There is no single most accurate AI. The answer depends on what the test asks the model to do, and the three most useful public tests give three different answers.
- Summarizing a document you hand it: the lowest hallucination rates on Vectara's leaderboard (updated October 16, 2026) belong to small and older models, at 1.8% to 3.3%. GPT-6 Astra scores 8.7%, Claude Opus 4.7 12.0% and Gemini 4 Argon 12.1%.
- Answering from memory: on Artificial Analysis's AA-Omniscience test, Claude Opus 5.5 has the highest overall score and Claude Fable 5.1 the highest accuracy. Gemini 4 Argon is far less likely to give a wrong answer when it doesn't know.
- Answering with web search on: in a ten-week test of five chatbots by the fact-checking charity Full Fact, several gave stale or wrong figures. Full Fact did not publish a ranking.
If you publish what an AI writes, the ranking matters less than one rule from Google: "It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing."
How is AI accuracy measured?
Three kinds of test dominate, and they measure different things.
- Faithfulness to a source. You give the model a text and ask for a summary. A checker model then looks for claims the text doesn't support. Vectara's leaderboard works this way. Vectara says it measures "factual consistency" with the supplied passage, not overall truth.
- Recall from memory. You ask a closed-book question and grade the answer. AA-Omniscience does this with 6,000 questions across 42 topics and six domains: business, health, law, software engineering, humanities and social sciences, and science, engineering and math.
- Real answers from real chatbots. Fact-checkers ask everyday questions, with search switched on, and check the replies against official sources. Full Fact and the European Broadcasting Union (EBU) run tests like this.
The rankings disagree partly because of how the scores are built. A 2025 research paper on why language models hallucinate argues that "the training and evaluation procedures reward guessing over acknowledging uncertainty." A test that counts a wrong answer the same as "I don't know" pushes models to guess. A test that punishes wrong answers rewards the model that stays quiet. Both are legitimate, and they crown different models.
Which AI is most accurate when summarizing a source?
On Vectara's leaderboard, the winners are not the famous flagships. The dataset holds over 7,700 articles, from 50 to 24,000 words, and the leaderboard ranks 109 models by how often a summary adds something the article doesn't say. Lower is better.
Model | Hallucination rate | What it is |
|---|---|---|
antgroup/finix_s1_32b | 1.8% | Best on the board |
GPT-5.4 nano | 3.1% | Small OpenAI model |
Gemini 2.5 Flash-Lite | 3.3% | Small Google model |
GPT-6 Astra | 8.7% | OpenAI flagship |
Claude Haiku 4.5 | 9.8% | Smallest Claude tested |
Gemini 3.1 Pro (preview) | 10.4% | |
Claude Sonnet 4.6 | 10.6% | Anthropic |
Claude Opus 4.7 | 12.0% | Newest Claude tested |
Gemini 4 Argon | 12.1% | Google flagship |
Grok 4.1 Fast (reasoning) | 19.2% | xAI |
Vectara Hallucination Leaderboard, updated October 16, 2026. It measures how often a summary contradicts or adds to the source document, not how true the source is. The worst of the 109 models scores 24.2%. Claude Opus 5.5, Sonnet 5.5 and Fable 5.1 were not on the board on October 20.
Two things stand out. Newer and bigger does not mean fewer errors here, and the flagships invent something in roughly one summary in ten when the source is sitting in front of them. Vectara notes that this also tells you about retrieval setups, since a model "deployed as a summarizer of the search results" faces the same task.
Which AI is most accurate when answering from memory?
AA-Omniscience, run by Artificial Analysis, asks closed-book questions and reports three numbers. The index rewards correct answers, penalizes wrong ones and does not penalize refusals. Accuracy is the share answered correctly. Hallucination rate is, of the questions a model did not answer correctly, the share where it gave a wrong answer instead of declining or answering partly. Lower is better for that last column.
Model (best-scoring setting) | Index | Accuracy | Hallucination rate |
|---|---|---|---|
Claude Opus 5.5 (Max) | 46 | 66% | 59% |
GPT-6 Astra (High) | 44 | 61% | 45% |
Claude Fable 5.1 (Max) | 43 | 67% | 73% |
Gemini 4 Argon (High) | 42 | 50% | 15% |
GPT-6.1 Sol (Max) | 42 | 62% | 54% |
Claude Sonnet 5.5 (Max) | 32 | 54% | 47% |
Grok 4.7 (Xhigh) | 32 | 47% | 29% |
Gemini 3.8 Flash (High) | 30 | 55% | 55% |
Artificial Analysis, read October 20, 2026, from the 32 models its page shows by default. Settings are reasoning-effort levels. GPT-6 Astra at Max effort has a 51% hallucination rate.
