BlogLLM Benchmark Scores as Published, Side by Side
LongFact: Results as Published
Every result the catalog holds for LongFact: 4 models from 3 developers, read from 2 pages. Each number is as printed by the company named beside it; nothing is ranked or rescored. Models are listed by developer, newest first.
Area General Knowledge and ReasoningMaintainer Jerry Wei and co-authors (Google DeepMind)Introduced 2024Data version 2026103001, as of 2026-10-30
LongFact is a prompt set for measuring factuality in long open-ended answers. GPT-4 generated the prompts across 38 topics, split into LongFact-Concepts and LongFact-Objects, with 30 prompts per topic and 1,140 prompts in each task. The SAFE method splits a response into individual facts and checks each fact by sending queries to Google Search and deciding whether the results support it. The aggregate score is an F1 that balances the percentage of supported facts with how many facts the response gives relative to a preferred length.
LongFact claim precision (LongFact claim-level precision)
LongFact claim-level precision
| Model | Value | Printed by | Setup as printed |
|---|---|---|---|
| Claude Sonnet 4.6Anthropic, released 2026-02-17 | 98% | Microsoft, Microsoft AI: MAI-Thinking-1 technical reportchecked 2026-10-19 | Effort: max |
| MAI-Thinking-1Microsoft, released 2026-08-12 | 98% | Microsoft, Microsoft AI: MAI-Thinking-1 technical reportchecked 2026-10-19 | - |
LongFact-Concepts (hallucination rate)
lower is better
| Model | Value | Printed by | Setup as printed |
|---|---|---|---|
| GPT-5OpenAI, released 2025-08-07 | 0.7% | OpenAI, Introducing GPT-5checked 2026-10-15 | Effort: high |
| o3OpenAI, released 2025-04-16 | 4.5% | OpenAI, Introducing GPT-5checked 2026-10-15 | Effort: high |
LongFact-Objects (hallucination rate)
lower is better
| Model | Value | Printed by | Setup as printed |
|---|---|---|---|
| GPT-5OpenAI, released 2025-08-07 | 0.8% | OpenAI, Introducing GPT-5checked 2026-10-15 | Effort: high |
| o3OpenAI, released 2025-04-16 | 5.1% | OpenAI, Introducing GPT-5checked 2026-10-15 | Effort: high |
Every number here is read from the same catalog file as the benchmark table, data version 2026103001, and is the number the table shows for that model and test. Where one page printed a model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.