blog
Gemini 3.7 Flash on every pad, no API key
Gemini 3.7 Flash is now included with every pad. It joins Gemma 4 31B and Gemini 3.6 Flash: three models in OpenCode, no provider account, no key to paste, usage included with the pad.
What you get
Each model is its own provider entry in OpenCode, served from the reachpad model gateway at agent.reachpad.dev. The picker shows all three on a pad.
| Model | Provider | Base path |
|---|---|---|
| Gemma 4 31B | reachpad | /v1 |
| Gemini 3.6 Flash | reachpad-gemini | /gemini/v1 |
| Gemini 3.7 Flash | reachpad-gemini-3-7 | /gemini-3-7/v1 |
How to use it
There is nothing to install and nothing to paste. A pad arrives with the providers configured and a credential of its own, so the models are ready before you open a shell. Ask OpenCode what the pad has:
$ opencode modelsThat prints one line per model, provider first. Run a one-shot against 3.7 with the same id:
$ opencode run -m reachpad-gemini-3-7/gemini-3-7-flash "explain this repo"In the OpenCode TUI, /model opens the picker and the status bar shows which provider answered — worth a glance, since each model is a separate entry and the wrong one answers just as fluently. To pin a default, set model in ~/.config/opencode/opencode.json; a pad keeps that choice.
The gateway speaks OpenAI-compatible chat completions, so anything that talks to that API works too. The credential is exported on the pad as ARK_TOKEN:
$ curl -H "Authorization: Bearer $ARK_TOKEN" https://agent.reachpad.dev/gemini-3-7/v1/modelsClaude Code and Codex are installed on every pad and keep using your own accounts. The included models are another option, not a replacement for an agent setup that already works for you.
What is included
There is no per-token charge and no separate model subscription. Usage is part of the pad, with a daily allowance that resets at midnight UTC so one runaway agent loop cannot consume an unbounded amount of inference. OpenCode surfaces a quota response if a pad reaches that ceiling.
No provider key is ever copied onto a pad. Each pad gets its own revocable credential, and the reachpad gateway adds the upstream key only while relaying the request. We meter the traffic rather than reading it: no prompt or response is parsed or stored to make the allowance work.
The included models accept text today. Bring your own provider account when a task needs a different model or input type.
Try it on a pad
All three models are available on new and existing pads, and a pad picks them up on its next converge. See the pad and workspace trial details, or create an account to start.