Claude Fable 5.1 answers the most questions correctly and also guesses the most when it is wrong. Gemini 4 Argon sits at the other end: it gets half right and rarely bluffs the rest. Which one is "more accurate" depends on whether you can check the answer afterwards. If you can, high accuracy is worth the bluffing. If you can't, a model that says "I don't know" is safer.
The domain split is the second lesson. For Claude Opus 5.5, the same index broken down by domain runs from 88.8 in software engineering and 48.1 in law down to 32.7 in business and 26.4 in health. The model on top of the overall index is weakest in health and business, where a wrong sentence costs the most. The benchmark's authors draw the same conclusion: models should be chosen according to the demands of the use case.
What about the AI chatbots people actually use?
A model score and a chatbot score are different things. Chatbots add web search, system prompts and a retrieval step that can go wrong.
Full Fact asked ChatGPT, Claude, Gemini, DeepSeek and Grok the same 22 UK questions every day from July 15 to September 22, 2026 (7,625 responses) with web search enabled. Several models quoted £8.86 for asylum support after it had been replaced by £9.95 in June 2025. Some answers changed from one day to the next. Full Fact's conclusion: "a model's confidence is no guide to its reliability." It also warns that its findings "are not a measure of how often those responses occur," so read it as a list of failure types, not a leaderboard.
The EBU and BBC study from October 2025 found something similar with journalists as judges. Across more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity, 45% had at least one significant issue, 31% had sourcing problems and 20% had major accuracy problems such as invented or outdated details. That study is a year old and the models have changed, but the failure types match. We collected documented cases, and the errors we caught in our own AI writer, in AI hallucination examples.
One availability note. Gemini 4 Argon was announced on September 30 with limited access. On October 13, Google made it available in the Gemini app for AI Pro and AI Ultra subscribers.
Which AI should you use for accuracy?
The data supports a short answer by task, not a single winner.
- Rewriting or summarizing text you supply: check Vectara's board for your model. A small, cheap model can beat a flagship on this job.
- Facts from memory: Claude Opus 5.5 leads Artificial Analysis's index. Gemini 4 Argon leads on not bluffing. Both are wrong often enough that you need to check.
- Anything that changes, such as prices, rules, laws or product specs: use a chatbot with search on, then open the sources it cites. Full Fact's stale-figure examples came from chatbots that had search enabled.
- Health and business topics: treat any model's answer as a first draft. The domain scores above show why.
Why model choice can't fix hallucinations
Picking the best model lowers the error rate. It doesn't bring it to zero. The best model on the summarization test still adds unsupported content now and then with the source in hand, and the best model on the memory test is wrong on most of what it doesn't know. Two points follow.
First, facts go stale. A model's training data stops at a cutoff. Anthropic lists June 2026 as the knowledge cutoff for its current Claude models, and OpenAI lists April 30, 2026 for GPT-6 Astra and GPT-6.1 Sol. Anything that changed after that, or that a source got wrong, can come out as confident prose.
Second, errors are not obvious. They sit in a price, a date, a name or a figure inside a fluent paragraph. Nothing in the tone tells you which sentence is wrong.
What accuracy means when you write with AI
For affiliate and review content, a wrong claim does direct damage. A reader who buys a product because your review gave it a feature, a rating or a price it doesn't have blames your page. Specs, prices and "what's in the box" are exactly the facts that models recall poorly and that change between product generations. Roundups add a second risk: facts from one product drifting onto another in a long list.
For small businesses, there is usually no editor between the draft and the publish button. The topics a small firm writes about, such as pricing, tax, health and local rules, include the domains where the top model scored lowest in the domain breakdown above, and they are facts that change often.
If you want to know which AI writing tools have verification built in, we compared seven in a separate post on AI writing tools with built-in fact-checking. This article stays with the models underneath them.
What a fact-checking workflow catches that model choice can't
A better model changes the odds. A fact check changes the outcome, because it reads the finished article and tests each claim against a source.
