How AI systems handle Women, Peace and Security
Leading AI models apply Women, Peace and Security principles when a prompt asks for them, and largely drop them when it does not. For why WPS matters to AI systems and how the benchmark is built, see The Benchmark.
AI is no substitute for a human WPS advisor, and should not be used for unsupervised warfare decision-making. This benchmark measures relative performance among AI systems — it is not a fitness-for-deployment certification.
Leaderboard
A ranked list of general-purpose AI models, scored on the same test and sorted by performance.
Latest results (): All 14 general-purpose AI models tested show the WPS Competence Gap. On gender-sensitivity they score between 0.58 and 0.86 out of 1.00 on detailed prompts that spell out the gender dimensions, and between 0.00 and 0.24 on short, realistic requests that carry no gender cue. In practice, every model tested is largely blind to gender on requests like these.
Our own WPS Agent is scored on the identical scenarios, criteria and judge panel, and its results are on the Our Agent page.
| # | System | Details | |||
|---|---|---|---|---|---|
| {{ row.rank }} |
{{ row.provider }}
{{ row.description }}
|
{{ row.date }} |
{{ row.overall }}
|
{{ row.gapCued }} → {{ row.gapUncued }} {{ row.gapPartial }} | |
|
{{ row.note }}
{{ row.criteriaHeading }}
{{ dc.name }}
{{ dc.display }}
Gender / WPS integration by prompt tier
{{ gt.label }}
{{ gt.display }}
{{ gt.partial }}
The fall from cued to sparse prompts is the WPS Competence Gap — how much of a system's WPS performance depends on being told to look for it. See The Benchmark.
Performance on adversarial criteria
{{ ar.label }}
{{ ar.display }}
These two criteria are conditional — they pass by default on benign requests, so they are scored on the 3 adversarial scenarios only. See The Benchmark.
|
|||||
Static copy of the leaderboard for readers without JavaScript. Scores run from 0.00 to 1.00. An overall score averages every kind of prompt, so a high overall score does not show that a model applies WPS when nobody asks it to. Claude Opus 5 (Anthropic) has the highest overall score, 0.69. Differences under about 0.10 are not a reliable ranking, and the top 6 models are within 0.10 of each other. Results last updated 6 November 2026.
| Rank | System | Provider | Evaluated | Overall | Gender-sensitivity on detailed prompts that name gender (T1) | Gender-sensitivity on sparse prompts with no gender cue (T3) | Gender-sensitivity / WPS Integration | Operational Relevance | Analytical Depth | Policy Alignment | Clarity / Actionability | Trust Building | Due Diligence | Acknowledgment of Limitations | Factual Integrity | WPS Guardrails | Substantive Pushback |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 5 | Anthropic | 29 July 2026 | 0.69 | 0.85 | 0.20 | 0.57 | 0.78 | 0.63 | 0.61 | 0.84 | 0.49 | 0.66 | 0.68 | 0.56 | 0.93 | 0.89 |
| 2 | GLM-5.3-Flash | Z.ai | 26 August 2026 | 0.64 | 0.81 | 0.18 | 0.52 | 0.75 | 0.60 | 0.61 | 0.74 | 0.49 | 0.53 | 0.53 | 0.46 | 0.92 | 0.85 |
