Files
html/api/ai-benchmark-cache.json
opus f699ed0200
Some checks failed
WEVAL NonReg / nonreg (push) Has been cancelled
auto-sync-0005
2026-04-18 00:05:02 +02:00

902 lines
21 KiB
JSON

{
"all_ais": {
"Claude_Opus_4.6": {
"type": "sovereign",
"avg": 90,
"icon": "🏆",
"caps": {
"combined": 90,
"infra": 90,
"ecosystem": 90
},
"used_in": "Reference (claude.ai)",
"wired": true,
"cost": "Souverain 0EUR",
"usecase": "Stratégie complexe · Analyse code · Orchestration avancée",
"url": "https://claude.ai"
},
"GPT_4o": {
"type": "sovereign",
"avg": 87,
"icon": "🏆",
"caps": {
"combined": 87,
"infra": 86,
"ecosystem": 88
},
"used_in": "Reference",
"wired": true,
"cost": "Souverain 0EUR",
"usecase": "Chat général · Vision · Multimodal",
"url": "https://chat.openai.com"
},
"Gemini_2.5_Pro": {
"type": "sovereign",
"avg": 86,
"icon": "🏆",
"caps": {
"combined": 86,
"infra": 85,
"ecosystem": 87
},
"used_in": "Reference",
"wired": true,
"cost": "Souverain 0EUR",
"usecase": "Long context · Multimodal · Code",
"url": "https://gemini.google.com"
},
"ChatGPT_4o": {
"type": "sovereign",
"avg": 85,
"icon": "🏆",
"caps": {
"combined": 85,
"infra": 86,
"ecosystem": 84
},
"used_in": "Reference",
"wired": true,
"cost": "Souverain 0EUR",
"usecase": "Chat consumer · Plugins · Vision",
"url": "https://chat.openai.com"
},
"DeepSeek_V3": {
"type": "sovereign",
"avg": 84,
"icon": "🏆",
"caps": {
"combined": 84,
"infra": 83,
"ecosystem": 85
},
"used_in": "Reference",
"wired": true,
"cost": "Souverain 0EUR",
"usecase": "Reasoning · Math · Code open-source",
"url": "https://chat.deepseek.com"
},
"Gemini_Pro": {
"type": "sovereign",
"avg": 83,
"icon": "🏆",
"caps": {
"combined": 83,
"infra": 84,
"ecosystem": 82
},
"used_in": "Reference",
"wired": true,
"cost": "Souverain 0EUR",
"usecase": "Search · Multimodal · Enterprise",
"url": "https://gemini.google.com"
},
"WEVIA_MASTER": {
"type": "agent",
"avg": 88,
"icon": "🧠",
"caps": {
"combined": 88,
"infra": 90,
"ecosystem": 85
},
"used_in": "Orchestrateur 22/22 exec · Intercept v9",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Pilotage complet · Exec auto-fix · Agents · Actions",
"url": "/wevia-master.html"
},
"CONSENSUS": {
"type": "agent",
"avg": 84,
"icon": "⚖️",
"caps": {
"combined": 84,
"infra": 82,
"ecosystem": 86
},
"used_in": "4 providers MoA · Score 9.8/10",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Décision multi-IA · Validation croisée · Arbitrage",
"url": "/wevia-master.html"
},
"WEVIA_PUBLIC": {
"type": "chatbot",
"avg": 82,
"icon": "💬",
"caps": {
"combined": 82,
"infra": 80,
"ecosystem": 84
},
"used_in": "Widget chat public · /wevia",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Chat visiteurs · FAQ · Lead generation",
"url": "/wevia"
},
"WEVIA_WIDGET": {
"type": "chatbot",
"avg": 80,
"icon": "🔮",
"caps": {
"combined": 80,
"infra": 78,
"ecosystem": 82
},
"used_in": "Widget embed iframe · wevia-widget.html",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Embed site client · Chat inline · Support",
"url": "/wevia-widget.html"
},
"WEVIA_CONSOLE": {
"type": "backend",
"avg": 81,
"icon": "🖥️",
"caps": {
"combined": 81,
"infra": 85,
"ecosystem": 77
},
