RP Bench Leaderboard
AI roleplay models ranked by blind, head-to-head votes from Mana players during real Realism and Hardcore story turns.
| Rank | Model | ||||
|---|---|---|---|---|---|
1 |  Mimo v2.6 Pro | 1,285 ±72 1.5% refused | 60.8% | 1.5% | 102 |
2 |  Gemini 3.8 Flash Provisional | 1,276 ±76 10.6% refused | 60.8% | 10.6% | 97 |
3 |  Kimi k3 Provisional | 1,273 ±70 3.9% refused | 60.6% | 3.9% | 99 |
4 |  GLM 5.1 | 1,271 ±64 2.6% refused | 59.4% | 2.6% | 101 |
5 |  Claude Opus 5 | 1,268 ±65 | 60.0% | 0.0% | 100 |
6 |  Muse Spark 1.2 Contributor | 1,260 ±72 0.8% refused | 57.0% | 0.8% | 100 |
7 | Gemini 3 Flash | 1,253 ±64 1.7% refused | 57.3% | 1.7% | 103 |
8 | Kimi k3.1 Pro Provisional | 1,249 ±91 | 57.1% | 0.0% | 63 |
9 | Claude Sonnet 4.5 | 1,246 ±66 2.4% refused | 57.4% | 2.4% | 101 |
10 |  Aion 3 Mini Provisional | 1,238 ±73 4.7% refused | 56.1% | 4.7% | 98 |
11 | Claude Opus 5.5 | 1,237 ±61 4.8% refused | 54.3% | 4.8% | 105 |
12 | Aion 3 Provisional | 1,227 ±76 1.8% refused | 54.4% | 1.8% | 90 |
13 | GLM 5.3 Flash Provisional | 1,227 ±81 3.6% refused | 55.9% | 3.6% | 93 |
14 | Gemini 3.1 Pro | 1,216 ±56 | 52.0% | 0.0% | 100 |
15 |  Longcat 2.0 Provisional | 1,215 ±70 0.8% refused | 52.0% | 0.8% | 98 |
16 | Gemini 3.5 Flash Provisional | 1,212 ±66 1.6% refused | 51.1% | 1.6% | 94 |
17 | GLM 5 | 1,211 ±74 3.8% refused | 52.3% | 3.8% | 111 |
18 | GLM 5.2 | 1,209 ±68 0.8% refused | 51.5% | 0.8% | 101 |
19 | Claude Opus 4.6 | 1,208 ±54 0.8% refused | 52.3% | 0.8% | 109 |
20 | Claude Sonnet 5.5 Provisional | 1,203 ±58 2.5% refused | 50.6% | 2.5% | 89 |
21 |  Deepseek v4.1 Flash 0731 Provisional | 1,203 ±65 0.9% refused | 49.5% | 0.9% | 97 |
22 | Gemini 4 Argon Provisional | 1,203 ±80 8.1% refused | 50.0% | 8.1% | 82 |
23 | GLM 5.3 | 1,200 ±52 5.6% refused | 51.5% | 5.6% | 101 |
24 | Mimo v2.5 Pro Provisional | 1,192 ±65 4.2% refused | 48.0% | 4.2% | 98 |
25 | Deepseek v4 Pro 0813 Provisional | 1,189 ±66 | 49.5% | 0.0% | 99 |
26 |  Grok 4.2 Provisional | 1,188 ±79 0.9% refused | 47.9% | 0.9% | 94 |
27 | Claude Sonnet 4.6 Provisional | 1,177 ±65 1.0% refused | 46.0% | 1.0% | 87 |
28 |  MiniMax M3 | 1,167 ±62 0.8% refused | 45.2% | 0.8% | 104 |
29 | Deepseek v4.1 Flash 1017 Provisional | 1,165 ±95 | 45.7% | 0.0% | 70 |
30 | Deepseek v4 Pro Preview Provisional | 1,159 ±68 1.8% refused | 44.9% | 1.8% | 89 |
31 | Aion 2 | 1,153 ±60 2.1% refused | 43.1% | 2.1% | 116 |
32 | Grok 4.7 Provisional | 1,144 ±63 9.9% refused | 40.9% | 9.9% | 93 |
33 | Muse Wonder 1.0 Provisional | 1,129 ±79 2.9% refused | 40.5% | 2.9% | 84 |
34 | Muse Spark 1.3 Contributor Provisional | 1,128 ±76 3.6% refused | 37.9% | 3.6% | 87 |
35 | Deepseek v4 Flash 0731 | 1,112 ±65 1.7% refused | 37.3% | 1.7% | 102 |
36 | Grok 4.6 Provisional | 1,079 ±82 3.4% refused | 31.2% | 3.4% | 93 |
37 | Claude Haiku 5.5 Provisional | 980 ±101 15.9% refused | 20.5% | 15.9% | 73 |
About the RP Bench leaderboard
RP Bench ranks AI language models on how well they write interactive roleplay. Every score comes from a blind vote: on a share of turns in Realism and Hardcore visual stories on Mana, two anonymous models answer the same message with the same story, characters and memories, and the player picks the response they prefer. The model names are revealed only after the pick, so no vote is swayed by a name. Every model runs with its reasoning effort set to High.
How to read the scores
Scores are Elo-style ratings fitted from all counted votes at once with a Bradley–Terry model, centred at 1200. A 100-point lead means the higher model is picked about 64% of the time the two meet. The ± is a 95% range: when two models’ ranges overlap, the votes cannot separate them yet. Win rate is the share of a model’s counted votes it won, and Votes is how many counted votes it has. Refusal Rate, in red, is the share of a model’s RP Bench responses that were safety refusals: it declined to write the reply, or it dodged the scene and the player marked it as breaking immersion because of a safety filter. A refusal never counts as a vote, so it does not change the score. Models with fewer than 100 counted votes are marked Provisional. The leaderboard is refitted about once an hour.
Safety Filter / Refusal
Some models decline to write a scene because of a safety filter. Every RP Bench response is checked for this before you see it: a quick text check catches plain refusals such as “I can’t continue this roleplay”, and an AI check reads each response next to your message and decides whether it narrated the scene or refused it. A refusal never reaches the vote. If one model refuses, the story continues with the other model’s response and no vote is asked for. If both refuse, your story’s own model writes the turn, as a normal turn.
Other models keep writing but steer the story away from the scene, which can get past the check. For those, the vote popup lets you mark a response as “Broke Narrative Immersion due to safety filter” instead of picking. The story continues with the other response, the flagged response counts as a refusal, and no vote is recorded. A flag made in under 5 seconds is not counted, the same as a pick.
Every refusal, whether the check caught it or a player flagged it, counts toward that model’s Refusal Rate. Refusals never change a score: scores come only from votes where both responses narrated the scene. A response that never arrives (an error, a timeout or an empty reply) is a failed call, not a refusal, so it does not count toward the Refusal Rate either.
Which models are ranked
The pool is a selection of the models in Mana’s story model picker; the default Free and Paid models are never in it. Models with fewer votes are matched more often, so a new model is rated quickly. A model Mana no longer offers or takes out of the benchmark is taken off the table, but its votes still count toward the scores of the models it met. Picks made in under 5 seconds are not counted, and a turn where a model fails or declines to write is never shown as a vote. The full rules, including what an RP Bench turn costs and how to take part, are in How RP Bench Works.