v3 vs v4: Choosing a Computer Use Model on the API
When you call Coasty to drive a desktop, browser, or terminal, you choose a cua_version. v3 is the default, it runs a stateless agent loop that reads a screenshot, predicts actions, and repeats until you tell it to stop. v4 is an upgrade that wraps that loop with an automatic pass/fail verifier, so the server decides whether the run succeeded or failed without manual review. Both use the same base endpoint and billing, but v4 gives you a higher assurance level for production workloads.
The v3 model
- ●You POST to /v1/runs with cua_version: "v3" (default).
- ●The server drives the agent step by step, billing $0.05 per agent step.
- ●You provide a task and optional instructions that append to the base prompt.
- ●You can set max_steps, deadline_seconds, and on_awaiting_human to pause, fail, or cancel.
- ●The run emits streaming events until the agent finishes or hits a limit.
- ●Use this for quick scripts, prototypes, or when you want full control over the loop.
The v4 model
- ●You POST to /v1/runs with cua_version: "v4".
- ●v4 includes a pass/fail verifier that watches the run and decides final status.
- ●Billing stays the same: $0.05 per agent step.
- ●You can still set max_steps, deadline_seconds, and on_awaiting_human.
- ●The final state is either succeeded or failed, and you get a clear success flag.
- ●Use this for production jobs where you want the server to decide success.
When to pick which
- ●Choose v3 for custom control: you decide when to stop, iterate, or intervene.
- ●Choose v4 for reliability: let the verifier handle success/failure logic.
- ●Both models see the screen and act like a human, so you get the benefits of computer use.
- ●The pricing and endpoints are identical except for the cua_version field.
- ●If you want an autonomous run with a clear outcome, start with v4.
Set cua_version to "v4" for an automatic pass/fail verifier, or stick with "v3" when you want to control the loop yourself.
Where computer use beats brittle selectors
API-only tools depend on stable URLs, element IDs, and JSON payloads. When a UI changes, a selector fails, or a page is dynamic, your automation breaks. Computer use models like v3 and v4 read the actual screenshot, understand the context, and click, type, or scroll based on what they see. They adapt to layout changes, missing IDs, and unexpected states without you needing to rewrite selectors. This makes them ideal for real-world dashboards, legacy UIs, and browser workflows that resist traditional tooling.
Start with v4 for a hands-off, reliable run, or use v3 when you want tight control over each step. Either way, you get a computer use agent that drives real desktops and browsers. Get your API key at https://coasty.ai/developers and try it today.