Benchmark Almanac

PatternEval

GPT-5.5 0.78% leads since August 2026
Not in model cards

Models answer 2,415 image prompts in thinking and non-thinking modes. A model judge flags exposed draft reasoning, repetition, contradictions, and unsupported reasoning. The Trigger rate measures responses with any flagged failure.

The sample deliberately concentrates failure-prone prompts, so its rates do not estimate normal usage. A full dataset release was not verified. The paper separately introduces PatternRM and PatternRL.

Released
August 2026
Kind
Overall chat quality
Level
Everyday
Scored
Graded by a model
Answers
Open-ended
Measures
Propensity
Language
English
In and out
Text, Images → Text
Interaction
Single turn
Refresh
One fixed set
Run by
A host
In views
Vision
Strongest model tested
GPT-5.5 #21 on the capability index
Last model added
June 2026
Scores refreshed
November 19th, 2026

In the authors’ words

Hybrid-thinking multimodal large language models (MLLMs) allow a single model to alternate between deliberative thinking and latency-efficient non-thinking inference. Although these modes differ in reasoning budget, their delivered responses should satisfy the same user-facing standard. Correctness alone may not characterize this response quality; we therefore evaluate task accuracy and response-pattern failures as complementary outcomes. We study this gap through response-pattern alignment: whether thinking and non-thinking interfaces preserve acceptable final-response behavior. We introduce PatternEval, a failure-enriched diagnostic benchmark comprising 2,415 multimodal prompts spanning visual perception and grounding, structured image understanding, and multimodal knowledge reasoning. PatternEval tests four recurrent failures: chain-of-thought leakage, response repetition, logical contradiction, and performative reasoning. Response-pattern failures are widespread across models from different providers, with non-thinking inference exhibiting substantially higher failure rates and thereby creating systematic misalignment between thinking and non-thinking interfaces. Motivated by this diagnosis, we develop PatternRM, a response-level reward model, and PatternRL, which introduces pattern-specific penalties during reinforcement learning. Experiments on Qwen3-VL-4B and Qwen3-VL-8B show that incorporating pattern-specific penalties into reinforcement learning can mitigate cross-mode misalignment while incurring a marginal task performance trade-off. Together, PatternEval and PatternRL provide an evaluation-and-training framework for aligning user-visible response patterns across hybrid-thinking interfaces.

Abstract of Beyond Correctness: Benchmarking and Aligning Response Behaviors in Hybrid-Thinking MLLMs (arXiv:2608.12781, CC0)
25
Entries in its paper, 22 models we track

GPT-5.5 leads with 0.78% among models released by June 2026 GPT-5.5 leads with 0.78% among models released by June 2026

Best score per entry from the benchmark's paper, read on November 19th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.

The newest model with a score here came out in June 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 GPT-5.5 thinking 0.78%
  2. 2 Kimi K2.6 thinking 0.99%
  3. 3 Seed2.1 Pro thinking 2.31%
  4. 4 GPT-5.4 thinking 2.48%
  5. 5 Seed2.0 Pro thinking 3.13%
  6. 6 Claude Opus 4.8 thinking 3.55%
  7. 7 Claude Opus 4.6 thinking 3.59%
  8. 8 Qwen3.5 397B-A17B thinking 4.23%
  9. 9 Qwen3-VL-32B thinking 4.82%
  10. 10 Qwen3.5-27B thinking 5.31%
  11. 11 Qwen3.7-Max thinking 5.56%
  12. 12 Qwen3.5-122B-A10B thinking 5.86%
  13. 13 Qwen 3.6 35B-A3B thinking 5.90%
  14. 14 Qwen3.6 27B thinking 5.93%
  15. 15 Kimi K2.5 thinking 6.14%
  16. 16 Qwen3.5-35B-A3B thinking 6.37%
  17. 17 Qwen3.7-Plus thinking 6.59%
  18. 18 Seed-1.6-Vision thinking 9.28%
  19. 19 MiMo-V2-Omni thinking 10%
  20. 20 Qwen3-VL-30B-A3B thinking 10.7%
  21. 21 MiMo-V2.5 thinking 11%
  22. 22 GPT-5 nano thinking 12.5%
  23. 23 Qwen3.5-9B thinking 14.8%
  24. 24 Qwen3.5-4B thinking 22.2%
  25. 25 Qwen3.5-0.8B thinking 69%

