Warp

The terminal is free; the agent’s model calls are not. The plans page says the Free plan includes no bundled usage for the Warp Agent, so you either buy usage, upgrade or bring your own inference. Ordinary shell commands and non-AI terminal features never draw on agent usage, according to the usage docs. Other AI features, such as Generate and AI Autofill in Workflows, do.

Buying usage on Free costs more than on a paid plan. The pricing page sells Free-plan usage at $12 for $10, $24 for $20, $60 for $50 and $120 for $100, while paid plans buy the same amounts at face value. Purchased usage rolls over for 12 months and stays available after moving back to Free.

In dollars for most self-serve accounts, after a change from credits. The usage docs say a plan bills either a US dollar balance or credits, and Enterprise contracts that use credits keep their terms. The pricing page still mixes both: the Build card promises $20 of usage while the comparison table lists $20 as 1,500 credits, and the Max card says $200 while the table says $240, or 18,000 credits. Your own Settings, Billing and usage screen shows the allowance that applies to your seat.

A run can draw on three parts. Inference is the model call, priced for paid self-serve plans at public API rates. Compute is the Warp-hosted sandbox for cloud runs. Platform usage bills agent hours on cloud runs, excluding queued and idle time; for the Warp Agent it counts time spent generating responses, while Claude Code and Codex runs count their whole active runtime. Below each response Warp shows the turn’s cost, with input, output and cache tokens where available.

The BYOK docs allow your own OpenAI, Anthropic or Google key on Free and every eligible paid plan, but only for individuals and organizations with 10 or fewer employees; larger companies need Business or Enterprise. Keys are stored on your device, and your prompt, context and key pass through Warp’s servers to the provider; Warp says it does not retain the key. Personal keys work for local runs, not cloud runs.

Two details decide whether the key actually saves money. The Auto models always use Warp-provided inference, so only a specific model marked with a key icon bills your provider. And on Business and Enterprise, local runs with your own inference add a platform charge. A ChatGPT or SuperGrok subscription can be connected, but a Claude consumer subscription cannot; if you pay for Claude and want its own terminal agent, the Claude profile covers that route.

With “Help improve Warp” switched on, the privacy docs say Warp may collect high-level usage events and the AI interactions and console inputs behind its AI features. Business and Enterprise are covered by a zero data retention agreement, so no AI interaction or console data is collected there. For model traffic Warp routes to OpenAI, Anthropic, Google or xAI, every plan is covered by zero data retention agreements with those providers; traffic through your own key follows your provider’s policy.

You can switch off telemetry and crash reports under Settings, Privacy, and the docs publish every telemetry event. The client source is public under AGPL v3. The license agreement itself is broader: it lets Warp use licensee data to improve the product, including training its algorithms and models, which sits alongside the narrower privacy-docs wording.

The refund policy covers mistaken sign-ups. A monthly plan is refunded in full if you cancel within 24 hours of the charge and used nothing; otherwise the plan runs to the end of the cycle. An annual plan is refunded in full within 15 days with no usage, and after that for the remaining full months, excluding the first. A usage purchase is refundable only if none of it was used. Refunds for a product defect need the conversation ID, logs and “Help improve Warp” enabled.

The Terms of Service, from Denver Technologies, Inc. doing business as Warp.dev, describe payments as non-cancelable and non-refundable except for a material breach by Warp. Warp also offers no guaranteed uptime agreement; incidents appear on its status page.

1. Install Warp, leave Settings, Privacy at your preferred telemetry choice, and add your Anthropic, OpenAI or Google key under API keys.
2. Open a small repository and pick a specific model with the key icon, not Auto.
3. Ask the agent to fix one failing test, approve each command, and read the diff before accepting it.
4. Expand the usage under the response and note input, output and cache tokens and the cost.
5. Repeat the task on Auto and compare what was drawn from Warp usage rather than your provider.

Accept if the test passes, no command ran without approval, and the per-task cost fits your monthly volume. Reject if the agent changes unrelated files or the cost on Auto is not visible before you top up. No result is recorded here.