We have two measurements of this, both internal. In a test with 29 planted errors (wrong figures, prices and names) in seven articles, ProofWrite's fact check flagged all 29 and cleared none. Separately, we checked 28 articles ranking on Google's first page on October 9, 2026, and 19 of them had verified errors, 213 wrong sentences in total. Those 28 articles were not known to be AI-written, so they say nothing about AI hallucination rates. They show how often published facts are wrong or out of date, whoever wrote them.
The check isn't perfect. In that same post, of the Likely wrong findings we could decide, 74% (88 of 119) were real errors. The rest are for you to review, which is the point: you see the sentence, the source passage and a verdict, and you decide.
ProofWrite is not a model. It is a research and fact-checking workflow around models. It gathers sources before drafting, checks every claim in the draft against them, shows what it could not confirm and leaves the decision with you. You can also import an article you wrote yourself, or check pages that are already live.
What Google says about AI content and accuracy
Google's guidance on generative AI content was updated on October 1, 2026. It doesn't ban AI content. It says the review step is on the publisher: "It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing." Google puts the weight on whether someone checked the draft, whichever model wrote it.
FAQ
What is the most accurate AI chatbot right now?
It depends on the test. Claude Opus 5.5 has the highest overall score on AA-Omniscience (46), Claude Fable 5.1 the highest accuracy (67%), and Gemini 4 Argon the lowest hallucination rate among the top scorers (15%). On Vectara's summarization test, small models lead. None of these tests covers the chatbot apps as you use them, with search on.
Which AI hallucinates the least?
On Vectara's summarization test, antgroup/finix_s1_32b at 1.8%. Among the large flagships on that board, GPT-6 Astra at 8.7%. On AA-Omniscience's closed-book questions, Gemini 4 Argon at 15%, among models that also score near the top overall. These are different tasks, so the lists don't merge.
Is ChatGPT, Claude or Gemini more accurate?
Each leads on some measure. Claude models lead AA-Omniscience on accuracy, Gemini 4 Argon leads on hallucination rate because it excels at not guessing, and GPT-6 Astra rounds out the top three by overall score.
Does web search stop AI from hallucinating?
It helps, and it doesn't remove the problem. Full Fact ran its test with web search enabled and still found outdated policy figures and answers that changed from day to day. Search gives the model a source. The model can still misread it, pick an old page or round a figure.
What is the best AI for fact-based writing?
Pick a strong model, then add a step that checks claims against sources, because no model is accurate enough to skip it. Look for a workflow that shows you each claim with its evidence and tells you which ones it couldn't confirm.
Check your articles before your readers do
If you already publish AI-assisted articles, start with three. Either they are clean, or you know which sentences to fix.
Methodology: all figures were read on October 20, 2026. Vectara figures come from its Hallucination Leaderboard README (updated October 16, 2026, HHEM-2.3 evaluator, over 7,700 documents, 109 models, summaries generated at temperature 0 where supported). AA-Omniscience figures come from the data behind Artificial Analysis's evaluation page, which shows 32 of 580 models by default; we report each model's highest-index setting and did not rerun any test. Artificial Analysis's hallucination rate is incorrect answers divided by all non-correct responses. The Full Fact and EBU studies measure chatbots, not models, and their authors state the limits of what the results mean. We left out FACTS and SimpleQA Verified scores because we could not read current figures from the benchmark owners' pages. ProofWrite's own figures are internal measurements, described in full in the linked post. Rankings change monthly; check the sources below for today's numbers.
Sources
- Vectara Hallucination Leaderboard and its redesign announcement
- Artificial Analysis, AA-Omniscience and Benchmarking GPT-6 Astra
- Full Fact, "Wrong on Sunday, right on Monday"
- EBU and BBC study via VRT
- Why Language Models Hallucinate (arXiv)
- Google Search Central, generative AI content
- Google, Gemini 4 Argon in the Gemini app
- Anthropic, models overview and OpenAI, models

Written by
Jussi Hyvarinen - Co-founder of ProofWrite
I built this platform to solve my own frustration with slow research and generic AI. I use it to write every article you see on this blog, including this one.
Publish articles where every claim is checked
ProofWrite researches and writes articles and checks every factual claim against its sources.
Try it on an article you already wrote.
Free Pilot: 1 article and 3 of your own articles. No credit card.
