Which Large Language Model Leads in Intelligence, Speed, and Cost? 2026 Rankings Revealed

Overview

The AI evaluation platform Artificial Analysis released a new global ranking of large language models (LLMs) covering three key dimensions: Intelligence Index, Output Tokens per Second (speed), and price in USD per million tokens. The ranking compares models from major providers such as Google, OpenAI, NVIDIA, Anthropic, as well as Chinese open‑source models.

Intelligence Index (Core Capability)

The Intelligence Index measures a model’s overall reasoning, knowledge, and creativity. The top scores are:

Gemini 3.1 Pro Preview (Google) : 57 points (tied for 1st)

GPT‑5.4 (xhigh, OpenAI) : 57 points (tied for 1st)

Claude Opus 4.6 (max, Anthropic) : 53 points

Claude Sonnet 4.6 (max, Anthropic) : 52 points

GLM‑5 (Zhipu AI) : 50 points – the highest among open‑source and Chinese models, ranking 5th globally

DeepSeek V3.2 : 42 points – moderate intelligence but excellent cost‑performance

Commentary: Gemini 3.1 Pro Preview and GPT‑5.4 remain at the intelligence ceiling, while GLM‑5 continues to improve, staying within the global top‑5, demonstrating that Chinese models can match the leading international tier.

Output Speed (Tokens per Second)

Speed directly affects the fluidity of chat, writing, and code generation. The fastest models are:

NVIDIA Nemotron 3 Super : 455 tokens/s (rank 1)

gpt‑oss‑120B (high) : 279 tokens/s (rank 2)

Grok 4.20 Beta 0309 : 216 tokens/s

Gemini 3 Flash : 166 tokens/s

Gemini 3.1 Pro Preview : 125 tokens/s

GPT‑5.4 (xhigh) : 73 tokens/s

GLM‑5 : 67 tokens/s

Claude Sonnet 4.6 (max) : 55 tokens/s

Claude Opus 4.6 (max) : 48 tokens/s

DeepSeek V3.2 : 28 tokens/s

Commentary: NVIDIA Nemotron 3 Super dominates speed, earning the “lightning” title. Grok 4.20 and Gemini Flash also offer strong performance for real‑time applications, while GLM‑5 provides moderate speed suitable for many use cases.

Price (USD per 1M Tokens)

Cost per million tokens influences large‑scale deployment budgets. The cheapest options are:

gpt‑oss‑120B : $0.30

DeepSeek V3.2 : $0.30 (tied for cheapest)

NVIDIA Nemotron 3 Super : $0.40

Gemini 3 Flash : $1.10

GLM‑5 : $1.60

Grok 4.20 Beta 0309 : $3.00

Gemini 3.1 Pro Preview : $4.50

GPT‑5.4 (xhigh) : $5.60

Claude Sonnet 4.6 (max) : $6.00

Claude Opus 4.6 (max) : $10.00

Commentary: DeepSeek V3.2 and gpt‑oss‑120B achieve the best price‑performance, while top‑tier models like Gemini, GPT‑5.4, and Claude Opus command higher prices suitable for scenarios demanding the highest intelligence.

Choosing the Right Model

Maximum intelligence (complex writing, reasoning) : Gemini 3.1 Pro Preview or GPT‑5.4 (57 points).

Maximum speed (real‑time chat, code generation) : NVIDIA Nemotron 3 Super (455 tokens/s).

Best cost‑performance (high‑frequency daily use, startups) : DeepSeek V3.2 or gpt‑oss‑120B ($0.30 per M tokens).

Best choice for Chinese users : GLM‑5 (top‑5 intelligence, reasonable price) and DeepSeek V3.2 (cheapest).

There is no absolute “best” model; the optimal choice depends on the specific trade‑off between intelligence, speed, and budget.