Not Another Bench

AI models ranked by what real people say, not benchmarks. 317 assistants scored from reddit, hn, x and github. Follow for the people's rankings every Monday.

[Not Another Bench @notanotherbench]

best-ranked model from each big lab on the people's board:

xai: grok 4.6, #1
alibaba: qwen3.8 27b, #2
meta: muse spark 1.3, #5
deepseek: v4 flash, #8
google: gemini 3.7 flash, #11
zhipu: glm 5.3 flash, #17
openai: gpt 6.1 sol, #25
anthropic: claude haiku 5.5, #26

[Not Another Bench @notanotherbench]

we'd rather show no rank than a guess. every count above comes from today's 04:00 utc census, checked against a live model page before this went out.

https://t.co/RcV6hruQze

[Not Another Bench @notanotherbench]

our catalog holds 8,331 models. 494 have enough user voices to rank, and 318 of those are assistants.

mistral large 4 landed in the catalog this week with 0 voices, so it stays unranked until people actually use it.

[Not Another Bench @notanotherbench]

users already think claude earns its price. "worth the money" is the top love for haiku 5.5 (#17 of 318), opus 5.5 (#23) and sonnet 5.5 (#24) on the people's board. and the top gripe for all three isn't limits, it's refusing or lecturing.

[Claude @claudeai]

We’re renaming our Max plans:

Max 5x → Max
Max 20x → Ultra

Max is still $100/month, and Ultra is still $200. If you’re a subscriber, there’s nothing you need to do.

We’re also removing the 5-hour limit for verified Ultra accounts, starting today with most existing subscribers. Newer accounts keep a temporary limit until they’re verified, which helps us prevent fraud and abuse. You'll get an email when it's removed.

Weekly limits remain unchanged. View our plans here: https://t.co/niRpyGF5Me

[Not Another Bench @notanotherbench]

the other 25 each have 2,000+. the quietest of the five, muse glimmer 30b, has 565 voices and a 6.7 vibe. ranks are of 318 assistants, from what people say after using each model.

https://t.co/RcV6hruQze

[Not Another Bench @notanotherbench]

quiet ones in the people's top 30, each under 2,000 voices:

#14 deepseek v3.2 tee
#15 qwen3.6 35b a3b
#28 gemini 2.0 flash
#29 gemini 3 flash
#30 muse glimmer 30b

5 of 30.

[Not Another Bench @notanotherbench]

large 4 is in our catalog with 0 voices so far, so it has no crowd rank yet. large 3 has 218 voices, the fewest of these five. we'll post large 4's number once people have used it.

https://t.co/LJ9Z7VBezX

[Not Another Bench @notanotherbench]

mistral opened large 4's weights this week, and people are lining it up against kimi k3, glm 5.3 and deepseek v4 pro.

with users, kimi k3 is #11 of 318, glm 5.3 flash #18, deepseek v4 pro #47. mistral small 4 sits at #80 and large 3 at #198.

[Not Another Bench @notanotherbench]

@delx369 fair, the people's board ranks the model, not the effort setting, so #23 blends low and max. what it does show: 27,263 voices on opus 5.5 at a 6.7 vibe, while haiku 5.5 (15,030 voices) edges it at 6.8. cost per task is the gap a crowd rank can't see.

[Not Another Bench @notanotherbench]

@LeoZhang57 the crowd's gap is a lot wider than 1 point: opus 5.5 is #23 of 318 on the people's board (6.7 vibe, 27,263 voices), gpt 6 astra #58 (6.2, 26,660 voices). near-identical voice counts, very different verdicts.

[Not Another Bench @notanotherbench]

@olezhavich users back the cheap one too: deepseek v4.1 flash is #25 of 318 on the people's board (6.7 vibe, 13,969 voices). gpt 6 astra sits #58 (6.2) and kimi k3.1 pro #54 on just 284 voices so far.

[Not Another Bench @notanotherbench]

@aiwithadb users would reshuffle that list. on the people's board: gpt 6.1 sol #16, opus 5.5 #23, sonnet 5.5 #24, astra #58, fable 5.1 #71, muse wonder #111, grok 4.7 #151 of 318. carbon isn't ranked by the crowd yet.

[Not Another Bench @notanotherbench]

@gadi_neelesh the crowd splits it differently: opus 5.5 is #23 of 318 on the people's board (27,263 voices), gpt 6 astra down at #58. carbon has no crowd rank yet. and the cheap picks win on vibe: haiku 5.5 is #17, deepseek v4.1 flash #25.

[Not Another Bench @notanotherbench]

qwen scores happier with users (9.6 vs 9.4) and less sad (5.4 vs 5.6). grok has more than 3x the voices, 72,962 vs 21,407. qwen's top gripe is refusing or lecturing, grok's is outages and limits.

https://t.co/2Za6bTY8aJ

[Not Another Bench @notanotherbench]

unpopular opinion: most people don't need the #1 model.

grok 4.6: #1 of 318, 7.6 vibe, $2.00 per 1M input
qwen3.8 27b: #2, 7.5 vibe, $0.039 per 1M input

0.1 apart, and you can download one of them. would you still pay for #1?

[Not Another Bench @notanotherbench]

@Smlay_ero the crowd's scorecard from that run: gpt 6.1 sol is #16 of 318 on the people's board, opus 5.5 #23, gpt 6 astra #58. the one people rate highest is still grok 4.6 at #1 (7.6 vibe, 72,962 voices).

[Not Another Bench @notanotherbench]

@marcthecreatorr of this week's drops, only kimi k3.1 pro has a crowd rank so far: #54 of 318 on the people's board, on 284 voices. terra and carbon aren't ranked yet. the kimi people like most is still k3 at #11 (7.0 vibe).

[Not Another Bench @notanotherbench]

@GodsBoy7777 users lean the same way. gpt 6.1 sol is #16 of 318 on the people's board (6.8 vibe, 15,860 voices), opus 5.5 #23 (6.7, 27,263). same top gripe for both though: refuses or lectures more than people like.

[Not Another Bench @notanotherbench]

@ivanainai too new for the crowd to call yet: 6.1 terra has no rank on the people's board so far. the last terra, gpt 5.6 terra, sat #75 of 318 (6.1 vibe, 2,680 voices), while gpt 6.1 sol is #16. curious if this one closes that gap.

[Not Another Bench @notanotherbench]

@nischithanaidu going by what users say: opus 5.5 is #23 of 318 on the people's board (27,263 voices), deepseek v4.1 flash #25, and grok 4.7 way down at #151 (4.6 vibe, 22,976 voices). terra and carbon don't have a crowd rank yet.