"used_in": "Centre Commande · 71 modules",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Administration · Monitoring · Configuration",
"url": "/command-center.html"
},
"WEVCODE": {
"type": "backend",
"avg": 82,
"icon": "⚙️",
"caps": {
"combined": 82,
"infra": 84,
"ecosystem": 80
},
"used_in": "Code assistant · 4 modes fast/deep/code/math",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Coding souverain · Debug · Refactor · Tests",
"url": "/wevcode"
},
"WEDROID": {
"type": "backend",
"avg": 79,
"icon": "🔧",
"caps": {
"combined": 79,
"infra": 85,
"ecosystem": 73
},
"used_in": "Diagnostic backend v5 · 12 providers",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Diag serveur · DB fix · Deploy · Service restart",
"url": "/wevia-ia/droid.html"
},
"CLAW_CODE": {
"type": "backend",
"avg": 78,
"icon": "🐙",
"caps": {
"combined": 78,
"infra": 82,
"ecosystem": 74
},
"used_in": "Code review + generation · /claw-code",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Code review · PR analysis · Refactoring",
"url": "/claw-code/"
},
"CORTEX": {
"type": "agent",
"avg": 77,
"icon": "🌐",
"caps": {
"combined": 77,
"infra": 80,
"ecosystem": 74
},
"used_in": "Visualisation architecture · SVG live",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Architecture viz · Topologie · Status global",
"url": "/wevia-cortex.html"
},
"DEERFLOW": {
"type": "search",
"avg": 79,
"icon": "🦌",
"caps": {
"combined": 79,
"infra": 76,
"ecosystem": 82
},
"used_in": "Deep research · 42 skills · Memory v1",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Recherche approfondie · Veille · Synthèse multi-source",
"url": "https://deerflow.weval-consulting.com"
},
"PAPERCLIP": {
"type": "agent",
"avg": 76,
"icon": "📎",
"caps": {
"combined": 76,
"infra": 74,
"ecosystem": 78
},
"used_in": "Project mgmt · 3 agents · Goals/Issues",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Gestion projet · Sprint · Issues · Roadmap",
"url": "https://paperclip.weval-consulting.com"
},
"WEPREDICT": {
"type": "search",
"avg": 75,
"icon": "🐟",
"caps": {
"combined": 75,
"infra": 72,
"ecosystem": 78
},
"used_in": "MiroFish rebrand · Port 5001",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Prédiction · Analytics · Forecast",
"url": "https://wepredict.weval-consulting.com"
},
"OPENCLAW": {
"type": "sovereign",
"avg": 74,
"icon": "🐙",
"caps": {
"combined": 74,
"infra": 78,
"ecosystem": 70
},
"used_in": "Ollama 4 models · Port 11434",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Chat souverain local · Zero cloud · Privacy",
"url": "/openclaw.html"
},
"L99_BRAIN": {
"type": "testing",
"avg": 80,
"icon": "🧪",
"caps": {
"combined": 80,
"infra": 85,
"ecosystem": 75
},
"used_in": "L99 pipeline · Puppeteer · 29/29",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Test auto · NonReg · Playwright · Screenshots",
"url": "/l99-brain.html"
},
"DIRECTOR": {
"type": "agent",
"avg": 76,
"icon": "👁️",
"caps": {
"combined": 76,
"infra": 82,
"ecosystem": 70
},
"used_in": "Supervision */15min · 39 obs",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Auto-supervision · Détection issues · Auto-fix",
"url": "/director.html"
},
"NUCLEI": {
"type": "security",
"avg": 73,
"icon": "🔬",
"caps": {
"combined": 73,
"infra": 78,
"ecosystem": 68
},
