AI Model Leaderboard, September 2026

How We Rank AI Models

Our leaderboard uses a composite quality index that combines three key benchmarks: MMLU Pro (measuring knowledge and reasoning across 57 subjects), HumanEval (measuring code generation ability), and MATH (measuring mathematical problem-solving). Scores are normalized to a 0-100 scale and cross-referenced against LMSYS Chatbot Arena ELO ratings for real-world validation.

We track speed (tokens per second), time-to-first-token (TTFT), pricing, and context window size to give you a complete picture. The Value Score divides quality by cost, showing you which models deliver the most capability per dollar.

Key Trends in AI Model Performance

• A new frontier #1: Anthropic's Claude Opus 5 (24 July 2026) tops the board — a step change over Opus 4.8 on deep reasoning and long-horizon agentic work, at unchanged $5/$25 pricing and roughly half the cost of Claude Fable 5. Thinking is on by default and the prompt-cache minimum halves to 512 tokens
• OpenAI split GPT-5.6 into three tiers: Sol ($5/$30), Terra ($2/$12), and Luna ($0.20/$1.20), GA 9 July. On 30 July OpenAI cut Terra 20% and Luna 80%, citing inference work that reduced serving cost 20% — the steepest cut of the year from a US lab, and a direct answer to the Chinese open-weight tier on price
• Open weights have caught the frontier: Moonshot's Kimi K3 (2.8T MoE) ranks #3 overall on the Artificial Analysis Intelligence Index — ahead of every proprietary model except Claude Fable 5 and GPT-5.6 Sol — and took #1 on the Frontend Code Arena. It is not a lone outlier: DeepSeek V4 Pro posts 80.6% on SWE-bench Verified (matching Gemini 3.1 Pro) and MiniMax M3 80.5%, while GLM-5.2 scores 62.1% on SWE-bench Pro against GPT-5.5's 58.6% — an MIT-licensed model you can self-host outscoring a $5/$30 US flagship on agentic coding
• "Open" now means datacentre-scale: the leading open models are trillion-parameter MoEs — GLM-5.2 needs ~1TB VRAM in BF16 (~8x H200 at FP8), Kimi K3 needs 64+ accelerators. The ecosystem has bifurcated into giant open MoEs and genuinely small models (Qwen 3.6 27B, Gemma 4, Nemotron 3 Nano Omni) that run on one GPU or a phone, with little in between
• Meta left the open frontier: Llama 5 has not shipped and is now forecast for 2027. Meta pivoted to its first closed frontier model, Muse Spark (April 2026), with no weights and no architecture paper — leaving Chinese labs (DeepSeek, Moonshot, Z.ai, Alibaba, MiniMax) as the effective owners of the open-weight frontier
• Reasoning is built in: Frontier models like GPT-5.6 Sol, Claude Opus 5, and Gemini 3.1 Pro ship extended thinking by default, while deep-research variants trade latency for accuracy on the hardest tasks
• Million-token context is standard: 1M-token windows are now table stakes for flagship models across OpenAI, Anthropic, Google, and xAI
• DeepSeek is the price floor: V4 Pro sits at $0.435/$0.87 per 1M tokens — launched at $1.74/$3.48, cut 75% as a promotion, then made permanent on 31 May 2026 — and V4 Flash at $0.14/$0.28 is the cheapest usable model anywhere, with cache hits at $0.0028. Measured cost-per-task is about $0.04 on V4 Pro against $0.94 on Kimi K3 and far more on the closed frontier. One dollar buys ~1.15M output tokens on V4 Pro, ~227K on GLM-5.2, ~67K on Kimi K3. Watch one thing: DeepSeek has announced 2x peak-hour pricing for two daily windows, not yet active as of 4 August

Choosing the Right Model

There is no single "best" model: it depends on your use case. For most applications, a model routing approach works best: route simple queries to fast, cheap models and complex queries to frontier models. This gives you the best of both worlds: low cost and high quality.