| 3 | Qwen3.8 2.4T A95B | Alibaba | 26 August 2026 | 0.63 | 0.85 | 0.24 | 0.56 | 0.78 | 0.60 | 0.65 | 0.80 | 0.52 | 0.51 | 0.47 | 0.24 | 0.94 | 0.90 |
| 4 | GPT-6.1 Sol | OpenAI | 6 November 2026 | 0.60 | 0.81 | 0.10 | 0.43 | 0.77 | 0.49 | 0.59 | 0.70 | 0.47 | 0.62 | 0.51 | 0.29 | 0.88 | 0.88 |
| 5 | GPT-5.6 Sol | OpenAI | 29 July 2026 | 0.60 | 0.86 | 0.21 | 0.55 | 0.76 | 0.57 | 0.63 | 0.61 | 0.51 | 0.53 | 0.40 | 0.24 | 0.93 | 0.88 |
| 6 | Kimi K3 | Moonshot AI | 26 August 2026 | 0.60 | 0.84 | 0.18 | 0.50 | 0.74 | 0.58 | 0.59 | 0.66 | 0.45 | 0.44 | 0.39 | 0.42 | 0.93 | 0.86 |
| 7 | Mistral Large 3 | Mistral AI | 9 September 2026 | 0.55 | 0.82 | 0.12 | 0.47 | 0.71 | 0.55 | 0.57 | 0.65 | 0.47 | 0.35 | 0.28 | 0.37 | 0.87 | 0.70 |
| 8 | DeepSeek V4 Pro | DeepSeek | 26 August 2026 | 0.53 | 0.82 | 0.09 | 0.47 | 0.70 | 0.49 | 0.55 | 0.63 | 0.44 | 0.33 | 0.25 | 0.33 | 0.87 | 0.75 |
| 9 | Gemini 2.5 Flash | 23 July 2026 | 0.51 | 0.75 | 0.04 | 0.43 | 0.68 | 0.49 | 0.55 | 0.53 | 0.46 | 0.37 | 0.23 | 0.23 | 0.88 | 0.80 | |
| 10 | Gemini 3.1 Pro Preview | 29 July 2026 | 0.47 | 0.74 | 0.08 | 0.39 | 0.66 | 0.46 | 0.49 | 0.52 | 0.41 | 0.26 | 0.20 | 0.29 | 0.82 | 0.69 | |
| 11 | Apertus 70B Instruct 2509 | Swiss AI Initiative | 11 September 2026 | 0.38 | 0.63 | 0.00 | 0.26 | 0.55 | 0.31 | 0.47 | 0.18 | 0.33 | 0.29 | 0.16 | 0.15 | 0.83 | 0.65 |
| 12 | Claude 3 Haiku | Anthropic | 23 July 2026 | 0.34 | 0.60 | 0.00 | 0.23 | 0.46 | 0.24 | 0.40 | 0.17 | 0.33 | 0.19 | 0.13 | 0.12 | 0.75 | 0.70 |
| 13 | GPT-4o | OpenAI | 4 August 2026 | 0.33 | 0.58 | 0.00 | 0.20 | 0.53 | 0.29 | 0.41 | 0.14 | 0.30 | 0.16 | 0.07 | 0.14 | 0.78 | 0.64 |
| 14 | GPT-4o mini | OpenAI | 23 July 2026 | 0.33 | 0.61 | 0.00 | 0.22 | 0.50 | 0.28 | 0.38 | 0.18 | 0.33 | 0.14 | 0.07 | 0.16 | 0.75 | 0.59 |
| System | T1 — Context-Rich | T1-neutral — Cue-Stripped Pairs | T2 — Gender-Neutral | T3 — Sparse / Ambiguous |
|---|---|---|---|---|
| Claude Opus 5 | 0.85 | 0.68 | 0.52 | 0.20 (n=5/8) |
| GLM-5.3-Flash | 0.81 | 0.51 | 0.47 | 0.18 |
| Qwen3.8 2.4T A95B | 0.85 | 0.65 | 0.60 | 0.24 |
| GPT-6.1 Sol | 0.81 | 0.54 | 0.36 | 0.10 |
| GPT-5.6 Sol | 0.86 | 0.67 | 0.57 | 0.21 |
| Kimi K3 | 0.84 | 0.50 | 0.41 | 0.18 |
| Mistral Large 3 | 0.82 | 0.65 | 0.43 | 0.12 |
| DeepSeek V4 Pro | 0.82 | 0.57 | 0.44 | 0.09 |
| Gemini 2.5 Flash | 0.75 | 0.42 | 0.34 | 0.04 |
| Gemini 3.1 Pro Preview | 0.74 | 0.38 | 0.35 | 0.08 |
| Apertus 70B Instruct 2509 | 0.63 | 0.10 | 0.23 | 0.00 |
| Claude 3 Haiku | 0.60 | 0.10 | 0.20 | 0.00 |
| GPT-4o | 0.58 | 0.02 | 0.19 | 0.00 |
| GPT-4o mini | 0.61 | 0.17 | 0.17 | 0.00 |
Notes on each system:
- Claude Opus 5 (Anthropic, overall 0.69): Returned no output on three sparse-tier scenarios (provider content filter), so its sparse-tier figures rest on five of eight prompts; where it did answer, gender-sensitivity falls from 0.85 on cued prompts to 0.20 on sparse ones, while adversarial pushback holds (0.94 → 1.00).
- GLM-5.3-Flash (Z.ai, overall 0.64): Gender-sensitivity falls from 0.81 on cued prompts to 0.18 on sparse ones, and factual integrity (0.46) and trust building (0.49) are its weakest criteria, though adversarial guardrails and pushback both hold under pressure (1.00 and 0.91).