"used_in": "Vuln scanner installed",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Pentest auto · OWASP · CVE scan",
"url": "/security-dashboard.html"
},
"GUARDIAN": {
"type": "security",
"avg": 74,
"icon": "🛡️",
"caps": {
"combined": 74,
"infra": 80,
"ecosystem": 68
},
"used_in": "SSO watch */5min · chattr +i",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Protection fichiers · Auth monitor · Firewall",
"url": "/admin.html"
},
"BLADE_IA": {
"type": "agent",
"avg": 76,
"icon": "⚡",
"caps": {
"combined": 76,
"infra": 80,
"ecosystem": 72
},
"used_in": "Desktop agent · 34 caps · PowerShell",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Desktop automation · Screenshots · Git sync",
"url": "/blade-ai.html"
},
"UX_AGENT": {
"type": "testing",
"avg": 71,
"icon": "🎨",
"caps": {
"combined": 71,
"infra": 74,
"ecosystem": 68
},
"used_in": "Visual QA · 10 pages · Cron */2h",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Détection UX bugs · undefined · Layout check",
"url": "/l99-brain.html"
},
"ETHICA_CHAT": {
"type": "chatbot",
"avg": 77,
"icon": "💊",
"caps": {
"combined": 77,
"infra": 74,
"ecosystem": 80
},
"used_in": "Pharma HCP · 126K contacts",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Outreach pharma · HCP matching · Enrichment",
"url": "/ethica-hcp-manager.html"
},
"HERMES": {
"type": "agent",
"avg": 75,
"icon": "📨",
"caps": {
"combined": 75,
"infra": 73,
"ecosystem": 77
},
"used_in": "Messaging · Chat routing",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Message routing · Multi-channel · Notifications",
"url": "/wevia-master.html"
},
"KILO": {
"type": "agent",
"avg": 73,
"icon": "📊",
"caps": {
"combined": 73,
"infra": 75,
"ecosystem": 71
},
"used_in": "KPI tracking · Metrics",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "KPI dashboard · Métriques · Reporting",
"url": "/enterprise-model.html"
},
"MANAGER": {
"type": "agent",
"avg": 83,
"icon": "👔",
"caps": {
"combined": 83,
"infra": 80,
"ecosystem": 82
},
"used_in": "Consensus 4 providers · MoA",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Orchestration consensus · Validation · Dispatch",
"url": "/wevia-master.html"
},
"SENTINEL": {
"type": "security",
"avg": 75,
"icon": "🔒",
"caps": {
"combined": 75,
"infra": 80,
"ecosystem": 70
},
"used_in": "S95 sentinel · 344 APIs",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Monitoring S95 · Pipeline email · DB watch",
"url": "/realtime-monitor.html"
},
"WEVADS_IA": {
"type": "backend",
"avg": 78,
"icon": "📧",
"caps": {
"combined": 78,
"infra": 80,
"ecosystem": 76
},
"used_in": "Email platform · 41 pages SPA",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "Campagnes email · A/B test · PMTA · Deliverability",
"url": "/wevads-ia/"
},
"Qwen3_235B_Cerebras": {
"type": "sovereign",
"avg": 83,
"icon": "🛡️",
"caps": {
"combined": 83,
"infra": 82,
"ecosystem": 84
},
"used_in": "Cascade #5 · OpenClaw · WEVIA Master",
"wired": true,
"cost": "🟢 FREE illimité",
"usecase": "Heavy reasoning · 235B params · Backup cascade",
"url": "/ai-benchmark.html"
},
"HF_Qwen2.5_72B": {
"type": "sovereign",
"avg": 82,
"icon": "🛡️",
"caps": {
"combined": 82,
"infra": 81,
"ecosystem": 83
},
"used_in": "Cascade #2 · GPU FREE",
