, but they fail differently. Gemini’s strength is Workspace integration (Gmail, Drive, Docs). ChatGPT’s strength is depth of multi-source synthesis on commercial topics.
What is the best free deep research tool?
Perplexity Deep Research on the free tier is the only fully unbounded free option – 5 deep research queries per day, no account required to start. Gemini free gives 5 reports per month with Workspace integration. ChatGPT free gives an o4-mini light version.
How long does a deep research query take?
Perplexity: 90 seconds to 4 minutes. Gemini: 5 to 15 minutes. ChatGPT: 5 to 30 minutes. The difference is roughly proportional to depth of synthesis – faster tools give you a brief, slower tools give you a 10 to 30 page report.
Can deep research agents hallucinate citations?
Yes, and the rate is worse than basic search. The 2026 multi-model study found 10.7 percent hallucination in deep research outputs versus 4.8 percent for basic search-augmented models – a 2.2x higher rate. The ICLR 2026 peer review process alone caught 50 hallucinated citations in 300 sampled papers.
The bottom line
The deep research category is mature. Every major lab ships one. Three are free at some tier. The April 2026 updates – Gemini 3.1 Pro Deep Research Max and Perplexity on Computer – raised the bar on what the free tier can do. The 10.7 percent hallucination rate is the cost you pay for the speed. Pick the tool that matches the work you do, run the verification step on every load-bearing citation, and treat the agent’s output as a draft rather than a deliverable.
For a deeper practitioner walkthrough of how to set up a research workflow that combines these tools, the Mr. Technology network’s Deep Research skill guide covers the prompt patterns and verification recipes in more detail than fits here.
For practitioners evaluating which underlying model to combine with your deep research workflow, the Claude Opus 4.7 vs GPT-5.4 vs Gemini 3.1 Pro 2026 frontier benchmark covers the model-level comparison that the deep research tools sit on top of. For the broader cost question of running research agents at scale, the AI Inference Cost in 2026 post breaks down the per-query economics across vendors.
For a deeper practitioner walkthrough of API-level research agent pricing and how to budget for production research tooling, the Mr. Technology network’s Deep Research skill guide covers the prompt patterns and verification recipes.