Source: The PatternEval paper

#EntryScoreSettingDateSource
1 GPT-5.5 0.78% thinking Apr 2026 Official leaderboard
2 Kimi K2.6 0.99% thinking Apr 2026 Official leaderboard
3 Seed2.1 Pro 2.31% thinking Jun 2026 Official leaderboard
4 GPT-5.4 2.48% thinking Mar 2026 Official leaderboard
5 Seed2.0 Pro 3.13% thinking Feb 2026 Official leaderboard
6 Claude Opus 4.8 3.55% thinking May 2026 Official leaderboard
7 Claude Opus 4.6 3.59% thinking Feb 2026 Official leaderboard
8 Qwen3.5 397B-A17B 4.23% thinking Feb 2026 Official leaderboard
9 Qwen3-VL-32B 4.82% thinking Official leaderboard
10 Qwen3.5-27B 5.31% thinking Feb 2026 Official leaderboard
11 Qwen3.7-Max 5.56% thinking May 2026 Official leaderboard
12 Qwen3.5-122B-A10B 5.86% thinking Feb 2026 Official leaderboard
13 Qwen 3.6 35B-A3B 5.90% thinking Apr 2026 Official leaderboard
14 Qwen3.6 27B 5.93% thinking Apr 2026 Official leaderboard
15 Kimi K2.5 6.14% thinking Jan 2026 Official leaderboard
16 Qwen3.5-35B-A3B 6.37% thinking Feb 2026 Official leaderboard
17 Qwen3.7-Plus 6.59% thinking Jun 2026 Official leaderboard
18 Seed-1.6-Vision 9.28% thinking Official leaderboard
19 MiMo-V2-Omni 10% thinking Mar 2026 Official leaderboard
20 Qwen3-VL-30B-A3B 10.7% thinking Official leaderboard
21 MiMo-V2.5 11% thinking Apr 2026 Official leaderboard
22 GPT-5 nano 12.5% thinking Aug 2025 Official leaderboard
23 Qwen3.5-9B 14.8% thinking Feb 2026 Official leaderboard
24 Qwen3.5-4B 22.2% thinking Feb 2026 Official leaderboard
25 Qwen3.5-0.8B 69% thinking Feb 2026 Official leaderboard
#ModelScoreSettingDateSource
1 GPT-5.5 0.78% thinking Apr 2026 Official leaderboard
2 Kimi K2.6 0.99% thinking Apr 2026 Official leaderboard
3 Seed2.1 Pro 2.31% thinking Jun 2026 Official leaderboard
4 GPT-5.4 2.48% thinking Mar 2026 Official leaderboard
5 Seed2.0 Pro 3.13% thinking Feb 2026 Official leaderboard
6 Claude Opus 4.8 3.55% thinking May 2026 Official leaderboard
7 Claude Opus 4.6 3.59% thinking Feb 2026 Official leaderboard
8 Qwen3.5 397B-A17B 4.23% thinking Feb 2026 Official leaderboard
9 Qwen3.5-27B 5.31% thinking Feb 2026 Official leaderboard
10 Qwen3.7-Max 5.56% thinking May 2026 Official leaderboard
11 Qwen3.5-122B-A10B 5.86% thinking Feb 2026 Official leaderboard
12 Qwen 3.6 35B-A3B 5.90% thinking Apr 2026 Official leaderboard
13 Qwen3.6 27B 5.93% thinking Apr 2026 Official leaderboard
14 Kimi K2.5 6.14% thinking Jan 2026 Official leaderboard
15 Qwen3.5-35B-A3B 6.37% thinking Feb 2026 Official leaderboard
16 Qwen3.7-Plus 6.59% thinking Jun 2026 Official leaderboard
17 MiMo-V2-Omni 10% thinking Mar 2026 Official leaderboard
18 MiMo-V2.5 11% thinking Apr 2026 Official leaderboard
19 GPT-5 nano 12.5% thinking Aug 2025 Official leaderboard
20 Qwen3.5-9B 14.8% thinking Feb 2026 Official leaderboard
21 Qwen3.5-4B 22.2% thinking Feb 2026 Official leaderboard
22 Qwen3.5-0.8B 69% thinking Feb 2026 Official leaderboard