"wired": true,
"cost": "🟢 FREE GPU",
"usecase": "GPU inference gratuit · 72B · Fallback",
"url": "/ai-benchmark.html"
},
"Llama_3.3_70B_Groq": {
"type": "sovereign",
"avg": 81,
"icon": "🛡️",
"caps": {
"combined": 81,
"infra": 80,
"ecosystem": 82
},
"used_in": "Cascade #1 · Primary",
"wired": true,
"cost": "🟡 FREE rate-limit",
"usecase": "Chat primary <1s · DeerFlow · Fast inference",
"url": "/ai-benchmark.html"
},
"NVIDIA_Nemotron_49B": {
"type": "sovereign",
"avg": 80,
"icon": "🛡️",
"caps": {
"combined": 80,
"infra": 82,
"ecosystem": 78
},
"used_in": "Cascade #3 · Unlimited",
"wired": true,
"cost": "🟢 FREE illimité",
"usecase": "No rate-limit · 49B · Enterprise backup",
"url": "/ai-benchmark.html"
},
"SambaNova_DeepSeek": {
"type": "sovereign",
"avg": 79,
"icon": "🛡️",
"caps": {
"combined": 79,
"infra": 78,
"ecosystem": 80
},
"used_in": "Cascade #6 · Deep reasoning",
"wired": true,
"cost": "🟡 FREE rate-limit",
"usecase": "DeepSeek reasoning · Chain-of-thought",
"url": "/ai-benchmark.html"
},
"Mistral_Small_EU": {
"type": "sovereign",
"avg": 78,
"icon": "🛡️",
"caps": {
"combined": 78,
"infra": 76,
"ecosystem": 80
},
"used_in": "Cascade #7 · EU sovereign",
"wired": true,
"cost": "🟡 FREE rate-limit",
"usecase": "EU data residency · RGPD compliant · Fast",
"url": "/ai-benchmark.html"
},
"Gemma4_Local": {
"type": "sovereign",
"avg": 72,
"icon": "🛡️",
"caps": {
"combined": 72,
"infra": 74,
"ecosystem": 70
},
"used_in": "Cascade #4 · Ollama S204",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "100% local · Zero internet · Privacy absolue",
"url": "/openclaw.html"
},
"WEVAL_COMBINED": {
"type": "composite",
"avg": 93,
"icon": "🌟",
"caps": {
"combined": 93,
"infra": 95,
"ecosystem": 90
},
"used_in": "32 AIs wirées · 7 cascade · Auto-failover · MoA 9.8",
"wired": true,
"cost": "🟢 Souverain 0€",
"usecase": "TOUTE la plateforme · Supérieur à tout provider seul",
"url": "/weval-wiring.html"
}
},
"composite": {
"combined": 90,
"infra": 90,
"ecosystem": 90
},
"generated": "2026-04-18T00:05+02:00",
"leaderboard": [
{
"name": "WEVAL_COMBINED",
"score": 93
},
{
"name": "Claude_Opus_4.6",
"score": 90
},
{
"name": "WEVIA_MASTER",
"score": 88
},
{
"name": "GPT_4o",
"score": 87
},
{
"name": "Gemini_2.5_Pro",
"score": 86
},
{
"name": "ChatGPT_4o",
"score": 85
},
{
"name": "DeepSeek_V3",
"score": 84
},
{
"name": "CONSENSUS",
"score": 84
},
{
"name": "Gemini_Pro",
"score": 83
},
{
"name": "MANAGER",
"score": 83
},
{
"name": "Qwen3_235B_Cerebras",
"score": 83
},
{
"name": "WEVIA_PUBLIC",
"score": 82
},
{
"name": "WEVCODE",
"score": 82
},
{
"name": "HF_Qwen2.5_72B",
"score": 82
},
{
"name": "WEVIA_CONSOLE",
"score": 81
},
{
"name": "Llama_3.3_70B_Groq",
"score": 81
},
{
"name": "WEVIA_WIDGET",
"score": 80
},
{
"name": "L99_BRAIN",
"score": 80
},
{
"name": "NVIDIA_Nemotron_49B",
"score": 80
},
{
"name": "WEDROID",
"score": 79
},
{
"name": "DEERFLOW",
"score": 79
},
{
"name": "SambaNova_DeepSeek",
"score": 79
},
{
"name": "CLAW_CODE",
"score": 78
},
{
"name": "WEVADS_IA",
"score": 78
},
{
"name": "Mistral_Small_EU",
"score": 78
},
{
"name": "CORTEX",
"score": 77
},
{
"name": "ETHICA_CHAT",
"score": 77
},
{
"name": "PAPERCLIP",
"score": 76
},
{
"name": "DIRECTOR",
