Everything here takes minutes, not hours. No card, no email, no account to create on this website — download the Sandbox app for Windows, install it, and you're comparing models in under a minute. (macOS build is coming soon — join the waitlist on the pricing page.)

1 · Download and install (30 seconds)

  1. Click Download for Windows on any page. The installer is a small .exe — no admin rights required on most machines.
  2. Run it. No email, no password, no account to set up here — Sandbox opens straight to your workspace.
  3. Your trial starts automatically: 300 trial credits and full Pro-level access for 3 days, completely free.
After your 3 free days: if you want to keep using Sandbox, a payment link appears inside the app — that's the only place billing details are ever collected. Nothing is charged before then, and if you do nothing, your sandbox simply pauses.
Minimum requirements: Windows 10 (64-bit, version 21H2+) or Windows 11, a dual-core 2.0 GHz+ processor, 4 GB RAM and 500 MB free disk space, plus a broadband connection — the models run in the cloud, the app is just the window into them.

2 · Turn on your models

Open Settings → Models. Every model in the directory is a toggle:

  • Chat & text: GPT-5, o3, Claude Opus/Sonnet, Gemini 2.5 Pro/Flash, Grok 4, DeepSeek V3, Llama 4, Mistral Large.
  • Image: Midjourney v7, DALL·E 3, Imagen 4, Stable Diffusion 3.
  • Voice & music: ElevenLabs voices, Suno v4.
  • Video: Runway Gen-4.

Enabled models appear as lanes in your workspace. Starter plans include the 8 core models; Pro and above unlock the full directory.

3 · Run your first parallel comparison

  1. Type any real task into the prompt bar — the one you'd normally ask one chatbot.
  2. Press ⌘/Ctrl + Enter to send it to all enabled models at once (or pick specific ones).
  3. Answers render side by side with latency, credit cost and a quality score per lane.
  4. Click Best answer to adopt one, or Merge to blend fragments from several into a final draft.
Why this matters: after a week of comparisons you'll know, from your own data, which model to route each task type to — and you'll stop paying attention to model hype entirely.

4 · Delegate to agents (Pro and up)

Open Agents and describe an outcome instead of a prompt:

  • Code Agent — point it at a GitHub repo, describe the change, approve the plan, review the diff.
  • Research Agent — brief a market or competitor study; receive a sourced report in the background.
  • Media Agent — brief a video campaign; approve scripts and frames at the gate; receive rendered video.
  • Copy Agent — brief a launch; receive multi-angle, multi-language copy scored to your brand voice.

Every agent run shows a plan before it starts, a spend cap you set, and a receipt after it finishes.

5 · Manage credits honestly

  • Costs are per answer, not per subscription: a standard text answer is 1–3 credits; images 4; video 12 per 5 seconds. The exact price shows before anything expensive runs.
  • Soft cap: set a monthly ceiling in Settings — heavy jobs pause at 90% instead of overspending.
  • Rollover: Pro credits roll over one month; Ultra three months.
  • Top-ups: $10 = 400 credits anytime, any plan, no plan change.

6 · Connect the API (Ultra and up)

Generate an API key under Settings → API. One endpoint, model chosen by parameter — drop-in compatible with OpenAI's format, so existing code changes to a base URL and a model name:

POST https://api.sandbox-ai.tech/v1/chat with "model": "claude-opus" — or "auto" to let routing pick the cheapest model that passes your quality bar. Webhooks deliver agent completions to your systems.

Trial timeline — what happens when

DayWhat happens
Day 0Full Pro access + 300 credits. No card, no email, no signup.
Day 3Trial ends. A payment link appears inside the app — pick a plan there, or ignore it and let it lapse. No dark patterns, no emails sent.
After Day 3 (no action)Workspace pauses; nothing is charged; your history waits 90 days.
Any time laterOpen the app, use the payment link, and everything resumes exactly where it stopped.

Troubleshooting

  • "Model busy" — shared lanes saturate at peak (rare, Pro+ has priority GPU). Retry, or route the prompt to the same model's Flash/Small variant.
  • Slow video renders — Runway queues are longest Monday mornings EU time; storyboard first so renders happen once.
  • Agent stopped at approval gate — that's by design: expensive steps wait for you. Check the gate in the Agents tab.
  • Imported custom GPT behaves differently — presets run on the model you choose; some instructions behave differently across models. The comparison mode shows you exactly where.

Still reading? You could be comparing models already.

Install takes 30 seconds. The first comparison takes 10.

Download for Windows

Next steps

More guides