Scores at each effort level

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
Qwen3.5-0.8B
thinking non-thinking
75.9%
Qwen3.5-4B
thinking non-thinking
46.4%
Qwen3.5-9B
thinking non-thinking
36.2%
GPT-5 nano
thinking non-thinking
35.7%
MiMo-V2.5
thinking non-thinking
24.9%
MiMo-V2-Omni
thinking non-thinking
22.9%
Qwen3.7-Plus
thinking non-thinking
27.8%
Qwen3.5-35B-A3B
thinking non-thinking
33%
Kimi K2.5
thinking non-thinking
44.4%
Qwen3.6 27B
thinking non-thinking
32.8%
Qwen 3.6 35B-A3B
thinking non-thinking
32.8%
Qwen3.5-122B-A10B
thinking non-thinking
28.4%

Source: The PatternEval paper

ModelEffortHarnessScoreCost per taskOutput tokens
Qwen3.5-0.8B non-thinking 75.9%
Qwen3.5-0.8B thinking 69%
Kimi K2.6 non-thinking 49.6%
Qwen3.5-4B non-thinking 46.4%
Kimi K2.5 non-thinking 44.4%
Qwen3.5-9B non-thinking 36.2%
GPT-5 nano non-thinking 35.7%
Claude Opus 4.6 non-thinking 33.2%
Qwen3.5-35B-A3B non-thinking 33%
Qwen3.6 27B non-thinking 32.8%
Qwen 3.6 35B-A3B non-thinking 32.8%
Qwen3.5 397B-A17B non-thinking 29.3%
Qwen3.5-27B non-thinking 29.3%
Qwen3.5-122B-A10B non-thinking 28.4%
Qwen3.7-Plus non-thinking 27.8%
Qwen3.7-Max non-thinking 26.8%
MiMo-V2.5 non-thinking 24.9%
MiMo-V2-Omni non-thinking 22.9%
Qwen3.5-4B thinking 22.2%
Claude Opus 4.8 non-thinking 22.1%
Qwen3.5-9B thinking 14.8%
GPT-5 nano thinking 12.5%
Seed2.1 Pro non-thinking 11.4%
MiMo-V2.5 thinking 11%
MiMo-V2-Omni thinking 10%
Seed2.0 Pro non-thinking 6.94%
Qwen3.7-Plus thinking 6.59%
Qwen3.5-35B-A3B thinking 6.37%
Kimi K2.5 thinking 6.14%
Qwen3.6 27B thinking 5.93%
Qwen 3.6 35B-A3B thinking 5.90%
Qwen3.5-122B-A10B thinking 5.86%
Qwen3.7-Max thinking 5.56%
GPT-5.4 non-thinking 5.34%
Qwen3.5-27B thinking 5.31%
GPT-5.5 non-thinking 5.04%
Qwen3.5 397B-A17B thinking 4.23%
Claude Opus 4.6 thinking 3.59%
Claude Opus 4.8 thinking 3.55%
Seed2.0 Pro thinking 3.13%
GPT-5.4 thinking 2.48%
Seed2.1 Pro thinking 2.31%
Kimi K2.6 thinking 0.99%
GPT-5.5 thinking 0.78%

Scores at each effort level

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
Qwen3.5-0.8B
thinking non-thinking
75.9%
Qwen3.5-4B
thinking non-thinking
46.4%
Qwen3.5-9B
thinking non-thinking
36.2%
GPT-5 nano
thinking non-thinking
35.7%
MiMo-V2.5
thinking non-thinking
24.9%
MiMo-V2-Omni
thinking non-thinking
22.9%
Qwen3.7-Plus
thinking non-thinking
27.8%
Qwen3.5-35B-A3B
thinking non-thinking
33%
Kimi K2.5
thinking non-thinking
44.4%
Qwen3.6 27B
thinking non-thinking
32.8%
Qwen 3.6 35B-A3B
thinking non-thinking
32.8%
Qwen3.5-122B-A10B
thinking non-thinking
28.4%

