Verdict: Gemini Deep Research is the DeepSearch AI most people should use in September 2026. Deep Research Max leads the published benchmark at 54.6% on Humanity's Last Exam against 50.4% for the standard tier (Google, April 2026, via this roundup), and AI Plus is the cheapest paid entry of any frontier lab at $4.99 per month (plan details). Pick Perplexity Pro for volume, ChatGPT for one long ambiguous report, Grok DeepSearch for live X coverage.
TL;DR
- Best overall: Gemini Deep Research, benchmark leader at a $4.99/mo entry, and the only turnkey frontier-lab deep-research API not on a retirement schedule.
- Best for volume: Perplexity Pro at $20/mo, listed at 20 runs per day against Gemini's much tighter allowance (PCMag).
- Best for one deep report: ChatGPT, though OpenAI retired its dedicated deep-research models on 23 July 2026 (migration guide).
- Best for real-time X: Grok DeepSearch on SuperGrok at $30/mo, with 15 to 60 second turnaround (xAI usage guide).
- Two deadlines reshaped the map: Perplexity's Sonar endpoint retires 27 September 2026 (Perplexity); OpenAI's deep-research models are already gone.
- Last verified: 2026-09-22.
What is DeepSearch AI, and why do the names differ?
DeepSearch AI is the general term for an agentic research mode: you ask a question, it plans searches, reads dozens of sources, and returns a cited report instead of a chat reply. Google, OpenAI, Perplexity, and xAI all ship one under different names. The mechanics are similar; the pricing, run limits, and API stability are not.
Which DeepSearch AI wins on price?
Gemini, clearly. AI Plus starts at $4.99 per month, with AI Pro at $19.99 and AI Ultra from $99.99 (plan table) - the cheapest route to the benchmark leader in this category.
OpenAI paused new signups for its $200 Pro plan on 10 September 2026, leaving $100 as the top purchasable tier (source), and its pricing page now says Limited, Expanded, or Maximum instead of run counts. The widely circulated "5 free, 25 Plus, 250 Pro" figures date from an April 2025 blog post and no longer describe what you buy.
| Tool | Entry paid price | Stated run allowance | Notable constraint |
|---|---|---|---|
| Gemini Deep Research | $4.99/mo (AI Plus) | Multiplier-based; limits refresh every 5 hours (plan table) | Tight monthly allowance on lower tiers |
| Perplexity Deep Research | $20/mo (Pro) | 20 runs/day on Pro, 5/day free (PCMag) | Sonar endpoint retires 27 Sep 2026 (Perplexity) |
| ChatGPT Deep Research | Plus tier upward | No published run counts since 2026 | Dedicated models shut down 23 Jul 2026 (details) |
| Grok DeepSearch | $30/mo (SuperGrok) | Not published as run counts | Optimised for speed, not depth (guide) |
Which DeepSearch AI is best if you run research daily?
Perplexity. Pro at $20 per month is listed at 20 Deep Research runs per day, with the free tier at five (PCMag). If your workflow is twenty short briefs a day rather than two long ones a month, that allowance beats any benchmark gap.
The engine underneath changed twice: since February 2026 advanced deep research runs on Claude Opus (Max gets a newer revision than Pro), and since June 2026 it executes inside Perplexity Computer, routing subtasks across more than 20 frontier models. Perplexity attributes a BrowseComp jump from 40.7% to 83.8% to that change (reported here) - company-reported, not an independent test.
What changed in September 2026 that makes older comparisons wrong?
Two retirement dates. Perplexity retires the Sonar chat-completions endpoint, including sonar-deep-research, on 27 September 2026; the Agent API becomes the single surface (Perplexity hub blog; migration guide). Perplexity reports its Agent API low preset beats Sonar Pro by about seven times on BrowseComp at three cents per query, and its fast preset roughly doubles Sonar's DSQA score at six-tenths of a cent per query - company-reported numbers published with the retirement notice (announcement).
