v3 vs v4: Choosing a Computer Use Model on the API
Priya Patel||6 min
Cmd+V
You wire a vision model to a desktop, but you still need to decide the level of control v3 or v4. v3 is a controllable loop where you drive capture, predict, act. v4 adds an autonomous run with a built‑in pass/fail verifier. Pick the right model for the job: tight control or full automation.
v3: controllable agent loop
- ●Use POST /v1/runs with cua_version set to "v3".
- ●Server queues the run and streams events via GET /v1/runs/{id}/events.
- ●States include queued, running, awaiting_human, succeeded, failed, cancelled, timed_out.
- ●You can cancel or resume a v3 run using POST /v1/runs/{id}/cancel and POST /v1/runs/{id}/resume.
- ●Billed $0.05 per agent step.
v4: autonomous runs with built-in verification
- ●Use POST /v1/runs with cua_version set to "v4".
- ●v4 runs autonomously with a pass/fail verifier.
- ●You can still cancel or resume a v4 run using the same endpoints.
- ●Billed $0.05 per agent step.
- ●Best for fully automated end-to-end tests or production workflows.
How to submit a v3 run
- ●Send POST /v1/runs with machine_id, task, cua_version: "v3", and optional instructions.
- ●Include max_steps and deadline_seconds to constrain the run.
- ●Set on_awaiting_human to "pause", "fail", or "cancel" when the agent asks for human input.
- ●Provide a webhook_url for async notifications.
- ●Use the idempotency-key header to safely retry writes.
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-12345",
"task": "Open YouTube, search for a video, and click the first result",
"cua_version": "v3",
"instructions": "Verify the video title matches the search term",
"max_steps": 20,
"deadline_seconds": 300,
"on_awaiting_human": "pause",
"webhook_url": "https://your-domain.com/v1/runs/callback"
}'
How to submit a v4 run
- ●Set cua_version to "v4" in the same POST /v1/runs payload.
- ●Use the same optional fields (instructions, max_steps, deadline_seconds, on_awaiting_human, webhook_url).
- ●v4 runs autonomously and reports success or failure through the run state.
- ●Billed $0.05 per agent step.
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-12345",
"task": "Open a web app, log in, and create a new project",
"cua_version": "v4",
"instructions": "Confirm the project is listed in the dashboard",
"max_steps": 25,
"deadline_seconds": 600,
"on_awaiting_human": "fail",
"webhook_url": "https://your-domain.com/v1/runs/callback"
}'
Use cua_version: "v3" for fine-grained control, or "v4" for fully autonomous runs with a built-in pass/fail verifier.
Why computer use beats brittle automation
- ●v3 and v4 drive real desktops, browsers, and terminals directly.
- ●They see the screen and act like a human, so they adapt to UI changes.
- ●No brittle selectors or API-only tools that break on layout shifts.
- ●v4 adds a verification step that gates success/failure on real behavior.
- ●Both models are billed $0.05 per agent step, no hidden per-call costs.
Pick the model that matches your automation pattern. v3 gives you tighter loop control for debugging and interactive workflows. v4 runs entirely on its own with built-in verification. Ready to build? Get a key at https://coasty.ai/developers.