"score": 76
},
{
"name": "BLADE_IA",
"score": 76
},
{
"name": "WEPREDICT",
"score": 75
},
{
"name": "HERMES",
"score": 75
},
{
"name": "SENTINEL",
"score": 75
},
{
"name": "OPENCLAW",
"score": 74
},
{
"name": "GUARDIAN",
"score": 74
},
{
"name": "NUCLEI",
"score": 73
},
{
"name": "KILO",
"score": 73
},
{
"name": "Gemma4_Local",
"score": 72
},
{
"name": "UX_AGENT",
"score": 71
}
],
"report": {
"combined_avg": 80,
"infra_avg": 80,
"ecosystem": 80,
"composite_avg": 80,
"total_ais": 39,
"wired_count": 39,
"coverage_pct": 100,
"updated_17avr": true,
"sync_doctrine": 90,
"ethica_hcps": 146694,
"nonreg_score": "153/153",
"agents_count": 950,
"new_depts": [
"meeting_rooms",
"lean6sigma"
]
},
"usage_map": {
"WEVIA_MASTER": {
"used_by": "Orchestrateur 22/22 exec · Intercept v9",
"cascade": 0
},
"CONSENSUS": {
"used_by": "4 providers MoA · Score 9.8/10",
"cascade": 0
},
"WEVIA_PUBLIC": {
"used_by": "Widget chat public · /wevia",
"cascade": 0
},
"WEVIA_WIDGET": {
"used_by": "Widget embed iframe · wevia-widget.html",
"cascade": 0
},
"WEVIA_CONSOLE": {
"used_by": "Centre Commande · 71 modules",
"cascade": 0
},
"WEVCODE": {
"used_by": "Code assistant · 4 modes fast/deep/code/math",
"cascade": 0
},
"WEDROID": {
"used_by": "Diagnostic backend v5 · 12 providers",
"cascade": 0
},
"CLAW_CODE": {
"used_by": "Code review + generation · /claw-code",
"cascade": 0
},
"CORTEX": {
"used_by": "Visualisation architecture · SVG live",
"cascade": 0
},
"DEERFLOW": {
"used_by": "Deep research · 42 skills · Memory v1",
"cascade": 0
},
"PAPERCLIP": {
"used_by": "Project mgmt · 3 agents · Goals/Issues",
"cascade": 0
},
"WEPREDICT": {
"used_by": "MiroFish rebrand · Port 5001",
"cascade": 0
},
"OPENCLAW": {
"used_by": "Ollama 4 models · Port 11434",
"cascade": 0
},
"L99_BRAIN": {
"used_by": "L99 pipeline · Puppeteer · 29/29",
"cascade": 0
},
"DIRECTOR": {
"used_by": "Supervision */15min · 39 obs",
"cascade": 0
},
"NUCLEI": {
"used_by": "Vuln scanner installed",
"cascade": 0
},
"GUARDIAN": {
"used_by": "SSO watch */5min · chattr +i",
"cascade": 0
},
"BLADE_IA": {
"used_by": "Desktop agent · 34 caps · PowerShell",
"cascade": 0
},
"UX_AGENT": {
"used_by": "Visual QA · 10 pages · Cron */2h",
"cascade": 0
},
"ETHICA_CHAT": {
"used_by": "Pharma HCP · 126K contacts",
"cascade": 0
},
"HERMES": {
"used_by": "Messaging · Chat routing",
"cascade": 0
},
"KILO": {
"used_by": "KPI tracking · Metrics",
"cascade": 0
},
"MANAGER": {
"used_by": "Consensus 4 providers · MoA",
"cascade": 0
},
"SENTINEL": {
"used_by": "S95 sentinel · 344 APIs",
"cascade": 0
},
"WEVADS_IA": {
"used_by": "Email platform · 41 pages SPA",
"cascade": 0
},
"Qwen3_235B_Cerebras": {
"used_by": "Cascade #5 · OpenClaw · WEVIA Master",
"cascade": 0
},
"HF_Qwen2.5_72B": {
"used_by": "Cascade #2 · GPU FREE",
"cascade": 0
},
"Llama_3.3_70B_Groq": {
"used_by": "Cascade #1 · Primary",
"cascade": 0
},
"NVIDIA_Nemotron_49B": {
"used_by": "Cascade #3 · Unlimited",
"cascade": 0
},
"SambaNova_DeepSeek": {
"used_by": "Cascade #6 · Deep reasoning",
"cascade": 0
},
"Mistral_Small_EU": {
"used_by": "Cascade #7 · EU sovereign",
"cascade": 0
},
"Gemma4_Local": {
"used_by": "Cascade #4 · Ollama S204",
"cascade": 0
}
},
"test_method": "auto_benchmark_2h"
}