The most accurate AI assistants
Short answer
Claude is the most accurate AI assistant on the published evidence. It repeated false news claims 10% of the time in NewsGuard's August 2025 audit, the lowest of 10 chatbots. Gemini ranks second and Perplexity third. Grok ranks last: Grok 3 named the wrong source for 94% of news excerpts in the Tow Center test.
Results as of October 2026
The top three
The leading assistants on the current evidence.
- Top pick
Claude, ranked 1
Mostly reliable10% false news claims repeated, lowest of 10 (NewsGuard)
Evidence
The ranking
| Rank | Tool | Key result (test) | Evidence date | Verdict |
|---|---|---|---|---|
| 1 | Claude | 10% false news claims repeated, lowest of 10 (NewsGuard) | Mostly reliable | |
| 2 |  Gemini | 16.67% false news claims repeated, second of 10 (NewsGuard) | Mixed | |
| 3 |  Perplexity | 37% wrong source lookups, best of eight (Tow Center) | Mixed | |
| 4 |  Microsoft Copilot | 36.67% false news claims repeated (NewsGuard) | Mixed | |
| 5 |  ChatGPT | 36% of news answers with a significant issue (EBU/BBC) | Mixed | |
| 6 |  Google AI Overviews | 91% correct on SimpleQA questions, 56% ungrounded (NYT and Oumi, via Futurism) | Mixed | |
| 7 |  Meta AI | 40% false news claims repeated (NewsGuard) | Mixed | |
| 8 |  DeepSeek | 83% fail rate on news prompts (NewsGuard) | Unreliable | |
| 9 |  Grok | 94% wrong source lookups, worst of eight (Tow Center) | Unreliable |
The key result is each tool's headline figure from one named test. Figures in different rows often come from different tests and are not compared with each other. The comparisons that set the order are in the results by test section below.
No AI assistant in these tests was right every time; the most accurate one still needs checking.
The assistants, one by one
Rank 1: Claude: the most accurate AI assistant
Mostly reliableby Anthropic- What it does best
- Holding back on hard questions. Claude Opus 5.5, a model behind Claude, scores 48 on the AA-Omniscience index, second highest when checked on 2026-10-29. The index penalizes wrong answers more than no answer.
- Where it slips
- Summaries. Claude Opus 5.5 added unsupported content to 9.9% of summaries on the Vectara leaderboard, mid-pack.
Choose it ifyou want the assistant least likely to repeat a false claim.
Is Claude accurate?Key resultClaude repeated false news claims 10% of the time in NewsGuard's August 2025 audit, the lowest of 10 chatbots (NewsGuard, 2025-09-04).Evidence dateRank 2: Gemini: strong on false claims, weak on sourcing
Mixedby Google- What it does best
- Hard knowledge questions. Gemini 4 Argon (High), a model behind Gemini, had the lowest hallucination rate on AA-Omniscience at 4%, and Gemini 4 Carbon (High) the top index score at 54 (Artificial Analysis).
- Where it slips
- News sourcing. 76% of Gemini's news answers had a significant issue in the EBU/BBC study, the worst of four, mostly from sourcing problems.
Choose it ifyou need hard factual answers and will check the sources yourself.
Is Gemini accurate?Key resultGemini repeated false news claims 16.67% of the time, second lowest of 10 in NewsGuard's audit.Evidence dateRank 3: Perplexity: the best citer
Mixedby Perplexity- What it does best
- Sourcing. Perplexity also had the fewest news answers with a significant issue in the EBU/BBC study, at 30%.
- Where it slips
- False stories on the web. Perplexity repeated false news claims 46.67% of the time in NewsGuard's August 2025 audit, second worst of 10, after scoring 0% in August 2024.
Choose it ifyou want links you can click and check for every answer.
Is Perplexity accurate?Key resultPerplexity named the wrong source for 37% of news excerpts, the lowest error rate of eight AI search tools in the Tow Center test.Evidence dateRank 4: Microsoft Copilot: middle of the pack, careful with quotes
Mixedby Microsoft- What it does best
- Quotes and restraint. Only 4% of Copilot responses with direct quotes had quote problems in the EBU/BBC study, and it was the only tool in the Tow Center test that declined more questions than it answered.
