-0.04
Confirmed task success · rank 25 of 46
- Unit
- IPS effect estimate, higher is better
- Range
- -0.06 to -0.03
- Sample
- 64406 observations
- Configuration
- GPT-5.5
- Measured
- 28 Sep 2026
- Not shown
- Not a success rate: the estimated effect of choosing this model as the agent's orchestrator, against an even mix of the models tested, inside Arena's own harness.
-0.03
Confirmed task success · rank 23 of 46
- Unit
- IPS effect estimate, higher is better
- Range
- -0.05 to -0.01
- Sample
- 42185 observations
- Configuration
- GPT-5.5
- Measured
- 28 Sep 2026
- Not shown
- Not a success rate: the estimated effect of choosing this model as the agent's orchestrator, against an even mix of the models tested, inside Arena's own harness.
0.05
Praise over complaint · rank 9 of 46
- Unit
- IPS effect estimate, higher is better
- Range
- 0.02 to 0.08
- Sample
- 21366 observations
- Configuration
- GPT-5.5
- Measured
- 28 Sep 2026
- Not shown
- Not a success rate: the estimated effect of choosing this model as the agent's orchestrator, against an even mix of the models tested, inside Arena's own harness.
0.08
Praise over complaint · rank 8 of 46
- Unit
- IPS effect estimate, higher is better
- Range
- 0.05 to 0.11
- Sample
- 14033 observations
- Configuration
- GPT-5.5
- Measured
- 28 Sep 2026
- Not shown
- Not a success rate: the estimated effect of choosing this model as the agent's orchestrator, against an even mix of the models tested, inside Arena's own harness.
0.02
Steerability · rank 11 of 46
- Unit
- IPS effect estimate, higher is better
- Range
- 0.01 to 0.04
- Sample
- 77644 observations
- Configuration
- GPT-5.5
- Measured
- 28 Sep 2026
- Not shown
- Not a success rate: the estimated effect of choosing this model as the agent's orchestrator, against an even mix of the models tested, inside Arena's own harness.
0.07
Steerability · rank 4 of 46
- Unit
- IPS effect estimate, higher is better
- Range
- 0.05 to 0.09
- Sample
- 47527 observations
- Configuration
- GPT-5.5
- Measured
- 28 Sep 2026
- Not shown
- Not a success rate: the estimated effect of choosing this model as the agent's orchestrator, against an even mix of the models tested, inside Arena's own harness.
0.00
Tool grounding · rank 1 of 46
- Unit
- IPS effect estimate, higher is better
- Range
- 0.00 to 0.00
- Sample
- 1990689 observations
- Configuration
- GPT-5.5
- Measured
- 28 Sep 2026
- Not shown
- Not a success rate: the estimated effect of choosing this model as the agent's orchestrator, against an even mix of the models tested, inside Arena's own harness.
0.00
Tool grounding · rank 1 of 46
- Unit
- IPS effect estimate, higher is better
- Range
- 0.00 to 0.00
- Sample
- 2518969 observations
- Configuration
- GPT-5.5
- Measured
- 28 Sep 2026
- Not shown
- Not a success rate: the estimated effect of choosing this model as the agent's orchestrator, against an even mix of the models tested, inside Arena's own harness.
1,483
Overall · rank 2 of 44
- Unit
- Arena rating, higher is better
- Range
- 1,476 to 1,489
- Sample
- 21279 votes
- Configuration
- GPT-5.5
- Measured
- 13 Sep 2026
- Not shown
- Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
1,480
Overall · rank 5 of 44
- Unit
- Arena rating, higher is better
- Range
- 1,474 to 1,487
- Sample
- 20944 votes
- Configuration
- GPT-5.5 (high)
- Measured
- 13 Sep 2026
- Not shown
- Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
1,224
Overall · rank 2 of 34
- Unit
- Arena rating, higher is better
- Range
- 1,218 to 1,229
- Sample
- 89873 votes
- Configuration
- GPT-5.5
- Measured
- 24 Aug 2026
- Not shown
- Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
1,482
Overall · rank 4 of 177
- Unit
- Arena rating, higher is better
- Range
- 1,480 to 1,485
- Sample
- 70523 votes
- Configuration
- GPT-5.5
- Measured
- 25 Sep 2026
- Not shown
- Not an error rate: claims that cannot be checked on the web are skipped, and preference still carries most of the weight.
