Itzi Roleplay Benchmark: Local Robustness
Models play eight scripted episodes. Judges check whether each model remembers instructions, respects player decisions, and keeps characters and locations consistent.
The maker reports local robustness for 36 models. It counts missing episode criteria as failures. External Arena scores supply the other 75% of overall quality.
- Made by
- Itzi
- Released
- September 2026
- Kind
- Creative writing
- Level
- Everyday
- Scored
- Graded by a model
- Answers
- Open-ended
- Measures
- Capability
- Subject
- Languages and literature
- Language
- English
- Interaction
- Multi-turn
- Refresh
- One fixed set
- Run by
- A host
- Strongest model tested
- GPT-6 Astra #4 on the capability index
- Last model added
- September 2026
- Scores refreshed
- November 20th, 2026
6 models share the top score of 100.0 among models released by September 2026
Best score per model from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.
The newest model with a score here came out in September 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6.1 Sol
- and 9 more
- 100.0 MiMo-V2.5
- 100.0 MiMo-V2.5-Pro
- 100.0 Claude Opus 4.8
- 100.0
Qwen 3.8 Max
- 100.0 GPT-6 Astra
- 100.0 Claude Opus 5.5
- 94.9 GPT-5.5
- 94.9 GPT-5.6 Sol
- 94.9
Gemini 3.6 Flash
- 94.9 GLM-5.3
- 94.9 GLM-5.3-Flash
- 94.9
Gemini 3.8 Flash
- 94.9 MiMo-V2.6 Pro
- 92.3 Claude Opus 4.7
- 92.3
MiniMax-M3
- 92.3 GPT-5.6 Terra
- 92.3 Claude Opus 5
- 92.3 Claude Fable 5.1
- 87.2 Claude Sonnet 4.6
- 87.2
Gemini 3.1 Pro
- 87.2 GLM-5.1
- 87.2 Claude Sonnet 5
- 84.6
Qwen 3.8 27B
- 79.5 Claude Fable 5
- 79.5 GPT-5.6 Luna
- 79.5 Kimi K3
- 77.8
Gemini 3.7 Flash
- 76.9
DeepSeek-V4-Flash
- 76.9
DeepSeek-V4-Pro
- 76.9 MiMo-V2.6 Flash
- 71.8 Claude Opus 4.6
- 71.8
Gemma 4 31B IT
- 66.7
Gemma 4 26B A4B
- 64.1 GLM-5.2
- 25.6 Grok 4.6
- 12.8 Grok 4.5
- 1 MiMo-V2.5 100.0
- 2 MiMo-V2.5-Pro 100.0
- 3 Claude Opus 4.8 100.0
- 4
Qwen 3.8 Max 100.0
- 5 GPT-6 Astra 100.0
- 6 Claude Opus 5.5 100.0
- 7 GPT-5.5 94.9
- 8 GPT-5.6 Sol 94.9
- 9
Gemini 3.6 Flash thinking 94.9
- 10 GLM-5.3 94.9
- 11 GLM-5.3-Flash 94.9
- 12
Gemini 3.8 Flash thinking 94.9
- 13 MiMo-V2.6 Pro 94.9
- 14 Claude Opus 4.7 92.3
- 15
MiniMax-M3 92.3
- 16 GPT-5.6 Terra 92.3
- 17 Claude Opus 5 92.3
- 18 Claude Fable 5.1 92.3
- 19 Claude Sonnet 4.6 87.2
- 20
Gemini 3.1 Pro 87.2
- 21 GLM-5.1 87.2
- 22 Claude Sonnet 5 87.2
- 23
Qwen 3.8 27B 84.6
- 24 Claude Fable 5 79.5
- 25 GPT-5.6 Luna 79.5
- 26 Kimi K3 79.5
- 27
Gemini 3.7 Flash thinking 77.8
- 28
DeepSeek-V4-Flash 76.9
- 29
DeepSeek-V4-Pro 76.9
- 30 MiMo-V2.6 Flash 76.9
- 31 Claude Opus 4.6 71.8
