Skip to main content
BenchLM
Data
Specialist vertical · BenchAlign v6.0 · 65 ranked (64 Supported, 1 Estimated)
Scored from the dataset of

Factuality and anti-hallucination leaderboard

As of November 21, 2026, Gemini 4 Carbon leads our factuality and anti-hallucination leaderboard at 78.8 (Supported, $2/$10 per 1M tokens), ahead of GPT-6 Astra at 77.9. Gemini 3.8 Flash (#6, 72.6) is the lowest-cost model near the top at $0.75/$3.75 per 1M tokens; DeepSeek V4.1 Flash 1017 leads open-weight models at #18 (61.9). We rank 65 models (64 Supported, 1 Estimated) across 10 domain protocols, and this vertical carries zero weight in the overall score.

Shortlist by workload and budget

Each pick comes directly from the live factuality rows below, with listed API rates and the case where another option wins.

Highest point score#1 · Supported

Google · $2/$10 per 1M tokens

Costs $2/$10 per 1M tokens; Gemini 3.8 Flash (#6, 72.6) cuts token spend to $0.75/$3.75 per 1M tokens.

Closest contender#2 · Supported

OpenAI · $10/$50 per 1M tokens

Sits 0.9 points behind Gemini 4 Carbon; test both on your own prompts when provider lock-in or latency matters.

Best API value#6 · Supported

Google · $0.75/$3.75 per 1M tokens

Keeps 92% of the top score (72.6 vs 78.8), but gives up 6.2 points on the hardest tail cases.

Top open-weight model#18 · Supported

DeepSeek · Open weights · $0.3/$1.2 per 1M tokens

Trails Gemini 4 Carbon by 16.9 points (#18 overall here), so pick it when weight control or on-prem residency outweighs peak score.

Ranked models

Ordered by BenchAlign v6.0 point estimate as of November 21, 2026. An Estimated row backed by fewer than two source families never ranks above a row scoring within 2 points below it. Conditional ranges do not establish rank confidence.

Factuality & Anti-Hallucination LLM Leaderboard
RankModelScoreEvidence
#1Google78.8Supported
#2OpenAI77.9Supported
#3Google77.1Supported
#4Google76.9Supported
#5OpenAI76.4Supported
#6Google72.6Supported
#7OpenAI72.3Supported
#8Google72.0Supported
#9OpenAI71.7Supported
#10OpenAI71.2Supported
#11Google70.9Supported
#12Google69.9Supported
#13Google68.0Supported
#14Anthropic66.3Supported
#15Anthropic65.9Supported
#16Google64.7Supported
#17OpenAI64.6Supported
#18DeepSeek61.9Supported
#19OpenAI61.3Supported
#20Anthropic61.3Supported
#21OpenAI61.1Supported
#22Anthropic61.0Supported
#23xAI60.4Supported
#24xAI59.3Supported
#25
GLM-5.3Open weights
Z.AI
59.0Supported
#26Anthropic58.9Supported
#27Anthropic58.9Supported
#28xAI58.8Supported
#29Anthropic58.5Supported
#30OpenAI58.1Supported
#31OpenAI57.2Supported
#32
Qwen3.8 MaxOpen weights
Alibaba
57.2Supported
#33Anthropic55.6Supported
#34Anthropic55.6Supported
#35OpenAI54.9Supported
#36Google54.6Supported
#37OpenAI53.1Supported
#38OpenAI52.8Supported
#39Anthropic51.8Supported
#40Anthropic51.6Supported
#41OpenAI49.7Supported
#42Google48.9Supported
#43
DeepSeek-R1Open weights
DeepSeek
47.5Supported
#44
DeepSeek V3Open weights
DeepSeek
47.2Supported
#45Anthropic46.7Supported
#46Anthropic46.4Supported
#47Anthropic46.4Supported
#48
DeepSeek V3.1Open weights
DeepSeek
46.3Supported
#49Google45.3Supported
#50Mistral44.2Supported
#51xAI43.7Supported
#52xAI43.4Supported
#53
Gemma 4 31BOpen weights
Google
43.0Supported
#54
Gemma 4 26B A4BOpen weights
Google
40.9Supported
#55Anthropic40.8Supported
#56Anthropic40.5Supported
#57Google40.3Supported
#58OpenAI40.3Supported
#59
GPT-OSS 120BOpen weights
OpenAI
38.3Supported
#60
Gemma 3 27BOpen weights
Google
37.4Supported
#61OpenAI37.0Supported
#62OpenAI33.6Supported
#63Moonshot AI33.3Estimated
#64
GPT-OSS 20BOpen weights
OpenAI
32.8Supported
#65OpenAI32.2Supported

Scored domain benchmarks (10)

Each percentage is the protocol's relative reference weight inside this vertical's calibrated BenchAlign v6.0 model, not a raw score average. Independent evaluator shares are capped at 40% whenever three or more independent evaluators contribute.

A benchmark that tests whether AI models challenge nonsensical, ill-posed, or logically flawed prompts instead of confidently generating incorrect answers. Measures the critical ability to push back on bad input.

16% ref. weight

FACTS Grounding v2 checks whether long-form answers remain supported by supplied documents. Its revised judge ensemble and public/private results remain separate from version 1.

16% ref. weight

Kaggle's continuously updated SimpleQA Verified board, with 56 configurations against the 13 rows transcribed from the author paper.

16% ref. weight

The official FACTS suite combines document grounding, visual factuality, parametric recall, and search-assisted answering. The published average stays separate from its component scores.

