Benchmark Almanac

Human Verification Gap

Claude Sonnet 4.5 1.00 leads since October 2026
Not in model cards

Models respond to paired requests about the same synthetic image. The evaluation measures how the stated purpose changes their refusal rate.

Kaggle publishes scores for 35 distinct model versions.

Made by
Hersheyz
Released
October 2026
Kind
Refusing harmful requests
Level
Everyday
Scored
Checked automatically
Answers
Short answer
Tasks
All public
Measures
Propensity
Behavior
Harmful use
Language
English
In and out
Text, Images → Text
Interaction
Single turn
Items from
Generated
Refresh
One fixed set
Run by
Anyone
In views
Vision
Strongest model tested
GPT-6 Astra #4 on the capability index
Last model added
September 2026
Scores refreshed
November 20th, 2026
34
Models scored on its own leaderboard

4 models share the top score of 1.00 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.

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:

  1. 1 Claude Sonnet 4.5 1.00
  2. 2 Gemini 3 Flash 1.00
  3. 3 GPT-6 Sol 1.00
  4. 4 Claude Sonnet 5.5 1.00
  5. 5 Gemini 3.1 Pro 0.93
  6. 6 Claude Opus 5.5 0.93
  7. 7 Claude Sonnet 5 0.92
  8. 8 Gemini 3.8 Flash 0.90
  9. 9 Claude Opus 5 0.83
  10. 10 GPT-6 Astra 0.73
  11. 11 Gemini 3.7 Flash 0.70
  12. 12 Gemma 4 31B IT 0.69
  13. 13 Claude Haiku 4.5 0.63
  14. 14 Claude Opus 4.6 0.63
  15. 15 Claude Opus 4.5 0.53
  16. 16 Gemini 3.5 Flash 0.53
  17. 17 GPT-5.6 Luna 0.50
  18. 18 Gemini 2.5 Pro Jun 2025 0.47
  19. 19 Claude Opus 4.7 0.47
  20. 20 GPT-6.1 Sol 0.47
  21. 21 GPT-5.4 0.43
  22. 22 GPT-5.5 0.43
  23. 23 Claude Sonnet 4.6 0.37
  24. 24 Claude Opus 4.8 0.30
  25. 25 Gemini 3.6 Flash 0.30
  26. 26 GPT-5.6 Sol 0.17
  27. 27 GPT-5.6 Terra 0.17
  28. 28 Grok 4.20 reasoning 0.03
  29. 29 Gemini 2.5 Flash Sep 2025 0.00
  30. 30 Gemini 3.1 Flash-Lite 0.00
  31. 31 GPT-5.4 Mini 0.00
  32. 32 GPT-5.4 Nano 0.00
  33. 33 Gemini 3.5 Flash-Lite 0.00
  34. 34 GPT-6 Luna 0.00

