Comparison

v3 vs v4: Choosing a Computer Use Model on the API

Michael Rodriguez||6 min
Ctrl+A

You want to automate real desktop work. A static API only covers a specific endpoint. A human-like computer use agent can see the screen and click, type, and scroll. The Coasty Computer Use API lets you choose between v3 and v4 models. v3 is a general-purpose agent. v4 adds a verifier and autonomous mode. This post covers when to use each.

What the models are

  • Both use the POST /v1/runs endpoint with cua_version to pick the model. v3 runs tasks with a fixed number of steps. v4 is autonomous and includes a pass/fail verifier.
  • v3 is ideal for scripted tasks where you know the steps ahead of time. You define a task, a max_steps limit, and a deadline.
  • v4 is ideal when you need tight alignment with success criteria. It runs until it sees a pass or fail result from the verifier. It also supports on_awaiting_human with pause, fail, or cancel options.
  • Both bill $0.05 per agent step. The verifier in v4 does not add extra cost per step.

v3 basics

  • Use cua_version: "v3" in the POST /v1/runs request.
  • Provide machine_id, task, and optional instructions, system_prompt, max_steps, deadline_seconds, on_awaiting_human, and webhook_url.
  • The server runs the agent and emits events via GET /v1/runs/{id}/events. You can pause, resume, or cancel the run with the respective endpoints.
  • States include queued, running, awaiting_human, succeeded, failed, cancelled, timed_out.

v4 basics

  • Use cua_version: "v4".
  • The verifier is built into the v4 model. It returns pass/fail after each completed step.
  • You can still set max_steps and deadline_seconds as guards, but the model may stop earlier if it receives a final result.
  • on_awaiting_human lets you decide how to handle pauses. Options: pause, fail, or cancel.
bash
curl -X POST https://coasty.ai/v1/runs \
  -H "X-API-Key: $COASTY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "machine_id": "vm-123",
    "task": "Open Chrome, navigate to https://google.com, and type a search query",
    "cua_version": "v4",
    "instructions": "Search for Coasty AI and open the first link.",
    "max_steps": 50,
    "deadline_seconds": 120,
    "on_awaiting_human": "pause",
    "webhook_url": "https://your-server.com/webhook"
  }'

How to inspect a run

  • GET /v1/runs shows all your recent runs.
  • GET /v1/runs/{id} gives the current state and metadata.
  • GET /v1/runs/{id}/events streams events as a Server-Sent Event stream. Reconnect with Last-Event-ID if needed.
  • POST /v1/runs/{id}/cancel stops the agent immediately. POST /v1/runs/{id}/resume lets you continue after a pause.

Set cua_version to v3 for scripted tasks. Set cua_version to v4 when you need a built-in pass/fail verifier and autonomous mode.

Where this beats brittle automation

  • A computer use agent sees the actual screen. It can handle layout changes, missing buttons, or unexpected popups.
  • You do not need fragile selectors like XPath or CSS classes that break on UI updates.
  • API-only tools only cover a fixed endpoint. A computer use agent can drive a browser, terminal, or desktop app like a human.
  • The v4 verifier ensures the task meets your success criteria, reducing false positives.

Pick v3 when you want a reliable, step-limited agent. Pick v4 when you need an autonomous agent with a pass/fail verifier. Try both by sending requests to POST /v1/runs with the appropriate cua_version. Get your API key at https://coasty.ai/developers and start automating real workflows.

Want to see this in action?

View Case Studies
Try Coasty Free