The same model can be measured at 0.7% and at 7.6% hallucination depending on which benchmark you run, and across 26 frontier models Stanford HAI recorded a range from 22% to 94%. That spread is the single most important fact about AI hallucination statistics in 2026: benchmark design determines the headline number far more than model quality does. Grounded summarisation, where the model is handed the source text, produces the flattering figures that appear in vendor materials. Open-recall testing, closer to how people actually use these systems, produces numbers an order of magnitude worse. The figures below come from Vectara’s hallucination leaderboard and the Stanford HAI 2026 AI Index.
TL;DR
- Reported 2026 hallucination rates span roughly 0.7% to 94% depending on benchmark (Vectara, Stanford HAI)
- Gemini-2.0-Flash-001 recorded 0.7% on grounded summarisation (Vectara)
- The same model scores 7.6% under Vectara’s FaithJudge benchmark (Vectara)
- That is roughly an elevenfold difference for one model (derived)
- Stanford HAI recorded 22% to 94% across 26 frontier models (Stanford HAI, 2026)
- Those figures measure sycophancy-induced hallucination (Stanford HAI, 2026)
- Vectara’s November 2025 refresh used more than 7,700 articles (Vectara)
- The refresh was designed to be significantly more challenging (Vectara)
- Measured rates rose as a direct result of the harder test (Vectara)
- Best rows on the grounded summarisation leaderboard sit near 1.8% (Vectara)
- Stanford’s findings appear in its Responsible AI chapter (Stanford HAI, 2026)
- Grounded summarisation supplies the source text to the model (Vectara)
- Open-recall benchmarks require the model to produce facts unaided (Stanford HAI)
1. The Range Is the Statistic
Anyone quoting a single hallucination rate is describing one benchmark, not the phenomenon. Published 2026 rates run from roughly 0.7% at the best end of Vectara’s grounded summarisation leaderboard to between 22% and 94% across the 26 frontier models Stanford HAI tested. Those are not competing estimates of the same quantity; they are measurements of different tasks, and the gap between them is more than two orders of magnitude.
| Metric | Value | Source |
|---|---|---|
| Lowest published rate, grounded summarisation | 0.7% | Vectara |
| Best-performing rows on that leaderboard | near 1.8% | Vectara |
| Range across 26 frontier models | 22% to 94% | Stanford HAI, 2026 |
| Model achieving the 0.7% figure | Gemini-2.0-Flash-001 | Vectara |
| Same model under FaithJudge | 7.6% | Vectara |
| Ratio between those two measurements | roughly 11x | Derived from Vectara figures |
| Span between lowest and highest published rates | more than two orders of magnitude | Derived |
| What determines the number | benchmark design | Vectara, Stanford HAI |
The elevenfold gap for a single model on two benchmarks from the same organisation is the cleanest demonstration available that these numbers describe tests rather than models.
This has a practical consequence that gets lost in benchmark discussion. If you are choosing a model for a retrieval-grounded application, where the system always supplies source documents, the grounded summarisation figures are the relevant ones and a sub-2% rate is a reasonable expectation. If you are choosing a model to answer questions from its own knowledge, those same figures are actively misleading and the 22% to 94% band is closer to what you should plan for. Most production disappointment with these systems traces to a team that benchmarked one way and deployed the other. The benchmark is not wrong; it was answering a different question than the one the deployment poses. Source: Vectara hallucination leaderboard.
2. Grounded Summarisation: The Optimistic Test
The benchmark that produces sub-1% figures is doing something specific and narrow. Grounded summarisation hands the model a source document and checks whether the resulting summary stays faithful to it, which removes the need for the model to know anything. It is a genuine and useful measurement of one failure mode, and it is the one vendors cite.
| Metric | Value | Source |
|---|---|---|
| Task measured | faithfulness of a summary to supplied source text | Vectara |
| Lowest recorded rate | 0.7% | Vectara |
| Typical best-performer band | near 1.8% | Vectara |
| Articles in the November 2025 refresh | more than 7,700 | Vectara |
| Design intent of the refresh | larger, more robust, more challenging | Vectara |
| Effect of the refresh on measured rates | higher | Vectara |
| What the task does not test | unaided factual recall | Derived |
| Why vendors cite it | it produces the lowest numbers | Derived |
The refresh detail matters more than it first appears: measured hallucination went up because the test got harder, not because models got worse, which means year-over-year comparisons on this leaderboard are unreliable across the version change. Source: Vectara on the next generation of its hallucination leaderboard.
3. Open Recall: The Pessimistic Test
The benchmarks producing double-digit and triple-digit-adjacent figures are testing something much closer to real use. Stanford HAI recorded hallucination rates from 22% to 94% across 26 frontier models on an accuracy benchmark reported in its 2026 Responsible AI chapter. When a model must supply facts from its own parameters rather than from a document in front of it, the failure rate rises dramatically.
| Metric | Value | Source |
|---|---|---|
| Lowest rate across the 26 models | 22% | Stanford HAI, 2026 |
| Highest rate across the 26 models | 94% | Stanford HAI, 2026 |
| Number of frontier models tested | 26 | Stanford HAI, 2026 |
| Spread between best and worst model | 72 points | Derived from Stanford figures |
| Chapter reporting the finding | Responsible AI | Stanford HAI, 2026 |
| Failure mode measured | sycophancy-induced hallucination | Stanford HAI, 2026 |
| Publication date of the AI Index | April 2026 | Stanford HAI |
| Comparison to grounded summarisation rates | far higher | Derived |
A 72-point spread between the best and worst frontier model on the same test is itself notable: model choice matters enormously on this task, and barely at all on grounded summarisation where everything clusters in low single digits. Source: Stanford HAI 2026 AI Index Report.
