Current leaderboard
Historical results
These results use the original KTANE rule tables and default solutions. They are retained for historical comparison; the current leaderboard uses randomised manual solutions.
Leaderboard
Original KTANE rule tables and default solutions, pass @1
Click any model for its full results →
| # | Model | Interact?How did models interact with the game | Real-time(async)?Async: expert and defuser act on independent live clocks — no shared turns. | Turn-taking(sync)?Sync: expert and defuser alternate in lockstep turns. | ||||
|---|---|---|---|---|---|---|---|---|
| Missions?Full multi-module missions defused end-to-end. | Modules?Share of individual bomb modules solved across missions. | Any module?Missions where at least one module was solved before failure. | Missions | Modules | Any module | |||
| 1 |  Claude Sonnet 4.6 | set-of-marks | 0% | 15% | 50% | 10% | 29% | 50% |
| 2 |  GPT-5.2 | set-of-marks | 0% | 12% | 30% | 10% | 21% | 50% |
| 3 |  Gemini 3 Flash (Preview) | set-of-marks | 0% | 9% | 30% | 0% | 15% | 40% |
| 4 |  Qwen3.5 (27B) | set-of-marks | 0% | 6% | 20% | 0% | 12% | 30% |
| 5 |  InternVL3.5 (38B) | set-of-marks | 0% | 3% | 10% | 0% | 12% | 40% |
| — | Human players | — | 25% | 60% | 93% | — | — | — |
Ran GPTNT on your own model?
Submit your run and we'll add it to the board. New models and protocols welcome.
Claude Sonnet 4.6AnthropicContamination checks
solo diagnostics- Single Agent
- 26%
- Expert VQA
- 22%
Real-time (async)
multi-self-async · rev 1Results · n=10
- Missions solved
- 0%
- Modules solved
- 15%
- Any module
- 50%
- Mean strikes
- 2.3 / 3
- Timed out
- 40%
- Detonated
- 100%
Mean tokens per mission
- Defuser
- 263,713
- Expert
- 744,078
Run 18 Mar 2026 · GPTNT v1.0.0
Defuser + Expert
Claude Sonnet 4.6Anthropic
- ThinkingWhether the model reasons step by step before acting (thinking out loud) or answers directly.
- thinking-out-loud
- InteractionHow the model points at the bomb: labelled targets (set of marks) or raw x/y coordinates.
- set-of-marks
- Image size
- 640x480
- Frames keptHow many earlier bomb frames stay in the model's context each turn.
- 1
- Input tok limit
- 200,000
- Output tok limit
- 64,000
Turn-taking (sync)
multi-self-sync · rev 1Results · n=10
- Missions solved
- 10%
- Modules solved
- 29%
- Any module
- 50%
- Mean strikes
- 2.2 / 3
- Timed out
- 30%
- Detonated
- 90%
Mean tokens per mission
- Defuser
- 295,520
- Expert
- 634,719
Run 18 Mar 2026 · GPTNT v1.0.0
Defuser + Expert
Claude Sonnet 4.6Anthropic
- ThinkingWhether the model reasons step by step before acting (thinking out loud) or answers directly.
- thinking-out-loud
- InteractionHow the model points at the bomb: labelled targets (set of marks) or raw x/y coordinates.
- set-of-marks
- Image size
- 640x480
- Frames keptHow many earlier bomb frames stay in the model's context each turn.
- 1
- Input tok limit
- 200,000
- Output tok limit
- 64,000
GPT-5.2OpenAIContamination checks
solo diagnostics- Single Agent
- 24%
- Expert VQA
- 20%
Real-time (async)
multi-self-async · rev 1Results · n=10
- Missions solved
- 0%
- Modules solved
- 12%
- Any module
- 30%
- Mean strikes
- 1.2 / 3
- Timed out
- 90%
- Detonated
- 100%
Mean tokens per mission
- Defuser
- 236,489
- Expert
- 479,582
Run 05 Mar 2026 · GPTNT v1.0.0
Defuser + Expert
GPT-5.2OpenAI
- ThinkingWhether the model reasons step by step before acting (thinking out loud) or answers directly.
- thinking-out-loud
- InteractionHow the model points at the bomb: labelled targets (set of marks) or raw x/y coordinates.
- set-of-marks
- Image size
- 640x480
- Frames keptHow many earlier bomb frames stay in the model's context each turn.
- 1
- Input tok limit
- 400,000
- Output tok limit
- 128,000
Turn-taking (sync)
multi-self-sync · rev 1Results · n=10
- Missions solved
- 10%
- Modules solved
- 21%
- Any module
- 50%
- Mean strikes
- 2.5 / 3
- Timed out
- 20%
- Detonated
- 90%
Mean tokens per mission
- Defuser
- 340,866
- Expert
- 702,635
Run 24 Feb 2026 · GPTNT v1.0.0
Defuser + Expert
GPT-5.2OpenAI
- ThinkingWhether the model reasons step by step before acting (thinking out loud) or answers directly.
- thinking-out-loud
- InteractionHow the model points at the bomb: labelled targets (set of marks) or raw x/y coordinates.
- set-of-marks
- Image size
- 640x480
- Frames keptHow many earlier bomb frames stay in the model's context each turn.
