AI Models

What people actually call

Ranked by tokens actually consumed — not our recommendation and not a benchmark score, but what people are really running.

Source: OpenRouter public rankings · snapshot 2026-08-21 · Scope: traffic on OpenRouter, whose users skew towards coding agents and free-tier usage — not the whole market. Rank movement compares against our own previous snapshot.

Popular is not the same as right for you — the 5 groups below are picked by use case.

⚠️ Usage data is 8 days old — the daily usage sync may have broken, so the ranking below reflects that day, not today.

Coding and agents

For letting the model edit files, run commands and chain many steps on its own. The hard requirement is tool calling — without it a model can only advise, not act.

Rule: tool calling + step-by-step reasoning + context ≥ 200k tokens + a metered first-party price + released within 365 days; newest 6 first.

• Qwen3.8 Flashalibaba

  ≈667k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output

  That job costs about $0.05

• GLM-5.3-Flashzhipuai

  ≈667k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output

  That job costs about $0.02

• DeepSeek V4 Flash Vision Expdeepseek

  ≈667k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output

  That job costs about $0.04

• GLM-5.3zhipuai

  ≈667k Chinese characters of context · Step-by-step reasoning · Calls tools / agents · Structured JSON output

  That job costs about $0.43

• Gemini 3.7 Flashgoogle

  ≈699k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output

  That job costs about $0.24

• Gemini Flash Latestgoogle

  ≈699k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output

  That job costs about $0.24

Long documents in one go

Drop a whole book, contract or a few hundred pages of PDF in at once instead of chunking it yourself. One million tokens is roughly 700,000 Chinese characters.

Rule: context ≥ 1M tokens, released within 365 days; largest context first, top 6.

• Grok 4.1 Fastxai

  ≈1333k Chinese characters of context · Reads images · Calls tools / agents

  That job costs about No public rate

• GPT-5.4openai

  ≈700k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output

  That job costs about $0.80

• GPT-5.4 Proopenai

  ≈700k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents

  That job costs about $9.54

• GPT-5.5openai

  ≈700k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output

  That job costs about $1.59

• GPT-5.5 Proopenai

  ≈700k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output

  That job costs about $9.54

• GPT-5.6 Lunaopenaiassessed

  ≈700k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output

  That job costs about $0.06

Cheap at volume

Simple work at volume — classification, extraction, bulk rewriting. Sorted by the vendor's own input price, cheapest first.

Rule: first-party input price above $0 and at most $0.5 per million tokens + tool calling + released within 365 days; cheapest 6. Plan-billed channels are excluded — that is not $0, just not per-token.

• GLM-4.7-FlashXzhipuai

  ≈133k Chinese characters of context · Step-by-step reasoning · Calls tools / agents · Open weights, self-hostable

  That job costs about $0.02

• GLM-5.3-Flashzhipuai

  ≈667k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output

  That job costs about $0.02

• Step 3.5 Flashstepfun

  ≈171k Chinese characters of context · Step-by-step reasoning · Calls tools / agents · Open weights, self-hostable

  That job costs about $0.03

• Step 3.5 Flash 2603stepfun

  ≈171k Chinese characters of context · Step-by-step reasoning · Calls tools / agents · Open weights, self-hostable

  That job costs about $0.03

• DeepSeek V4 Flashdeepseekassessed

  ≈667k Chinese characters of context · Step-by-step reasoning · Calls tools / agents · Structured JSON output · Open weights, self-hostable

  That job costs about $0.04

• DeepSeek V4 Flash Vision Expdeepseek

  ≈667k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output

  That job costs about $0.04

Reads images

Takes screenshots, photos and scans as input — invoices, tables, UI mockups. This is reading images, not generating them.

Rule: image in the input modalities, released within 365 days; newest 6 first.