1,486
Overall · rank 3 of 177
- Unit
- Arena rating, higher is better
- Range
- 1,483 to 1,489
- Sample
- 68926 votes
- Configuration
- GPT-5.5 (high)
- Measured
- 25 Sep 2026
- Not shown
- Not an error rate: claims that cannot be checked on the web are skipped, and preference still carries most of the weight.
1,486
Business, management and finance · rank 7 of 402
- Unit
- Arena rating, higher is better
- Range
- 1,479 to 1,492
- Sample
- 14097 votes
- Configuration
- GPT-5.5
- Measured
- 25 Sep 2026
- Not shown
- Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
1,489
Business, management and finance · rank 3 of 402
- Unit
- Arena rating, higher is better
- Range
- 1,483 to 1,496
- Sample
- 13739 votes
- Configuration
- GPT-5.5 (high)
- Measured
- 25 Sep 2026
- Not shown
- Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
1,449
Creative writing · rank 20 of 407
- Unit
- Arena rating, higher is better
- Range
- 1,442 to 1,456
- Sample
- 13078 votes
- Configuration
- GPT-5.5
- Measured
- 25 Sep 2026
- Not shown
- Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
1,452
Creative writing · rank 18 of 407
- Unit
- Arena rating, higher is better
- Range
- 1,445 to 1,458
- Sample
- 12802 votes
- Configuration
- GPT-5.5 (high)
- Measured
- 25 Sep 2026
- Not shown
- Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
1,511
Expert prompts · rank 10 of 359
- Unit
- Arena rating, higher is better
- Range
- 1,503 to 1,519
- Sample
- 7644 votes
- Configuration
- GPT-5.5
- Measured
- 25 Sep 2026
- Not shown
- Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
1,513
Expert prompts · rank 10 of 359
- Unit
- Arena rating, higher is better
- Range
- 1,505 to 1,520
- Sample
- 7384 votes
- Configuration
- GPT-5.5 (high)
- Measured
- 25 Sep 2026
- Not shown
- Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
1,475
Instruction following · rank 13 of 409
- Unit
- Arena rating, higher is better
- Range
- 1,470 to 1,480
- Sample
- 24807 votes
- Configuration
- GPT-5.5
- Measured
- 25 Sep 2026
- Not shown
- Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
1,478
Instruction following · rank 11 of 409
- Unit
- Arena rating, higher is better
- Range
- 1,473 to 1,483
- Sample
- 24017 votes
- Configuration
- GPT-5.5 (high)
- Measured
- 25 Sep 2026
- Not shown
- Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
1,477
Overall · rank 19 of 409
- Unit
- Arena rating, higher is better
- Range
- 1,473 to 1,480
- Sample
- 70635 votes
- Configuration
- GPT-5.5
- Measured
- 25 Sep 2026
- Not shown
- Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
1,481
Overall · rank 13 of 409
- Unit
- Arena rating, higher is better
- Range
- 1,478 to 1,485
- Sample
- 69008 votes
- Configuration
- GPT-5.5 (high)
- Measured
- 25 Sep 2026
- Not shown
- Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
1,467
Writing, literature and language · rank 15 of 408
- Unit
- Arena rating, higher is better
- Range
- 1,461 to 1,472
- Sample
- 18231 votes
- Configuration
- GPT-5.5
- Measured
- 25 Sep 2026
- Not shown
- Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
1,470
Writing, literature and language · rank 11 of 408
- Unit
- Arena rating, higher is better
- Range
- 1,464 to 1,475
- Sample
- 17866 votes
- Configuration
- GPT-5.5 (high)
- Measured
- 25 Sep 2026
- Not shown
- Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
44.2%
Consistency (pass^5) · rank 1 of 2
- Unit
- % of tasks, higher is better
- Configuration
- gpt-5.5
- Measured
- 1 Jul 2026
- Not shown
- Not the model alone: it runs inside STATE-Bench's agent loop against simulated users, and versions of the benchmark are not comparable.