- Where it slips
- News answers. 37% had a significant issue in the EBU/BBC study, and NewsGuard found it citing Pravda network content via VK.
Choose it ifyou would rather get no answer than a confident wrong one.
Is Microsoft Copilot accurate?Key resultCopilot repeated false news claims 36.67% of the time in NewsGuard's audit.Evidence dateRank 5: ChatGPT: useful, but confidently wrong too often
Mixedby OpenAI- What it does best
- Everyday news questions, where it beat Copilot by a point and Gemini by a wide margin in the EBU/BBC study.
- Where it slips
- Sources and false claims. ChatGPT Search got 134 of 200 source lookups wrong in the Tow Center test and never declined. It repeated false news claims 40% of the time in NewsGuard's audit.
Choose it ifyou want a fast first draft and will verify the facts.
Is ChatGPT accurate?Key result36% of ChatGPT's news answers had a significant issue in the EBU/BBC study, second best of four.Evidence dateRank 6: Google AI Overviews: right often, grounded less often
Mixedby Google- What it does best
- Short factual questions, where accuracy rose from 85% to 91% after the move to Gemini 3.
- Where it slips
- Grounding. 56% of answers were not supported by the sources they cited. Google disputes the study.
Choose it ifyou only need a quick fact and will click through to confirm it.
Are Google AI Overviews accurate?Key resultAI Overviews answered 91% of SimpleQA questions correctly in the New York Times and Oumi analysis, reported by Futurism on 2026-04-08.Evidence dateRank 7: Meta AI: tied with ChatGPT on false claims
Mixedby Meta- What it does best
- Summaries, at the model level. Llama 4 Maverick, a model behind Meta AI, added unsupported content to 8.2% of summaries on the Vectara leaderboard.
- Where it slips
- Breaking news. Meta admitted that some answers in July 2024 said the Trump assassination attempt had not happened.
Choose it ifyou already use it inside WhatsApp or Instagram and will check before you share.
Is Meta AI accurate?Key resultMeta AI repeated false news claims 40% of the time in NewsGuard's August 2025 audit, tied with ChatGPT.Evidence dateRank 8: DeepSeek: poor on news, good at summaries
Unreliableby DeepSeek- What it does best
- Summaries. DeepSeek-V4.1-Flash, a model behind DeepSeek, added unsupported content to only 4.2% of summaries on Vectara.
- Where it slips
- Sources. DeepSeek Search misattributed 115 of 200 news sources in the Tow Center test.
Choose it ifyou need a summary of a document you supply, not facts about the world.
Is DeepSeek accurate?Key resultDeepSeek failed 83% of news prompts in NewsGuard's January 2025 audit: 30% false claims and 53% non-answers.Evidence dateRank 9: Grok: the least accurate in published tests
Unreliableby xAI- What it does best
- The older Grok 3 model scored 5.8% on Vectara's summary test, better than the newer Grok 4 models.
- Where it slips
- 154 of 200 Grok 3 citations led to error pages. The Grok 4 Fast reasoning model, a model behind Grok, added unsupported content to 20.2% of summaries on Vectara. xAI's own 4.22% hallucination figure for Grok 4.1 is a vendor claim.
Choose it ifyou only want a starting point and will verify every fact elsewhere.
Is Grok accurate?Key resultGrok 3 named the wrong source for 94% of news excerpts in the Tow Center test, the worst of eight.Evidence date
Results by test
These tables set the order. Each one compares tools within a single test, never across tests.
NewsGuard AI False Claims Monitor, August 2025 (published 2025-09-04): rate of repeating false news claims
| Chatbot | Result |
|---|---|
| Claude | 10% |
| Gemini | 16.67% |
| Microsoft Copilot | 36.67% |
| ChatGPT | 40% |
| Meta AI | 40% |
| Perplexity | 46.67% |
| Average of 10 chatbots | 35% |
Source: NewsGuard. The worst chatbot in the audit was Inflection Pi at 56.67%.
EBU/BBC News Integrity in AI Assistants (2025-10-21): news answers with a significant issue
| Assistant | Result |
|---|---|
| Perplexity | 30% |
| ChatGPT | 36% |
| Microsoft Copilot | 37% |
| Google Gemini | 76% |
Source: EBU and BBC. 2,709 core responses from 22 public service media in 18 countries.