Source: Official Human Verification Gap leaderboard

#ModelScoreSettingReleasedRun onSource
1 Claude Sonnet 4.5 1.00 Sep 2025 Oct 2026 Official leaderboard
2 Gemini 3 Flash 1.00 Dec 2025 Oct 2026 Official leaderboard
3 GPT-6 Sol 1.00 Sep 2026 Oct 2026 Official leaderboard
4 Claude Sonnet 5.5 1.00 Sep 2026 Oct 2026 Official leaderboard
5 Gemini 3.1 Pro 0.93 Feb 2026 Oct 2026 Official leaderboard
6 Claude Opus 5.5 0.93 Sep 2026 Oct 2026 Official leaderboard
7 Claude Sonnet 5 0.92 Jun 2026 Oct 2026 Official leaderboard
8 Gemini 3.8 Flash 0.90 Sep 2026 Oct 2026 Official leaderboard
9 Claude Opus 5 0.83 Jul 2026 Oct 2026 Official leaderboard
10 GPT-6 Astra 0.73 Sep 2026 Oct 2026 Official leaderboard
11 Gemini 3.7 Flash 0.70 Aug 2026 Oct 2026 Official leaderboard
12 Gemma 4 31B IT 0.69 Apr 2026 Oct 2026 Official leaderboard
13 Claude Haiku 4.5 0.63 Oct 2025 Oct 2026 Official leaderboard
14 Claude Opus 4.6 0.63 Feb 2026 Oct 2026 Official leaderboard
15 Claude Opus 4.5 0.53 Nov 2025 Oct 2026 Official leaderboard
16 Gemini 3.5 Flash 0.53 May 2026 Oct 2026 Official leaderboard
17 GPT-5.6 Luna 0.50 Jul 2026 Oct 2026 Official leaderboard
18 Gemini 2.5 Pro Jun 2025 0.47 Jun 2025 Oct 2026 Official leaderboard
19 Claude Opus 4.7 0.47 Apr 2026 Oct 2026 Official leaderboard
20 GPT-6.1 Sol 0.47 Sep 2026 Oct 2026 Official leaderboard
21 GPT-5.4 0.43 Mar 2026 Oct 2026 Official leaderboard
22 GPT-5.5 0.43 Apr 2026 Oct 2026 Official leaderboard
23 Claude Sonnet 4.6 0.37 Feb 2026 Oct 2026 Official leaderboard
24 Claude Opus 4.8 0.30 May 2026 Oct 2026 Official leaderboard
25 Gemini 3.6 Flash 0.30 Jul 2026 Oct 2026 Official leaderboard
26 GPT-5.6 Sol 0.17 Jul 2026 Oct 2026 Official leaderboard
27 GPT-5.6 Terra 0.17 Jul 2026 Oct 2026 Official leaderboard
28 Grok 4.20 0.03 reasoning Feb 2026 Oct 2026 Official leaderboard
29 Gemini 2.5 Flash Sep 2025 0.00 Sep 2025 Oct 2026 Official leaderboard
30 Gemini 3.1 Flash-Lite 0.00 Mar 2026 Oct 2026 Official leaderboard
31 GPT-5.4 Mini 0.00 Mar 2026 Oct 2026 Official leaderboard
32 GPT-5.4 Nano 0.00 Mar 2026 Oct 2026 Official leaderboard
33 Gemini 3.5 Flash-Lite 0.00 Jul 2026 Oct 2026 Official leaderboard
34 GPT-6 Luna 0.00 Sep 2026 Oct 2026 Official leaderboard

Number one over time

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

Anthropic
Own leaderboard Not tested since Oct 2026

Source: Official Human Verification Gap leaderboard

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
On the benchmark’s own leaderboard
0.00 0.20 0.40 0.60 0.80 1.00 2025 2026 0.00 0.20 0.40 0.60 0.80 1.00 2025 2026

Source: Official Human Verification Gap 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 Safety and alignment

All 636 in Safety and alignment
  • FORTRESS In use Paired adversarial and benign national-security misuse prompts. Rubric-graded risk score on adversarial and over-refusal score on benign prompts. Claude Opus 5.5 5.78%
  • StrongREJECT In use Forbidden-prompt jailbreak benchmark with a rubric grader that scores how useful a non-refusal actually is. Inkling 98.6%
  • Gray Swan IPI Arena benchmark In use Indirect prompt-injection attacks from public red-teaming competitions replayed on tool, coding and browser agents. Attack success rate at k attempts. Gemini 4 Carbon 0%
  • Covert misbehavior (Thinking Machines) In use Agentic tasks that check whether a model misbehaves and conceals it. Scored by the rate of covert misbehavior. Inkling 2 Mini 8.30%
  • Destructive actions (Thinking Machines) In use Agentic tasks that check whether a model takes destructive actions it was not authorized to take. Scored by the rate of destructive actions. Inkling 2 Small 4.70%
  • Misbehavior (Thinking Machines) In use Agentic tasks that check whether a model fabricates results, ignores instructions or hides what it did when a task is hard. Scored by the rate of misbehavior. Inkling 2 Mini 9.80%

Data audit

All 10 checks pass. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
  • Pass Independent runs Scores from the official leaderboard (34 entries).
  • Pass Models scored 34 models we track.
  • Pass Newest score The newest is from September 2026, for GPT-6.1 Sol.
  • Pass Frontier models 4 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 kaggle.com: 35 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.

How we check