59.0%
Customer support task success (pass@1) · rank 1 of 2
- Unit
- % of tasks, higher is better
- Configuration
- gpt-5.5
- Measured
- 1 Jul 2026
- Not shown
- Not the model alone: it runs inside STATE-Bench's agent loop against simulated users, and versions of the benchmark are not comparable.
55.0%
Shopping assistant task success (pass@1) · rank 1 of 2
- Unit
- % of tasks, higher is better
- Configuration
- gpt-5.5
- Measured
- 1 Jul 2026
- Not shown
- Not the model alone: it runs inside STATE-Bench's agent loop against simulated users, and versions of the benchmark are not comparable.
58.9%
Task success (pass@1) · rank 1 of 2
- Unit
- % of tasks, higher is better
- Configuration
- gpt-5.5
- Measured
- 1 Jul 2026
- Not shown
- Not the model alone: it runs inside STATE-Bench's agent loop against simulated users, and versions of the benchmark are not comparable.
62.0%
Travel task success (pass@1) · rank 1 of 2
- Unit
- % of tasks, higher is better
- Configuration
- gpt-5.5
- Measured
- 1 Jul 2026
- Not shown
- Not the model alone: it runs inside STATE-Bench's agent loop against simulated users, and versions of the benchmark are not comparable.
3.62
User experience · rank 2 of 2
- Unit
- score from 1 to 5, higher is better
- Configuration
- gpt-5.5
- Measured
- 1 Jul 2026
- Not shown
- Not real customer satisfaction: the ratings come from a model-based judge.
65.9%
Average · rank 5 of 34
- Unit
- % resolved, higher is better
- Configuration
- gpt-5.5
- Measured
- 30 Jun 2026
- Not shown
- Not the model alone: the score is for the model inside the OpenHands agent, at the SDK version shown.
38.2%
Front end (SWE-Bench Multimodal) · rank 10 of 34
- Unit
- % resolved, higher is better
- Configuration
- gpt-5.5
- Measured
- 30 Jun 2026
- Not shown
- Not the model alone: the score is for the model inside the OpenHands agent, at the SDK version shown.
43.8%
Greenfield (Commit0) · rank 8 of 34
- Unit
- % resolved, higher is better
- Configuration
- gpt-5.5
- Measured
- 30 Jun 2026
- Not shown
- Not the model alone: the score is for the model inside the OpenHands agent, at the SDK version shown.
86.1%
Information gathering (GAIA) · rank 1 of 34
- Unit
- % resolved, higher is better
- Configuration
- gpt-5.5
- Measured
- 30 Jun 2026
- Not shown
- Not the model alone: the score is for the model inside the OpenHands agent, at the SDK version shown.
78.2%
Issue resolution (SWE-Bench) · rank 5 of 34
- Unit
- % resolved, higher is better
- Configuration
- gpt-5.5
- Measured
- 30 Jun 2026
- Not shown
- Not the model alone: the score is for the model inside the OpenHands agent, at the SDK version shown.
83.4%
Testing (SWT-Bench) · rank 3 of 34
- Unit
- % resolved, higher is better
- Configuration
- gpt-5.5
- Measured
- 30 Jun 2026
- Not shown
- Not the model alone: the score is for the model inside the OpenHands agent, at the SDK version shown.
29.9%
Consistency (pass^4) · rank 4 of 21
- Unit
- % of tasks, higher is better
- Configuration
- gpt-5.5
- Measured
- 22 Jul 2026
- Not shown
- Not your policies or customers: the tasks and policies are tau2's own, and the customer is simulated by gpt-5.2.
44.6%
Task success (pass^1) · rank 5 of 21
- Unit
- % of tasks, higher is better
- Configuration
- gpt-5.5
- Measured
- 22 Jul 2026
- Not shown
- Not your policies or customers: the tasks and policies are tau2's own, and the customer is simulated by gpt-5.2.
42.0%
Requested-length error · rank 295 of 370
- Unit
- % off the requested word count, lower is better
- Configuration
- GPT-5.5 (high)
- Measured
- 25 Apr 2026
- Not shown
- Not other format limits such as character counts or bullet counts.
96.0%
Requested-length error · rank 346 of 370
- Unit
- % off the requested word count, lower is better
- Configuration
- gpt-5.5
- Measured
- 25 Apr 2026
- Not shown
- Not other format limits such as character counts or bullet counts.