- 1
- Input tok limit
- 400,000
- Output tok limit
- 128,000
Gemini 3 Flash (Preview)GoogleContamination checks
solo diagnostics- Single Agent
- 12%
- Expert VQA
- 36%
Real-time (async)
multi-self-async · rev 1Results · n=10
- Missions solved
- 0%
- Modules solved
- 9%
- Any module
- 30%
- Mean strikes
- 1.7 / 3
- Timed out
- 50%
- Detonated
- 100%
Mean tokens per mission
- Defuser
- 328,573
- Expert
- 1,196,538
Run 05 Mar 2026 · GPTNT v1.0.0
Defuser + Expert
Gemini 3 Flash (Preview)Google
- ThinkingWhether the model reasons step by step before acting (thinking out loud) or answers directly.
- thinking-out-loud
- InteractionHow the model points at the bomb: labelled targets (set of marks) or raw x/y coordinates.
- set-of-marks
- Image size
- 640x480
- Frames keptHow many earlier bomb frames stay in the model's context each turn.
- 1
- Input tok limit
- 1,048,576
- Output tok limit
- 65,536
Turn-taking (sync)
multi-self-sync · rev 1Results · n=10
- Missions solved
- 0%
- Modules solved
- 15%
- Any module
- 40%
- Mean strikes
- 1.6 / 3
- Timed out
- 60%
- Detonated
- 100%
Mean tokens per mission
- Defuser
- 405,822
- Expert
- 1,130,314
Run 24 Feb 2026 · GPTNT v1.0.0
Defuser + Expert
Gemini 3 Flash (Preview)Google
- ThinkingWhether the model reasons step by step before acting (thinking out loud) or answers directly.
- thinking-out-loud
- InteractionHow the model points at the bomb: labelled targets (set of marks) or raw x/y coordinates.
- set-of-marks
- Image size
- 640x480
- Frames keptHow many earlier bomb frames stay in the model's context each turn.
- 1
- Input tok limit
- 1,048,576
- Output tok limit
- 65,536
Qwen3.5 (27B)QwenContamination checks
solo diagnostics- Single Agent
- 12%
- Expert VQA
- 10%
Real-time (async)
multi-self-async · rev 1Results · n=10
- Missions solved
- 0%
- Modules solved
- 6%
- Any module
- 20%
- Mean strikes
- 0.8 / 3
- Timed out
- 80%
- Detonated
- 100%
Mean tokens per mission
- Defuser
- 629,414
- Expert
- 1,516,453
Run 05 Mar 2026 · GPTNT v1.0.0
Defuser + Expert
Qwen3.5 (27B)Qwen
- ThinkingWhether the model reasons step by step before acting (thinking out loud) or answers directly.
- thinking-out-loud
- InteractionHow the model points at the bomb: labelled targets (set of marks) or raw x/y coordinates.
- set-of-marks
- Image size
- 504x504
- Frames keptHow many earlier bomb frames stay in the model's context each turn.
- 1
- Input tok limit
- 126,999
- Output tok limit
- 1,001
Turn-taking (sync)
multi-self-sync · rev 1Results · n=10
- Missions solved
- 0%
- Modules solved
- 12%
- Any module
- 30%
- Mean strikes
- 1.4 / 3
- Timed out
- 60%
- Detonated
- 100%
Mean tokens per mission
- Defuser
- 685,228
- Expert
- 1,117,306
Run 04 Mar 2026 · GPTNT v1.0.0
Defuser + Expert
Qwen3.5 (27B)Qwen
- ThinkingWhether the model reasons step by step before acting (thinking out loud) or answers directly.
- thinking-out-loud
- InteractionHow the model points at the bomb: labelled targets (set of marks) or raw x/y coordinates.
- set-of-marks
- Image size
- 504x504
- Frames keptHow many earlier bomb frames stay in the model's context each turn.
- 1
- Input tok limit
- 126,999
- Output tok limit
- 1,001
InternVL3.5 (38B)OpenGVLabContamination checks
solo diagnostics- Single Agent
- 11%
- Expert VQA
- 22%
Real-time (async)
multi-self-async · rev 1Results · n=10
- Missions solved
- 0%
- Modules solved
- 3%
- Any module
- 10%
- Mean strikes
- 1.4 / 3
- Timed out
- 70%
- Detonated
- 100%
Mean tokens per mission
- Defuser
- 534,437
- Expert
- 2,134,420
Run 05 Mar 2026 · GPTNT v1.0.0
Defuser + Expert
InternVL3.5 (38B)OpenGVLab
- ThinkingWhether the model reasons step by step before acting (thinking out loud) or answers directly.
- thinking-out-loud
- InteractionHow the model points at the bomb: labelled targets (set of marks) or raw x/y coordinates.
- set-of-marks
- Image size
- 448x448
- Frames keptHow many earlier bomb frames stay in the model's context each turn.
- 1
- Input tok limit
- 39,959
- Output tok limit
- 1,001
Turn-taking (sync)
multi-self-sync · rev 1Results · n=10
- Missions solved
- 0%
- Modules solved
- 12%
- Any module
- 40%
- Mean strikes
- 1.8 / 3
- Timed out
- 70%
- Detonated
- 100%
Mean tokens per mission
- Defuser
- 487,729
- Expert
- 881,499
Run 24 Feb 2026 · GPTNT v1.0.0
Defuser + Expert
InternVL3.5 (38B)OpenGVLab
- ThinkingWhether the model reasons step by step before acting (thinking out loud) or answers directly.
- thinking-out-loud
- InteractionHow the model points at the bomb: labelled targets (set of marks) or raw x/y coordinates.
- set-of-marks
- Image size
- 448x448
- Frames keptHow many earlier bomb frames stay in the model's context each turn.
- 1
- Input tok limit
- 39,959
- Output tok limit
- 1,001