15% ref. weight

FACTS Parametric asks hard factoid questions from real Google searches and checks whether a model answers them from memory, without tools. It is the parametric-knowledge part of the FACTS suite.

14% ref. weight

Kaggle's run of OpenAI's original SimpleQA, separate from provider-reported SimpleQA rows and from the Verified subset.

13% ref. weight

A verified and revised Humanity's Last Exam set that removes uncertain items and repairs fixable questions before evaluating frontier expert reasoning.

5% ref. weight

The archived FACTS Grounding V1 protocol tests factual responses grounded in supplied documents. Its original judge setup and results remain separate from Grounding V2.

2% ref. weight

SimpleQA Verified measures short-form factual recall on 1,000 prompts without tools. This dated author-paper table ranks source models by F1 and keeps accuracy, accuracy given attempted, attempted, and hedged separate.

2% ref. weight

A benchmark that evaluates the ability of language models to answer short, fact-seeking questions accurately. Focuses on factual correctness rather than reasoning complexity.

1% ref. weight

Grounded faithfulness, parametric precision, and pushback against false premises

Ten protocols drive this ranking, led by BullshitBench v2 (16%), FACTS Grounding v2 (16%), SimpleQA Verified (Kaggle) (16%) together with the full FACTS Benchmark Suite, FACTS Parametric, and HLE-Verified. Unlike broad knowledge exams that reward confident guessing on multiple-choice items, these suites penalize unforced fabrications: SimpleQA Verified grades short-form factual precision, FACTS Grounding checks whether every claim stays inside the supplied passage, and BullshitBench v2 tests whether a model challenges nonsensical premises instead of playing along.

Sixty-three of the 65 ranked models carry Supported status here because independent Kaggle and benchmark-owner evaluations cover almost the entire frontier cohort on identical prompts.

Low benchmark hallucination does not fix stale cutoffs or broken retrieval

Even Gemini 4 Carbon at 78.8 still misses roughly one in five frontier factuality stress items on raw benchmark scales. Parametric recall freezes at each model's training cutoff, and grounded faithfulness benchmarks only verify consistency with the text you paste into the prompt—if your retriever fetches an outdated policy page, a faithful model will repeat that outdated rule accurately.

Treat this leaderboard as a measure of calibration discipline and instruction adherence under uncertainty, not as a substitute for citation verification, live search, or deterministic database lookups on regulated facts.

Picking a factual model for RAG pipelines versus closed-book chat

Gemini 4 Carbon (78.8, Supported, $2/$10 per 1M tokens) and GPT-6 Astra (77.9, Supported, $10/$50 per 1M tokens) sit 0.9 points apart at the top. If your application supplies retrieved passages and runs millions of daily checks, Gemini 3.8 Flash (#6, 72.6 at $0.75/$3.75 per 1M tokens) preserves over 90% of the leader's calibrated score at a fraction of the API bill.

Open-weight buyers should note the wider spread on closed-book recall: DeepSeek V4.1 Flash 1017 leads open weights at #18 (61.9), because smaller parameter counts store fewer long-tail entity facts even when their grounded RAG discipline is solid.

Questions

Which LLM is best for factuality right now?

Gemini 4 Carbon holds #1 on our factuality leaderboard at 78.8 (Supported) as of November 21, 2026, ahead of GPT-6 Astra at 77.9 and Gemini 4 Argon at 77.1. Both top models hold Supported status across independent public evaluations. Check the ranked table and benchmark weights on this page to verify source coverage before production deployment.

What is the best open-weight LLM for factuality?

DeepSeek V4.1 Flash 1017 from DeepSeek leads open-weight models on this factuality board at #18 with a calibrated score of 61.9 (Supported). That trails Gemini 4 Carbon by 16.9 points, so choose DeepSeek V4.1 Flash 1017 when self-hosting, VPC data residency, or weight customization matters more than peak frontier accuracy.

Which API model offers the best value for factuality?

Gemini 3.8 Flash (#6, 72.6) is our top API value pick for factuality, priced at $0.75/$3.75 per 1M tokens compared with Gemini 4 Carbon at $2/$10 per 1M tokens. You give up 6.2 calibrated points, making Gemini 3.8 Flash the better buy whenever those extra points do not cut downstream retries.

Which benchmarks make up the factuality leaderboard score?

BenchAlign v6.0 calibrates 10 open and primary domain protocols for this board, led by BullshitBench v2 (16%), FACTS Grounding v2 (16%), SimpleQA Verified (Kaggle) (16%), FACTS Benchmark Suite (15%). Each percentage reflects share of the vertical's reference weight inside our item-response model rather than a raw score average, and no single independent evaluator can exceed 40% when three

What is the difference between parametric recall and grounded factuality?

Parametric recall tests whether a model answers closed-book questions accurately from memory without guessing, as measured by SimpleQA Verified and FACTS Parametric. Grounded factuality tests whether answers stay strictly inside supplied source documents without inventing unsupported claims, measured by FACTS Grounding and BullshitBench v2.

Does a high factuality score eliminate hallucinations in production?

No. Even Gemini 4 Carbon at 78.8 still produces errors on adversarial closed-book prompts and cannot know events after its training cutoff without retrieval. Pair top-ranked models with verified source passages, require inline citations, and abstain or route to human review when retrieved context is missing.

Factuality leaderboard updates

Weekly updates when new models or domain evaluations move the factuality rankings.

Read a sample issue

Join 7,000+ subscribers.

One email each week. Unsubscribe anytime.