OpenAI went further: o3-deep-research and o4-mini-deep-research were deprecated on 22 April 2026 and shut down on 23 July 2026, replaced by gpt-5.5-pro on the Responses API with the web_search tool - you orchestrate the research loop yourself (details). OpenAI's own deep-research guide page and the o3-deep-research model page are stale and still present the models as available, so trust the deprecations page (noted here).
That leaves Gemini's Deep Research Agent API, on the Interactions API with background=true, as the only turnkey frontier-lab deep-research API without a published end date - roughly $1 to $3 per task, or $3 to $7 for Max, with a documented 60-minute cap (reference).
How long do these tools actually take?
Long enough that you should not wait at the keyboard. In PCMag's four-way test the full version of one tool took 49 minutes while a lightweight variant finished in about five, and the reviewer favoured Gemini or ChatGPT (PCMag). Grok DeepSearch sits at the fast end at 15 to 60 seconds (guide), trading depth for freshness with live X posts alongside the web (review).
Our own testing agrees on the speed pattern. Across three trials each on an identical seven-constraint planning task (n=6 total, measured 2026-09-22), Gemini 3.8 Flash (High) and Claude Opus 4.6 (Thinking) both scored 17 of 17 on machine-checked constraint adherence; median wall time was 23 seconds for Gemini against 67 seconds for Opus. A narrower task than a full research run, but the pattern holds: comparable quality, less waiting.
Google added Deep Research to Gemini Live on 19 August 2026: start a report by voice, let it run as a background task, get a notification when it finishes, and keep the report in context across voice and text (9to5Google; Digital Trends).
Where each tool falls short
Gemini's weak point is volume: allowances are multipliers that refresh every five hours, and lower tiers run out quickly (plan table). Perplexity's is churn: two engine changes and an endpoint retirement inside eight months. ChatGPT's is that the managed research models are gone from the API side and its documentation has not caught up. Grok's is depth - academic research belongs elsewhere (comparison).
One caution on benchmarks: Kimi K3 claims 56.0% on Humanity's Last Exam, above Gemini Max's 54.6% (Google's figures via this roundup) - a vendor claim, not an independently reproduced result. Our Tencent HY4 Preview vs Kimi K3 vs GLM 5.3 comparison covers where those claims hold up.
Related reading
- Self-hosting your search stack: Perplexica vs Perplexity
- Grok as an agent, not a search mode: Grok bot vs Hermes agent
- Model choice for code: Gemini 3.8 Flash vs Opus 4.6
- Cutting API spend: cheapest AI API verdict
FAQ
Q: Which DeepSearch AI is cheapest in 2026?
A: Gemini, with a $4.99 per month AI Plus entry tier, the lowest paid entry of any frontier lab's research agent (plan table).
Q: Is Gemini Deep Research faster than the alternatives?
A: On our seven-constraint task, yes: a 23-second median against 67 seconds for Claude Opus 4.6 at identical constraint scores. For full runs, Grok DeepSearch is faster still at 15 to 60 seconds but far shallower (guide).
Q: Why do published benchmark scores for Gemini Deep Research disagree?
A: Two tiers: Max scored 54.6% on Humanity's Last Exam, the standard tier 50.4%, and many articles quote the Max figure under the plain product name (source).
Q: Can I still build on Perplexity's deep-research API?
A: Not through Sonar after 27 September 2026, when the endpoint including sonar-deep-research retires; the Agent API replaces it (Perplexity).
Q: What happened to OpenAI's deep-research models?
A: They were deprecated on 22 April 2026 and shut down on 23 July 2026, replaced by gpt-5.5-pro with the web_search tool on the Responses API - you build the research loop yourself (details).
Q: Is Grok DeepSearch a fair comparison to the others?
A: Only partly: it is the one option that searches live X posts alongside the web, so it wins on social freshness, but it is not built for depth (review).
Last verified: 2026-09-22. Pricing, run allowances, and retirement dates in this category changed several times during 2026; check vendor pricing pages before committing to a plan.
Corrections log: No corrections yet. Spotted an error? See how we work for our sourcing, verification, and AI-disclosure policy.
Top comments (0)