- 32
Gemma 4 31B IT 71.8
- 33
Gemma 4 26B A4B 66.7
- 34 GLM-5.2 64.1
- 35 Grok 4.6 25.6
- 36 Grok 4.5 12.8
Source: Official Itzi Roleplay Benchmark: Local Robustness leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | MiMo-V2.5 | 100.0 | Apr 2026 | Official leaderboard | |
| 2 | MiMo-V2.5-Pro | 100.0 | Apr 2026 | Official leaderboard | |
| 3 | Claude Opus 4.8 | 100.0 | May 2026 | Official leaderboard | |
| 4 | | 100.0 | Aug 2026 | Official leaderboard | |
| 5 | GPT-6 Astra | 100.0 | Sep 2026 | Official leaderboard | |
| 6 | Claude Opus 5.5 | 100.0 | Sep 2026 | Official leaderboard | |
| 7 | GPT-5.5 | 94.9 | Apr 2026 | Official leaderboard | |
| 8 | GPT-5.6 Sol | 94.9 | Jul 2026 | Official leaderboard | |
| 9 | | 94.9 | thinking | Jul 2026 | Official leaderboard |
| 10 | GLM-5.3 | 94.9 | Aug 2026 | Official leaderboard | |
| 11 | GLM-5.3-Flash | 94.9 | Aug 2026 | Official leaderboard | |
| 12 | | 94.9 | thinking | Sep 2026 | Official leaderboard |
| 13 | MiMo-V2.6 Pro | 94.9 | Sep 2026 | Official leaderboard | |
| 14 | Claude Opus 4.7 | 92.3 | Apr 2026 | Official leaderboard | |
| 15 | | 92.3 | Jun 2026 | Official leaderboard | |
| 16 | GPT-5.6 Terra | 92.3 | Jul 2026 | Official leaderboard | |
| 17 | Claude Opus 5 | 92.3 | Jul 2026 | Official leaderboard | |
| 18 | Claude Fable 5.1 | 92.3 | Sep 2026 | Official leaderboard | |
| 19 | Claude Sonnet 4.6 | 87.2 | Feb 2026 | Official leaderboard | |
| 20 | | 87.2 | Feb 2026 | Official leaderboard | |
| 21 | GLM-5.1 | 87.2 | Apr 2026 | Official leaderboard | |
| 22 | Claude Sonnet 5 | 87.2 | Jun 2026 | Official leaderboard | |
| 23 | | 84.6 | Aug 2026 | Official leaderboard | |
| 24 | Claude Fable 5 | 79.5 | Jun 2026 | Official leaderboard | |
| 25 | GPT-5.6 Luna | 79.5 | Jul 2026 | Official leaderboard | |
| 26 | Kimi K3 | 79.5 | Jul 2026 | Official leaderboard | |
| 27 | | 77.8 | thinking | Aug 2026 | Official leaderboard |
| 28 | | 76.9 | Apr 2026 | Official leaderboard | |
| 29 | | 76.9 | Apr 2026 | Official leaderboard | |
| 30 | MiMo-V2.6 Flash | 76.9 | Sep 2026 | Official leaderboard | |
| 31 | Claude Opus 4.6 | 71.8 | Feb 2026 | Official leaderboard | |
| 32 | | 71.8 | Apr 2026 | Official leaderboard | |
| 33 | | 66.7 | Apr 2026 | Official leaderboard | |
| 34 | GLM-5.2 | 64.1 | Jun 2026 | Official leaderboard | |
| 35 | Grok 4.6 | 25.6 | Aug 2026 | Official leaderboard | |
| 36 | Grok 4.5 | 12.8 | Jul 2026 | Official leaderboard |
The best score, 100.0, costs $0.005 for the full run
Runs on the cost frontier score higher than any run that's cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.