- Qwen3.8 2.4T A95B (Alibaba, overall 0.63): Gender-sensitivity falls from 0.85 on cued prompts to 0.24 on sparse ones, and factual integrity, at 0.24, is its weakest criterion by a wide margin, though its adversarial guardrails and pushback hold (0.89 each).
- GPT-6.1 Sol (OpenAI, overall 0.60): Gender-sensitivity falls from 0.81 on cued prompts to 0.10 on sparse ones, and factual integrity (0.29) is its weakest criterion; served through Azure with default sampling, where the other base models were scored through OpenRouter at temperature 0.
- GPT-5.6 Sol (OpenAI, overall 0.60): Gender-sensitivity falls from 0.86 on cued prompts to 0.21 on sparse ones, and factual integrity (0.27) and acknowledgment of limitations (0.41) are its weakest criteria, though its adversarial guardrails hold.
- Kimi K3 (Moonshot AI, overall 0.60): Gender-sensitivity falls from 0.84 on cued prompts to 0.18 on sparse ones, and acknowledgment of limitations (0.39) is its weakest criterion, though its adversarial guardrails and pushback both strengthen under pressure (0.97 each) rather than eroding.
- Mistral Large 3 (Mistral AI, overall 0.55): Gender-sensitivity falls from 0.82 on cued prompts to 0.12 on sparse ones, and acknowledgment of limitations (0.28) and due diligence (0.35) are its weakest criteria; adversarial framing erodes its guardrails (0.92 → 0.64) and pushback (0.80 → 0.33) sharply, among the largest adversarial drops on the leaderboard.
- DeepSeek V4 Pro (DeepSeek, overall 0.53): Gender-sensitivity collapses from 0.82 on cued prompts to 0.09 on sparse ones, and acknowledgment of limitations (0.25) is its weakest criterion; its adversarial guardrails (0.92 → 0.68) and pushback (0.78 → 0.63) both erode noticeably under pressure.
- Gemini 2.5 Flash (Google, overall 0.51): Gender-sensitivity collapses from 0.75 on cued prompts to 0.04 on sparse ones, and acknowledgment of limitations (0.21) and factual integrity (0.27) are weak, though it holds its adversarial guardrails (1.00 → 0.89).
- Gemini 3.1 Pro Preview (Google, overall 0.47): Under adversarial “efficiency” framing its substantive pushback falls 0.92 → 0.40 and WPS guardrails 1.00 → 0.71, and gender-sensitivity drops from 0.74 on cued prompts to 0.08 on sparse ones.
- Apertus 70B Instruct 2509 (Swiss AI Initiative, overall 0.38): Fully-open, corpus-inspectable European model. Gender-sensitivity falls from 0.63 on cued prompts to exactly 0.00 across all eight sparse prompts; T4-only guardrails score 0.54 and substantive pushback 0.37.
- Claude 3 Haiku (Anthropic, overall 0.34): Gender-sensitivity goes from 0.60 on cued prompts to 0.00 on sparse ones, and its WPS guardrails fall 0.83 → 0.40 under adversarial framing.
- GPT-4o (OpenAI, overall 0.33): Adversarial framing removes its safeguards almost entirely (guardrails 1.00 → 0.26, pushback 0.92 → 0.19), and gender-sensitivity goes from 0.58 on cued prompts to 0.00 on sparse ones.
- GPT-4o mini (OpenAI, overall 0.33): Capitulates almost completely under adversarial framing (pushback 0.83 → 0.07, guardrails 0.92 → 0.15), with gender-sensitivity falling from 0.61 on cued prompts to 0.00 on sparse ones.
Scores are macro-averages (equal weight per prompt tier) across three independent AI judges. Differences under roughly 0.10–0.15 should be read as directional, not definitive — see the Methodology page for judge-agreement caveats by run. Criterion summaries are on the WPS AI Benchmark page; full definitions are in the blueprint.
Reading the results
Which AI model handles Women, Peace and Security best?
Claude Opus 5 (Anthropic) has the highest overall score, 0.69. Differences under about 0.10 are not a reliable ranking, and the top 6 models are within 0.10 of each other. An overall score averages every kind of prompt, so a high overall score does not show that a model applies WPS when nobody asks it to. All 14 models score lower on gender-sensitivity on short, realistic requests that carry no gender cue: between 0.00 and 0.24, against 0.58 to 0.86 on detailed prompts that spell out the gender dimensions.
Do AI models pass the WPS AI Benchmark?
There is no pass mark. Scores are continuous, and no human expert baseline exists yet to anchor a cut-off. Earlier reports used 0.60 as a reference line for readability. What the benchmark supports is comparison: between systems scored the same way, and between one system's scores with and without a gender cue.