How often AI gets it right
How often today’s AI models get facts right, and how often they make things up
Independent test results for current AI models (October 2026): how many hard fact questions each one answers correctly, and how often it guesses wrong instead of saying “I don’t know.” Even the best models still miss many questions, so check anything important.
Checked Oct 16, 2026
What these numbers mean.
- Gets it right: out of 6,000 hard fact questions in law, health, business, science and software, the share the model answered correctly. Higher is better. (Artificial Analysis, AA-Omniscience Accuracy)
- Makes things up: when the model did not give the right answer, how often it gave a wrong answer instead of saying it didn’t know. Lower is better. (AA-Omniscience Hallucination Rate)
- Summary errors: when asked to summarize a short document, how often the summary added something that wasn’t in it. Lower is better. (Vectara Hallucination Leaderboard)
- These are different tests, so their numbers can’t be compared with each other, and a test score is not a promise about your own questions. A model that guesses less often can score lower on “gets it right” but make up fewer answers.
- We list only current models and only numbers published by independent testers, copied from their own pages. “Not published yet” means we found no result for that exact model. We never estimate.
Current models, side by side
Sorted by “gets it right.” “Top model” marks each company’s most capable current model.
| Model | Maker | Gets it right higher is better | Makes things up lower is better | Summary errors lower is better |
|---|---|---|---|---|
| Claude Fable 5.1 | Anthropic | 67% | 73% | Not published yet |
| Claude Opus 5.5 Top model | Anthropic | 66% | 59% | Not published yet |
| GPT-6.1 Sol | OpenAI | 62% | 54% | Not published yet |
| GPT-6 Astra Top model | OpenAI | 61% | 45% | 8.7% |
| Gemini 3.8 Flash | 55% | 55% | Not published yet | |
| Claude Sonnet 5.5 | Anthropic | 54% | 47% | Not published yet |
| Gemini 4 Argon Top model | 50% | 15% | Not published yet | |
| Kimi K3 | Moonshot | 48% | 53% | Not published yet |
| Grok 4.7 Top model | xAI | 47% | 29% | Not published yet |
| DeepSeek V4.1 Flash | DeepSeek | 46% | 96% | Not published yet |
| GPT-6 Luna | OpenAI | 44% | 77% | Not published yet |
| Muse Wonder Full Top model | Meta | 43% | 23% | Not published yet |
| Step 5 (preview) | StepFun | 42% | 43% | Not published yet |
| Muse Wonder | Meta | 41% | 27% | Not published yet |
| Muse Spark Open (open weights) | Meta | 39% | 24% | Not published yet |
| Claude Haiku 5.5 | Anthropic | 36% | 40% | Not published yet |
| MiMo-V2.6-Pro | Xiaomi | 35% | 41% | Not published yet |
| GLM-5.3 | Z.ai | 34% | 30% | Not published yet |
| Reflection Beam 1.0 | Reflection | 33% | 47% | Not published yet |
| Qwen3.8 Max Top model | Alibaba | 32% | 29% | Not published yet |
| Mistral Large 4 (preview) Top model | Mistral | 26% | 42% | Not published yet |
| Nemotron 3.5 Super VL | NVIDIA | 23% | 36% | Not published yet |
| MiniMax-M3 | MiniMax | 17% | 18% | Not published yet |
| Qwen3.8 27B (open weights) | Alibaba | 16% | 30% | Not published yet |
| DeepSeek V4 Pro Top model | DeepSeek | Not published yet | Not published yet | 8.6% |
Top models at a glance
Gets it right (share of 6,000 hard fact questions)
Makes things up when it doesn’t know (lower is better)
Sources
- Artificial Analysis, AA-Omniscience: Knowledge and Hallucination Benchmark. Independent tester. Figures copied Oct 16, 2026, for the setting Artificial Analysis lists for each model (for example “High” or “Max”).
- Vectara, Hallucination Leaderboard. Independent tester. Leaderboard last updated Sep 22, 2026; checked Oct 16, 2026.
Related: Is AI always right if it sounds sure? · How AI works: why it makes things up · Companies & models