Tow Center AI search citation test (2025-03-06): source lookups that went wrong, 200 queries per tool
| Tool | Result |
|---|---|
| Perplexity (free), wrong answers | 37% |
| DeepSeek Search, misattributed | 115 of 200 |
| ChatGPT Search, wrong answers | 134 of 200 (67%) |
| Grok 3, wrong answers | 94% |
Source: CJR, Tow Center. Copilot declined more questions than it answered.
Vectara Hallucination Leaderboard (2026-10-22): summaries with unsupported content, models behind each assistant
| Model | Result |
|---|---|
| DeepSeek-V4.1-Flash (DeepSeek) | 4.2% |
| Gemini 2.5 Pro (Gemini) | 7.0% |
| Llama 4 Maverick (Meta AI) | 8.2% |
| GPT-5.5 (ChatGPT) | 9.3% |
| Claude Opus 5.5 (Claude) | 9.9% |
| Grok 4 Fast reasoning (Grok) | 20.2% |
Source: Vectara. This test measures summary faithfulness only. The full table is on AI hallucination rates by model.
How we ranked
The order comes from published results only: independent studies, news audits and public benchmarks, each with its date and source. Vendor figures appear on the tool pages, labelled as the vendor's claim, and do not move a tool up or down.
- Compare within one test at a time. A tool's place in each test is what counts, not its raw number next to a number from another test.
- Weight tests of the consumer product most. NewsGuard, the EBU/BBC study, the Tow Center test and the NYT and Oumi analysis tested the apps people use. Vectara and AA-Omniscience test the models behind them and break ties.
- Weight repeating false claims heavily. A confident false statement does more harm than a weak citation, so NewsGuard's audit carries the most weight among the news tests.
- Break close calls on breadth of evidence. Claude ranks first because it leads the false-claims audit, the test weighted most, and the model behind it sits second on AA-Omniscience. Meta AI tied ChatGPT in NewsGuard's audit; ChatGPT ranks higher because it also placed second of four in the EBU/BBC study, while Meta's own statement records a breaking-news failure. Google AI Overviews was tested on its own, so its place rests on its verdict: right on 91% of questions, but ungrounded on 56%.
Known weaknesses of this method: the news tests are from 2025, and tools change faster than studies are published. NewsGuard's prompt sets per chatbot are small. The EBU/BBC samples per organization are small too, so we use the pooled figures. The Tow Center tested March 2025 versions, including Grok 3 beta. The full method is on how we score AI accuracy, and the rules for sourcing are on editorial standards.
Sources
Every result in this ranking comes from one of these.
- News audit
- News audit
- Independent study
- Independent study
- Benchmark
- Benchmark
- News audit
- Vendor claim
Questions
Short answers, with links to the dated evidence.
Which AI is the most accurate?
Claude is the most accurate AI assistant on the published evidence. It repeated false news claims 10% of the time in NewsGuard's August 2025 audit, the lowest of 10 chatbots, and Claude Opus 5.5 scores 48 on the AA-Omniscience index, second highest when checked on 2026-10-29. Gemini ranks second and Perplexity third.
Which AI chatbot makes the fewest mistakes?
Claude made the fewest false-claim mistakes in NewsGuard's August 2025 audit, at 10%. For citing news sources, Perplexity made the fewest mistakes, wrong on 37% of lookups in the Tow Center test. Which chatbot is best depends on the kind of mistake you most need to avoid, so check the test that matches your task.
Is the most accurate AI also the best?
Not always. Accuracy is one measure; speed, features, price and how a tool fits your work matter too. This ranking covers only how often each assistant gets facts right. A tool lower on the list may suit a task better, as DeepSeek's newer models do for summarizing a document you supply.
Does paying make AI more accurate?
Not reliably. The Tow Center found premium AI search tiers gave more confidently wrong answers than free ones in March 2025, and Grok's paid tier was not more accurate than the free one. Paid plans may unlock larger models, but model results vary, so check paid answers exactly as you would check free ones.
How often do these rankings change?
The ranking changes when new published results change a tool's place within a test. Live leaderboards such as Vectara and AA-Omniscience move often, while large news studies appear less often. Every figure on this page carries its test date, so you can see how recent the evidence for each tool is.