• Qwen3.8 Flash Nextalibaba

  ≈175k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output · Open weights, self-hostable

  That job costs about No public rate

• Qwen3.8 Flashalibaba

  ≈667k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output

  That job costs about $0.05

• GLM-5.3-Flashzhipuai

  ≈667k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output

  That job costs about $0.02

• DeepSeek V4 Flash Vision Expdeepseek

  ≈667k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output

  That job costs about $0.04

• Ornith 1.5 35B A3Bdeepreinforce

  ≈175k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Open weights, self-hostable

  That job costs about No public rate

• Qwen3.8 27Balibaba

  ≈175k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output · Open weights, self-hostable

  That job costs about No public rate

Open weights

The weights can be downloaded and run on your own hardware — one of the few options when data cannot leave your perimeter. Self-hosting is not automatically cheaper.

Rule: open weights + a metered first-party price (so a hosted API also exists) + released within 365 days; newest 6 first.

• Kimi K3moonshotaiassessed

  ≈699k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output · Open weights, self-hostable

  That job costs about $0.94

• GLM-5.2zhipuaiassessed

  ≈667k Chinese characters of context · Step-by-step reasoning · Calls tools / agents · Structured JSON output · Open weights, self-hostable

  That job costs about $0.43

• Kimi K2.7 Codemoonshotai

  ≈175k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output · Open weights, self-hostable

  That job costs about $0.30

• Kimi K2.7 Code Highspeedmoonshotai

  ≈175k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output · Open weights, self-hostable

  That job costs about $0.59

• MiMo-V2.5-Pro-UltraSpeedxiaomi

  ≈699k Chinese characters of context · Step-by-step reasoning · Calls tools / agents · Open weights, self-hostable

  That job costs about $0.40

• MiniMax-M3minimax

  ≈699k Chinese characters of context · Reads images · Step-by-step reasoning · Calls tools / agents · Structured JSON output · Open weights, self-hostable

  That job costs about $0.09

All models

Every model in the snapshot is below, each with its own page. The reference-job column prices one fixed scenario (Read a 200,000-Chinese-character book and write a 2,000-Chinese-character summary) at each vendor's own rate; the method is in the glossary.

Video and image generation

The table above comes from models.dev, which is overwhelmingly text APIs — a handful of video, image and speech models are in it, but without prices or guidance, and models like Seedance are not in it at all. This short section is hand-verified and hand-maintained, refreshed when someone checks, not daily.

• Seedance 2.5ByteDance / Volcano Engine

  Text or an image in, video out. Give it a description (or a still plus a description) and it produces a clip; the vendor advertises narratives up to 30 seconds.

• MidjourneyMidjourney Inc.

  Text (optionally with a reference image) in, images out. The same subscription can also turn a still into a short clip — a separate product line — but the core is still image generation.

• 可灵 AI (Kling AI)Kuaishou

  Text or an image in, video (or images) out. The membership plan grants a monthly pool of "inspiration credits"; each image or video generation spends some of them.

Glossary

token

The unit models bill and measure in. One Chinese character is about 1.5 tokens, one English word about 0.75. Prices are quoted per million tokens.

context

How much the model can hold in view at once, your input and its output together. Past that you have to chunk it yourself.

tool calling / agents

Whether the model can look things up, run commands and edit files on its own. Without it, it can only advise.

structured output

Whether it reliably emits JSON a program can parse. In an automated pipeline this decides how much defensive code you write.

open weights

Whether the model file can be downloaded and run on your own hardware. Self-hostable is not the same as cheaper — hardware for a model this size usually costs more than the API.

first-party / resale

“First-party” is the model vendor's own price. Untagged rows are intermediaries: price may be higher or lower, and so are the terms.

How the “that job costs about” figure is computed

The scenario is fixed at “Read a 200,000-Chinese-character book and write a 2,000-Chinese-character summary”, converted at 1.5 tokens per Chinese character — 300,000 input and 3,000 output tokens — times the vendor's own rate. It is an order-of-magnitude estimate, not a quote: the gap between $0.9 and $90 is decision-grade, the gap between $0.9 and $1.1 is not. Models without a public first-party rate read “No public rate”, never $0.