Twenty current models, every provider, every rate

A price list goes stale twice: when a price changes, and when a model is replaced. GLM-5.2 is still for sale, but once GLM-5.3 is out nobody starting today should be pricing 5.2. So every model here is the newest version of its family. A Flash or a mini is its own family, because it is its own product at its own price, so GLM-5.3 Flash sits beside GLM-5.3 rather than under it.

The twenty are a lineup by maker, chosen by rule. From each of ten makers we took its main line: the strongest general model and the fast tier. xAI, Moonshot and MiniMax each have one current model, so they give one. The three slots that frees go to the two makers outside the ten with the most use that week, Xiaomi and StepFun.

The prices come from everywhere they are published, not from one marketplace: OpenRouter's list of hosts, the sellers' own published rates (Amazon Bedrock, Google Vertex, Azure, Fireworks, Together, Nebius and dozens more), the routers that name the host behind each price (Hugging Face, Requesty, Experiential, the European router Cortecs), resellers such as NanoGPT and OpenCode Zen, and hosts that publish their own lists: DeepInfra, Chutes, Synthetic, SambaNova, Telnyx, LithosAI, Qubrid, Supernova Labs, PrimaLabs and River.

The closed models are short lines in one colour: the maker's price, repeated by the clouds that resell it. The open-weight models are long lines in every colour. GLM 5.3 has 52 hosts, GLM 5.3 Flash has 50 hosts, DeepSeek V4.1 Flash has 49 hosts, and the dearest charges 6 to 8 times the cheapest.

Use does not follow the length of the line. The most-used model that week, DeepSeek V4.1 Flash, is also one of the cheapest, at $0.11 blended. Use here is OpenRouter's weekly count, the only public measure of use there is; the makers and clouds do not publish theirs.

Claude Fable 5.1 costs $10.00 / $50.00 from Anthropic, Amazon, Google and Microsoft alike. The clouds resell the maker's model at the maker's price, and add a tenth when you ask for the work to stay in one region. That is the whole of the price difference between sellers of a closed model.

The real choice is speed. OpenAI sells GPT-6 Astra at five prices:

Twelve times from the slowest to the fastest, for the same answer. Google sells Gemini the same way, at half price for flex and 1.8 times for priority, and xAI charges double for priority on Grok 4.7. Work that can wait an hour, like an overnight batch or a backlog of documents, should almost never pay the standard price.

One more line hides in the fine print: a long prompt costs more. Once a prompt passes 272,000 tokens, GPT-6 Astra charges double to read and half as much again to write, so a job that sends whole contracts is priced at $20.00 / $75.00, not $10.00 / $50.00.

An open-weight model can be served by anyone with the hardware. Each bar below is one GLM-5.3 host's cheapest standard offer, with how it runs the weights, in the colour of how it is sold. Z.ai's own is the bar in bold.

The maker's price works as the going rate: 30 offers charge exactly $1.40 / $4.40, and 37 undercut it. Most of the undercutting comes from running the weights at lower precision. An fp4 endpoint stores each number in four bits where fp8 uses eight, which halves the memory and cuts the cost, and can make the answers measurably worse on hard work. "Not stated" means the host does not say.

The same host can also charge differently by route. Wafer sells GLM-5.3 at $0.079 / $4.80 through OpenRouter and $1.19 / $4.40 through Experiential. A router's shelf is the router's prices, not the host's.

A host has two prices, reading and writing, and they do not move together. InferenceNet via OpenRouter charges $0.080 to read and $4.40 to write. Baidu via OpenRouter charges $0.35 and $1.10. Which is cheaper depends on what the job does with its tokens. Here are two jobs on GLM-5.3, priced on every offer:

The host that wins the reading job costs 3.8 times as much on the writing job. A single "blended" number, ours included, assumes a job reads three tokens for each one it writes. Price your own ratio: an agent that reads whole repositories is a reading job, a model drafting reports is a writing job, and they want different hosts.

Every number on this page came from two questions, and hedwigai ask can now answer both for any model, at the moment you ask. Its model-prices program lists the current models, newest version per family and never one it replaced, and for any one of them lists every seller it can find, each with its reading, writing and cache prices, how it runs the weights, the route it is sold through, and the cheapest for reading, for writing and blended. You can name a model loosely: "glm 5.3", "claude opus", or just "kimi" for the newest Kimi.

ask is a tool on hedwigai's MCP server, the standard way AI assistants plug in outside tools. In Claude Code:

claude mcp add --transport http --scope user hedwigai https://mcp.hedwigai.com

Turn on model-prices for your account, then ask, for example:

Which host serves GLM-5.3 cheapest for a job that reads ten tokens for every one it writes? Say how each runs the weights.