71.0%
Requested-length error · rank 330 of 370
- Unit
- % off the requested word count, lower is better
- Configuration
- gpt-5.5
- Measured
- 25 Apr 2026
- Not shown
- Not other format limits such as character counts or bullet counts.
98.0%
Requested-length error · rank 347 of 370
- Unit
- % off the requested word count, lower is better
- Configuration
- gpt-5.5
- Measured
- 25 Apr 2026
- Not shown
- Not other format limits such as character counts or bullet counts.
13.0%
Requested-length error · rank 103 of 370
- Unit
- % off the requested word count, lower is better
- Configuration
- gpt-5.5
- Measured
- 25 Apr 2026
- Not shown
- Not other format limits such as character counts or bullet counts.
0.41
Style adherence · rank 11 of 370
- Unit
- score from 0 to 1, higher is better
- Configuration
- GPT-5.5 (high)
- Measured
- 25 Apr 2026
- Not shown
- Not brand voice on your own examples: UGI's prompts are private and lean towards creative writing.
0.41
Style adherence · rank 17 of 370
- Unit
- score from 0 to 1, higher is better
- Configuration
- gpt-5.5
- Measured
- 25 Apr 2026
- Not shown
- Not brand voice on your own examples: UGI's prompts are private and lean towards creative writing.
0.41
Style adherence · rank 14 of 370
- Unit
- score from 0 to 1, higher is better
- Configuration
- gpt-5.5
- Measured
- 25 Apr 2026
- Not shown
- Not brand voice on your own examples: UGI's prompts are private and lean towards creative writing.
0.40
Style adherence · rank 33 of 370
- Unit
- score from 0 to 1, higher is better
- Configuration
- gpt-5.5
- Measured
- 25 Apr 2026
- Not shown
- Not brand voice on your own examples: UGI's prompts are private and lean towards creative writing.
0.40
Style adherence · rank 19 of 370
- Unit
- score from 0 to 1, higher is better
- Configuration
- gpt-5.5
- Measured
- 25 Apr 2026
- Not shown
- Not brand voice on your own examples: UGI's prompts are private and lean towards creative writing.
67.9
Writing score · rank 42 of 370
- Unit
- score out of 100, higher is better
- Configuration
- GPT-5.5 (high)
- Measured
- 25 Apr 2026
- Not shown
- Not business copy quality: UGI's prompts are private and lean towards creative writing, and models that often refuse get no score.
58.7
Writing score · rank 109 of 370
- Unit
- score out of 100, higher is better
- Configuration
- gpt-5.5
- Measured
- 25 Apr 2026
- Not shown
- Not business copy quality: UGI's prompts are private and lean towards creative writing, and models that often refuse get no score.
64.8
Writing score · rank 67 of 370
- Unit
- score out of 100, higher is better
- Configuration
- gpt-5.5
- Measured
- 25 Apr 2026
- Not shown
- Not business copy quality: UGI's prompts are private and lean towards creative writing, and models that often refuse get no score.
51.5
Writing score · rank 150 of 370
- Unit
- score out of 100, higher is better
- Configuration
- gpt-5.5
- Measured
- 25 Apr 2026
- Not shown
- Not business copy quality: UGI's prompts are private and lean towards creative writing, and models that often refuse get no score.
69.7
Writing score · rank 27 of 370
- Unit
- score out of 100, higher is better
- Configuration
- gpt-5.5
- Measured
- 25 Apr 2026
- Not shown
- Not business copy quality: UGI's prompts are private and lean towards creative writing, and models that often refuse get no score.
100.0%
Answer rate · rank 1 of 108
- Unit
- % of documents, higher is better
- Configuration
- GPT-5.5
- Measured
- 22 Sep 2026
- Not shown
- Not a quality score: a low rate usually means content filters were triggered, and hallucination rates are measured on answered documents only.
9.3%
Hallucination rate · rank 47 of 108
- Unit
- % of summaries, lower is better
- Configuration
- GPT-5.5
- Measured
- 22 Sep 2026
- Not shown
- Not errors in open questions or other tasks: only summarisation, judged by Vectara's own model (HHEM-2.3), not by people, on news-style documents rather than your data.