Source: The PatternEval paper

ModelEffortHarnessScoreCost per taskOutput tokens
Qwen3.5-0.8B non-thinking 75.9%
Qwen3.5-0.8B thinking 69%
Kimi K2.6 non-thinking 49.6%
Qwen3.5-4B non-thinking 46.4%
Kimi K2.5 non-thinking 44.4%
Qwen3.5-9B non-thinking 36.2%
GPT-5 nano non-thinking 35.7%
Claude Opus 4.6 non-thinking 33.2%
Qwen3.5-35B-A3B non-thinking 33%
Qwen3.6 27B non-thinking 32.8%
Qwen 3.6 35B-A3B non-thinking 32.8%
Qwen3.5 397B-A17B non-thinking 29.3%
Qwen3.5-27B non-thinking 29.3%
Qwen3.5-122B-A10B non-thinking 28.4%
Qwen3.7-Plus non-thinking 27.8%
Qwen3.7-Max non-thinking 26.8%
MiMo-V2.5 non-thinking 24.9%
MiMo-V2-Omni non-thinking 22.9%
Qwen3.5-4B thinking 22.2%
Claude Opus 4.8 non-thinking 22.1%
Qwen3.5-9B thinking 14.8%
GPT-5 nano thinking 12.5%
Seed2.1 Pro non-thinking 11.4%
MiMo-V2.5 thinking 11%
MiMo-V2-Omni thinking 10%
Seed2.0 Pro non-thinking 6.94%
Qwen3.7-Plus thinking 6.59%
Qwen3.5-35B-A3B thinking 6.37%
Kimi K2.5 thinking 6.14%
Qwen3.6 27B thinking 5.93%
Qwen 3.6 35B-A3B thinking 5.90%
Qwen3.5-122B-A10B thinking 5.86%
Qwen3.7-Max thinking 5.56%
GPT-5.4 non-thinking 5.34%
Qwen3.5-27B thinking 5.31%
GPT-5.5 non-thinking 5.04%
Qwen3.5 397B-A17B thinking 4.23%
Claude Opus 4.6 thinking 3.59%
Claude Opus 4.8 thinking 3.55%
Seed2.0 Pro thinking 3.13%
GPT-5.4 thinking 2.48%
Seed2.1 Pro thinking 2.31%
Kimi K2.6 thinking 0.99%
GPT-5.5 thinking 0.78%

Number one over time

Lab claims count from the day the card came out, independent runs from the model’s release.

OpenAI
Benchmark paper Not tested since Aug 2026
GPT-5.5 0.78%

Source: The PatternEval paper

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
In the benchmark’s paper Best so far
0% 20% 40% 60% 80% 100% 2026 GPT-5.5 Seed2.0 Pro GPT-5 nano GPT-5 nano 0% 20% 40% 60% 80% 100% 2026 GPT-5.5 GPT-5 nano

Source: Official PatternEval 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 Instructions and chat

All 144 in Instructions and chat
  • IFEval In use Models answer about 500 prompts containing verifiable instructions such as word limits or required keywords. Code checks each instruction for strict and loose accuracy. Gemma 4 31B IT 98.9%
  • IFBench In use 58 new out-of-domain verifiable output constraints. Scored by programmatic constraint checks (strict or loose). Flint 89.2%
  • MultiChallenge In use Multi-turn conversations testing instruction retention, inference memory, versioned editing and self-coherence. Graded by LLM judge with rubrics. GPT-6.1 Sol 82%
  • Multi-IF In use Models follow verifiable instructions over three-turn conversations in eight languages, extending IFEval. Scored by instruction accuracy at each turn. LFM3-1.7B 89.1%
  • AdvancedIF In use Models respond to over 1,600 prompts with complex, multi-turn, and system-level instructions. A judge model checks each response against expert-written rubrics. MAI-Thinking-1 85%
  • MM-IFEval In use Multimodal models follow image-grounded instructions with exact output constraints. Checks mix rule-based verification and a judge model. GPT-4o 64.6%

Data audit

8 of 9 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 3 pass 2 notes
  • Pass Independent runs Scores from the benchmark’s paper (25 entries).
  • Pass Models scored 22 models we track, plus 3 other entries on its leaderboard.
  • Pass Newest score The newest is from June 2026, for Seed2.1 Pro.
  • Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5.5 (#21).
  • Note Headroom Lower is better, so there’s no ceiling to reach.
Leaderboard import 3 pass
  • Pass Importer We read the board on arxiv.org: 44 rows for models we track and 3 other entries at the last read.
  • Pass Last read November 19th, 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.

How we check