AI WAR ROOM — 350+ AI models, ranked by live arena ELO

AI WAR ROOM is an independent, daily-updated leaderboard of the strongest large language models in the world. We pull live ELO scores from arena.ai's human-vote battle system, combine them with current pricing from OpenRouter, and present the entire field — 350+ models from OpenAI, Anthropic, Google, xAI, Meta, Alibaba, DeepSeek, Moonshot, Z.ai, Mistral, Baidu, ByteDance, Xiaomi, Tencent and more — in one ranked, filterable view.

Every score comes from real human preference: anonymous head-to-head battles where voters pick the better of two responses. The result is a relative skill rating — the same ELO system used in competitive chess — that is harder to game than benchmark numbers and more useful than marketing claims.

What you'll find here

• War Room — the command-center overview: strongest, fastest, cheapest and best-value models across language, code, image, video and voice, with 7-day movement.
• LLM Rankings — every active model ranked by arena ELO, with confidence intervals, votes, Intelligence Index, live input / output / cached pricing, context window, max output, speed and latency. Sort, filter, choose columns, or switch to interactive charts (intelligence vs price, ELO vs price, context vs price, Pareto frontier).
• Code Ops — coding agents and models on SWE-bench (Verified, Lite, Multimodal, Multilingual): resolved rate, cost per task, calls per task, with agent and model kept separate and a model-by-harness matrix.
• Image Arena & Video Arena — text-to-image, image editing, text-to-video and image-to-video models ranked by quality ELO with price per image / per second, audio support and a side-by-side Battle Mode.
• Voice Comms — speech-to-text word error rate and throughput from the Open ASR Leaderboard; text-to-speech and speech-to-speech boards when a measurement source is configured.
• Provider War — the same model across inference providers: price, context, quantisation and uptime, with the cheapest and best-value host marked.
• Benchmarks — a capability matrix across human preference, composite indexes, coding, agentic, design, reasoning, math, long context, multimodal and speech, each with source and methodology.
• Battle Mode, Speed Race, Mission Planner — compare up to four models head to head, watch real measurements as a race, or get a mathematically derived recommendation for your workload.
• The Guide — a plain-English walkthrough of how arena ELO works, the 2026 lab landscape, how to pick a model, and common pitfalls.
• FAQ — honest answers to the questions we get most often about ratings, pricing, thinking variants, open-weight models, and arena mechanics.
• Blog — long-form analysis on reading ELO leaderboards, open vs closed weights in 2026, and when thinking models help or hurt.
• Model and lab profiles — in-depth pages for the top frontier models and the labs building them.
• Methodology — the exact rules behind the rankings, the ELO tiers (S/A/B/C), and how pricing data is sourced.

How to choose a model

The #1 model is rarely the model you want. Top-ranked frontier models are the most expensive, slowest, and built for the hardest reasoning. For most real workloads — chat, summarisation, classification, RAG pipelines — a model in positions 5–20 will perform indistinguishably for a fraction of the cost. Start by sorting on ELO to identify the top tier, then filter on context window, licence (open vs proprietary), and price. Run your own evals on the three candidates with the best fit. The arena gives you the shortlist; your own judgement picks the winner.

How the data is sourced

ELO ratings come from arena.ai's public leaderboards (text, vision, search, image and video arenas), refreshed every six hours. Pricing, context and provider data come from OpenRouter's public catalogue. Coding results come from the SWE-bench leaderboard, speech-recognition results from the Hugging Face Open ASR Leaderboard, and intelligence indexes from Artificial Analysis (relayed via OpenRouter, or directly when its Data API is configured). Every metric carries its source; values we cannot obtain from a permitted source are shown as N/A, never estimated. Model descriptions are maintained by hand from public lab announcements and model cards. AI WAR ROOM is not affiliated with any AI lab or any of these data providers — we surface their public data and add editorial context.

Start exploring

Enter the War Room · Open the leaderboard · Code Ops · Image Arena · Video Arena · Voice Comms · Provider War · Benchmarks · Read the guide · Browse the FAQ · Visit the blog · See the methodology

Loading interactive view… If you're seeing this for more than a few seconds, JavaScript may be disabled in your browser. The content above is a summary; the full leaderboard requires JavaScript to load.