4. Sycophancy Is a Distinct Failure Mode
Stanford’s framing is worth separating from ordinary factual error. Sycophancy-induced hallucination means the model conforms to a premise the user has stated, even when that premise is false, which is a different mechanism from a model simply not knowing something. It also means the measured rate depends partly on how the question is posed, not only on what the model knows.
| Metric | Value | Source |
|---|---|---|
| Failure mode | conforming to a false user premise | Stanford HAI, 2026 |
| Rate range across models | 22% to 94% | Stanford HAI, 2026 |
| Models evaluated | 26 frontier models | Stanford HAI, 2026 |
| Distinction from ordinary factual error | mechanism differs | Stanford HAI, 2026 |
| Dependence on prompt framing | high | Derived |
| Reporting chapter | Responsible AI | Stanford HAI, 2026 |
| Broader AI Index finding on reporting | responsible AI disclosure lags capability | Stanford HAI, 2026 |
| Implication for evaluation | prompt design is part of the measurement | Derived |
The practical consequence is uncomfortable for anyone deploying these systems: a model can be highly accurate when asked neutrally and highly unreliable when a user asserts something wrong first. Deployment context sits in our enterprise AI adoption statistics. Source: Stanford HAI, 12 takeaways from the 2026 AI Index.
5. Reading a Vendor Claim
The practical upshot is a short checklist. A sub-1% hallucination claim is almost certainly grounded summarisation, where the model was handed the source material, and it says nothing about unaided recall. The right question is never “what is the hallucination rate” but “on which benchmark, on which task, at what sample size”.
| Metric | Value | Source |
|---|---|---|
| What a sub-1% claim usually measures | grounded summarisation | Vectara |
| What it does not measure | unaided factual recall | Derived |
| Rate for the same model on a harder in-house test | 7.6% | Vectara |
| Rate band on open-recall style testing | 22% to 94% | Stanford HAI, 2026 |
| Articles in the current Vectara test set | more than 7,700 | Vectara |
| Models covered by the Stanford range | 26 | Stanford HAI, 2026 |
| Questions to ask of any claim | which benchmark, which task, what sample | Derived |
| Whether cross-source comparison is valid | no, without matching methodology | Derived |
Retrieval-grounded deployments genuinely do sit closer to the optimistic end, because they replicate the conditions of the optimistic benchmark, which is the one legitimate way to use the low figures. The reverse also holds: a chatbot answering open questions without retrieval should be evaluated against the pessimistic band, and quoting the grounded number for that product is a category error rather than an exaggeration. Search and retrieval context sits in our AI search statistics, and model-landscape context in our open-source AI statistics. Source: Benchmarking LLM faithfulness in RAG with evolving leaderboards.
Summary: AI Hallucination by the Numbers
| Metric | Value | Source |
|---|---|---|
| Lowest published rate | 0.7% | Vectara |
| Best-performer band, grounded summarisation | near 1.8% | Vectara |
| Same model under FaithJudge | 7.6% | Vectara |
| Ratio between those measurements | roughly 11x | Derived |
| Range across 26 frontier models | 22% to 94% | Stanford HAI |
| Spread between best and worst model | 72 points | Derived |
| Models tested by Stanford HAI | 26 | Stanford HAI |
| Failure mode Stanford measured | sycophancy-induced hallucination | Stanford HAI |
| Articles in Vectara’s refreshed test set | more than 7,700 | Vectara |
| Design intent of the refresh | more challenging | Vectara |
| Effect on measured rates | higher | Vectara |
| Task in grounded summarisation | faithfulness to supplied source | Vectara |
| Task in open-recall benchmarks | unaided factual production | Stanford HAI |
| Reporting chapter at Stanford HAI | Responsible AI | Stanford HAI |
| AI Index publication date | April 2026 | Stanford HAI |
| Span across all published 2026 rates | more than two orders of magnitude | Derived |
Methodology and Sources
- Grounded summarisation rates, the FaithJudge comparison, and the November 2025 test set refresh come from Vectara’s public hallucination leaderboard and its accompanying documentation (leaderboard repository, next-generation leaderboard announcement).
- The 22% to 94% range across 26 frontier models, the sycophancy framing, and the Responsible AI chapter context come from the Stanford HAI 2026 AI Index, published April 2026 (report, takeaways, AI Index home).
- Methodological background on why leaderboard design drives measured faithfulness comes from published research on evolving RAG benchmarks (arXiv).
- Data watch: the figures in this roundup must not be averaged, compared, or presented as a single rate. Vectara measures faithfulness to a supplied document; Stanford HAI measures a different failure mode under different conditions. Vectara’s November 2025 refresh deliberately increased test difficulty, so rates before and after that change are not comparable and apparent worsening may reflect the harder test rather than worse models. Model-specific figures move quickly as new versions ship, and any named model result more than a quarter old should be treated as stale. Stanford’s sycophancy measurement depends on prompt framing, which is part of the experimental design rather than a fixed property of the models. Vectara is a commercial vendor of retrieval-grounded AI products, which gives it an interest in benchmarks where grounding performs well. Rows marked as derived are arithmetic or direct inference from published figures.
- Last updated: August 2, 2026. We update this roundup quarterly as Vectara refreshes its leaderboard and Stanford HAI publishes new AI Index editions.