Source: Official Itzi Roleplay Benchmark: Local Robustness leaderboard
| Model | Effort | Harness | Score | Cost of the full run | Output tokens |
|---|---|---|---|---|---|
| MiMo-V2.5 | 100.0 | $0.005 | |||
| MiMo-V2.5-Pro | 100.0 | $0.01 | |||
| Claude Opus 4.8 | 100.0 | $0.94 | |||
| 100.0 | $0.46 | ||||
| GPT-6 Astra | 100.0 | $1.40 | |||
| Claude Opus 5.5 | 100.0 | $0.88 | |||
| GPT-5.5 | 94.9 | $0.60 | |||
| GPT-5.6 Sol | 94.9 | $1.02 | |||
| thinking | 94.9 | $0.14 | |||
| GLM-5.3 | 94.9 | $0.29 | |||
| GLM-5.3-Flash | 94.9 | $0.04 | |||
| thinking | 94.9 | $0.08 | |||
| MiMo-V2.6 Pro | 94.9 | $0.02 | |||
| Claude Opus 4.7 | 92.3 | $1.18 | |||
| 92.3 | $0.02 | ||||
| GPT-5.6 Terra | 92.3 | $0.38 | |||
| Claude Opus 5 | 92.3 | $0.75 | |||
| Claude Fable 5.1 | 92.3 | $1.35 | |||
| Claude Sonnet 4.6 | 87.2 | $0.17 | |||
| 87.2 | $0.32 | ||||
| GLM-5.1 | 87.2 | $0.07 | |||
| Claude Sonnet 5 | 87.2 | $0.85 | |||
| 84.6 | $0.16 | ||||
| Claude Fable 5 | 79.5 | $2.37 | |||
| GPT-5.6 Luna | 79.5 | $0.04 | |||
| Kimi K3 | 79.5 | $0.20 | |||
| thinking | 77.8 | $0.05 | |||
| 76.9 | $0.01 | ||||
| 76.9 | $0.03 | ||||
| MiMo-V2.6 Flash | 76.9 | $0.004 | |||
| Claude Opus 4.6 | 71.8 | $0.37 | |||
| 71.8 | $0.006 | ||||
| 66.7 | $0.005 | ||||
| GLM-5.2 | 64.1 | $0.07 | |||
| Grok 4.6 | 25.6 | $0.06 | |||
| Grok 4.5 | 12.8 | $0.03 |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official Itzi Roleplay Benchmark: Local Robustness leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official Itzi Roleplay Benchmark: Local Robustness leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Writing and taste
- Design Bench (Moonshot) In use Models make graphic designs from direct prompts or few-shot reference images. An automated grader rates the results as Elo scores. Kimi K3.1 Pro 957
- WritingBench Fading Models complete 1,000 real-world writing requests across 6 domains and 100 subdomains. A judge model scores each reply on five query-specific criteria. GPT-5 83.9
- Creative Writing (MiniMax in-house) Only in system cards Models write creative pieces for real Hailuo AI user requests. Automatic assessments score the writing.
MiniMax-Text-01 81.3%
- DeepSeek Chinese functional writing evaluation Only in system cards Models answer 3,170 everyday Chinese writing prompts such as reports, emails and speeches. Pairwise judging against a baseline model gives a win rate.
DeepSeek-V4-Pro 62.6%
- Slide generation human evaluation (Z.ai) Only in system cards People compare HTML slide decks from two models on content, layout and visual quality. Metric is overall win rate. GLM-5 67.5%
- Task-dependent output homogenization evaluation Only in system cards Models answer prompts several times. A judge scores functional diversity where it is wanted and consistency where one answer is right.
Muse Spark 0.21
Data audit
9 of 10 checks pass, 1 warning. Look first at the Task access check. We don’t know whether the tasks are public. Every benchmark
Catalog record 1 warning 1 pass
- Warning Task access We don’t know whether the tasks are public.
- Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
- Pass Independent runs Scores from the official leaderboard (36 entries).
- Pass Models scored 36 models we track.
- Pass Newest score The newest is from September 2026, for Claude Opus 5.5.
- Pass Frontier models 3 of the top 10 have a score, led by GPT-6 Astra (#4).
- Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
- Pass Importer We read the board on itzi.app: 36 rows for models we track at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.