Reading a reply: its model, provider, speed, cost and code

Every answer on tryingopen names the model that wrote it, and most also show which company ran it, how long it took, how many tokens it used and what it cost us. Cloud replies also carry the request itself, ready to send with your own OpenRouter key. Here is each one, as of October 2026.

Who answered

Above each reply sit the model's name and its maker, such as GLM 5.3 Flash from Z.ai. When we know which company ran the model, a label follows: "via CoreWeave", "via Groq" and so on. If a provider is set in the composer, whether you picked it or the chat started on CoreWeave, its name is there as soon as the reply starts. With none set, the name OpenRouter reports fills in when the reply finishes, unless the turn offered web search or went through ngrok's gateway; those replies carry no label. Choosing the company that runs your model covers how the provider gets set.

Replies from the default pick, LFM2.5-VL 3B ("Runs on an edge device"), say more. A reply our small model wrote reads "via Offline Base", with a line saying where it ran: "Answered by the small model on an Offline Base edge device. The d1 router judged this one small enough for it." When d1 sends a turn to the cloud, the line names the cloud model instead. These replies also draw their route through d1. How the d1 router picks between our small model and the cloud explains it.

How this reply was produced

Since September 20, 2026, each finished reply has a small info icon at the end of its action row, labelled "How this reply was produced". Hover over it or tap it, and a panel opens with the model and company at the top ("GLM 5.3 Flash on CoreWeave", for example) and these rows:

Speed leaves out the wait before the first token on purpose: a big model can take many seconds to start, and counting that wait would report a rate nobody saw.

Thinking counts as output: a reasoning model's thinking tokens go into Tokens out and Speed, because the provider bills for them.

A share link keeps each reply's model name and leaves out the company and the measurements, so a chat opened from a link still says who answered and shows no info icon. Sharing a chat, a model setup or a preset as a link covers what else a link holds.

The cost line and "Free for you here"

The Cost row comes from OpenRouter's usage accounting for that reply, added up across each step when the model searched the web or wrote a file along the way. It reflects whichever company served the reply.

Below a cent, the figure runs to five decimal places, so a small reply still shows a real number in place of $0.00. At a cent or more it shows four.

Only replies that went through OpenRouter carry the Cost row. Replies from our own boards, or through ngrok's gateway, show the other rows without it.

The "us" in "What this reply cost us" is deliberate. You pay nothing. The model picker says so at the top: "$ to $$$ is what each model costs to run. Free for you here." Each model in the picker carries a band of $, $$ or $$$, and hovering it reads "Cheap to run", "Mid-priced to run" or "Expensive to run". We set each band from the model's output price when we added it.

The band gives a rough price before you send anything, and the Cost row gives what one reply cost once it's done. A tour of the 24 models in the picker lists which models sit where.

Get the code for this reply

Since September 22, 2026, each cloud reply has a code icon in the same row, labelled "Get the code for this reply". It opens a panel titled "Run this yourself" with a curl command: the OpenRouter API request for that model and, when one was set, that provider. The panel says "You need your own OpenRouter key", and the command reads it from an environment variable. For GLM 5.3 Flash on CoreWeave, asked why the sky is blue, it looks like this:

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "z-ai/glm-5.3-flash",
  "provider": {
    "only": [
      "coreweave"
    ],
    "allow_fallbacks": false,
    "zdr": true
  },
  "messages": [
    {
      "role": "user",
      "content": "Why is the sky blue?"
    }
  ]
}'

In the provider block, a provider set in the composer becomes only, with fallbacks off. Where OpenRouter has a zero-retention host for the model, the block adds zdr: true and the panel says so. For a model with no such host, the flag is left out and the panel says the model "has no zero-retention host yet". A model that reached you through ngrok's gateway, such as GLM 5.3 Flash with no provider picked, still gets an OpenRouter command. OpenRouter's docs on provider routing and zero data retention explain each field.

The command is the smallest request that works: your question (a placeholder stands in past 600 characters), the model and the provider settings. Earlier turns, our system prompt, tools and attachments stay out, so a rerun gets a fresh answer from the same model (and the same company, when one was set), and its wording may differ from the page's. The panel also links to the model's OpenRouter page, such as GLM 5.3 Flash, and to the provider's own site when a provider was set.

A reply written on our boards has no Get the code. The command targets OpenRouter, and our fleet can't be reached with an OpenRouter key. When d1 sends a turn to the cloud, that reply gets the button for the cloud model that wrote it. Replies from a model on your own computer in Self Chat are marked "On your computer" and have no button either (Use Ollama, LM Studio, llama.cpp or Jan from Self Chat).

A note on /developers

We've also published the Tryingopen API: "An OpenAI-compatible API for open models, where your key is a name you pick." The docs and the model list are public, and GET /v1/models answers without a key. Accounts aren't open yet. As of October 14, 2026, the Keys tab reads "Accounts aren't available here yet. Check back soon." Until then, Get the code is the way to take a model from this page into your own code.