GenuineBench

July 2026

GenuineBench

How often do LLMs say genuinely?

We gave 13 frontier models 80 everyday writing tasks — texts, tweets, emails, LinkedIn posts — 5× each at default settings, and counted. claude-haiku-4-5 says it in 38% of everything it writes.

Anthropic Everyone else

The top cluster — four Claudes plus DeepSeek and Kimi — is clearly separated from the rest of the field; the ordering within it sits inside the error bars.

The receipts

# Tweet just hit my 1-month streak with [app name] and i'm genuinely impressed 🚀 the [specific feature] has completely changed h
claude-haiku-4-5 · tweet-05
# Book Launch Thank You Tweet Here's a warm, genuine option: --- 🎉 What an incredible night! Thank you to ever
claude-sonnet-5 · tweet-03
u can customize: --- One month in with [App Name] and I'm genuinely impressed. My to-do list actually gets *done* now instead o
claude-fable-5 · tweet-05
Full results & method
ModelProvider% w/ hit95% CIPer 1k words% w/ "genuinely" onlyHitsn
claude-haiku-4-5Anthropic38.0%[32.0, 44.5]2.6625.2%203400
claude-sonnet-5Anthropic35.5%[29.2, 41.5]1.7321.8%182400
claude-fable-5Anthropic35.2%[29.0, 41.5]1.5924.5%173400
deepseek-v4-proDeepSeek34.3%[28.2, 40.2]1.2120.8%166399
claude-opus-4-8Anthropic33.2%[27.5, 39.2]1.6420.0%164400
kimi-k2.6Moonshot AI31.6%[26.2, 37.2]1.1221.6%149399
grok-4.5xAI25.8%[20.0, 31.5]0.969.2%115400
qwen3.7-maxAlibaba15.5%[10.8, 20.5]0.379.2%74400
gemini-3.1-pro-previewGoogle15.2%[10.0, 19.2]0.409.2%43250
gpt-5.4-miniOpenAI11.5%[7.2, 16.0]0.598.0%47400
gemini-3.5-flashGoogle11.2%[7.5, 15.2]0.293.5%50400
gpt-5.5OpenAI11.0%[6.2, 16.2]0.528.2%44400
llama-4-maverickMeta2.5%[0.8, 4.8]0.090.8%10400
claude-sonnet-4-5 prev genAnthropic22.5%[17.2, 28.2]1.2510.8%100400
o3 prev genOpenAI15.8%[11.2, 20.8]0.587.2%64400

Method. 80 composition prompts across 8 categories (texts, tweets, emails, LinkedIn, one-pagers, reviews, essays, speeches), 5 samples per prompt per model, provider-default settings, no system prompt, no prompt containing the target word.

Grading is deterministic — a case-insensitive word-boundary regex for genuine / genuinely / genuineness; no LLM judge. The only randomness is the models' own sampling, which the repeated samples average over. CIs: 95% stratified bootstrap over prompts within categories (n=10,000). Rankings are unchanged if you count only the adverb "genuinely" (see table).

Is this new? Mixed. Previous-generation models on the same tasks: claude-sonnet-4-5 22% → claude-sonnet-5 36% · o3 16% → gpt-5.5 11%. Claude's rate rose generation-over-generation; OpenAI's fell.

The word clusters in performative formats — reviews, essays, LinkedIn, tweets — and is rare in casual texts and structured one-pagers. Models: 13 current-generation across Anthropic, OpenAI, Google, xAI, DeepSeek, Moonshot AI, Alibaba, and Meta, plus 2 previous-generation for the trend. Sample-size notes: gemini-3.1-pro-preview n=250 (API quota), kimi-k2.6 n=399; one empty deepseek response excluded.