LLM comparison: price vs. intelligence

How to read the table and choose a model

What the columns mean

• Intel — Intelligence Index — Artificial Analysis’s aggregate score: higher means a smarter model on average.
• Mode — Reasoning / non-reasoning and the effort level. Different modes of one model are separate rows with their own score.
• Rsn? — A quick flag: is this a reasoning model (✓) or not. It mirrors the Mode column, but it’s handy for sorting and filtering by type.
• Providers — How many platforms on OpenRouter serve the model. Click the number to see each provider’s price, latency, uptime and quantization.
• in $/M and out $/M — Price per million input and output tokens (from OpenRouter).
• $/intel — Input price divided by the Intelligence Index: the lower it is, the cheaper a unit of intelligence.
• TTFT (s) and t/s — Speed: time to first token, and output tokens per second.
• Context (k) — Context-window size in thousands of tokens.
• Tokenizer — The model’s tokenizer family; it affects the real cost of non-English text (see below).
• Maker — The company that built the model (Anthropic, OpenAI, Google…) — not the inference provider.

What the Intelligence Index is

The Intelligence Index from Artificial Analysis rolls several benchmarks into one number, so you can compare models on a single scale instead of juggling ten leaderboards. Treat it as a solid average, not gospel: a model that trails by a point or two may still win on your specific task. Use it to shortlist, then test the finalists on your own prompts.

Why $/intel can mislead for reasoning models

Reasoning modes score higher at the same per-token price, so their $/intel looks better. But reasoning is far more verbose — it burns a lot more output tokens — so the real cost of an answer is higher than the column suggests. Read $/intel together with the mode and the output price, not on its own.

Tokenizers and non-English text: what’s cheaper

OpenRouter prices are per token, so the real cost of, say, Russian or Arabic depends on how finely a tokenizer splits the script. Rough order, cheapest to priciest: Gemini/Gemma and GPT (o200k) → Llama 3, Mistral, Command, Claude → Chinese models (Qwen, DeepSeek, GLM, Kimi, MiniMax), which chop non-Latin text into more tokens. It’s a qualitative guide, not exact multipliers — only a measurement on your own text gives the real number.

Maker vs. provider — don’t mix them up

Two different things share this table. The maker is who built the model (Anthropic, OpenAI, Google, Xiaomi). A provider is a platform that serves that model over OpenRouter, each with its own price, latency and uptime. One model can have a dozen providers — click the number in the Providers column to compare them side by side.

How to pick in practice

Start from the task, not the hype. Set a minimum Intelligence Index you’re comfortable with, sort by $/intel, and scan the top rows. Need long documents? Filter by context. Watching latency? Sort by TTFT. Working in a non-English language? Lean toward a friendlier tokenizer. Then pass the model you chose to hotdoc and run it against real files.