Is T3 Code free? Yes, and here is the bill it hides

The short answer

T3 Code costs nothing. It is released under the MIT licence, there is no paid tier as of August 2026, and the project has said directly that it charges nothing because it built the tool to have the best possible development experience with agents. There is no trial, no seat, no credit balance, and no usage meter belonging to T3 Code, because T3 Code performs no inference and never handles a token.

The bill T3 Code assumes you are already paying

T3 Code drives five providers and installs none of them. Every thread you open is a thread on somebody’s subscription. So the real question is not what T3 Code costs, it is what a bring-your-own-key agent stack costs, and that number moves by an order of magnitude depending on how hard you run it.

Two structural facts change the arithmetic more than any individual price. First, both Claude and ChatGPT meter subscription agent work with a rolling five-hour window and a weekly cap running at once, so a heavy Monday still costs you on Thursday, and the way you hit a wall is usually the weekly cap rather than the session one. Second, running agents in parallel is exactly the workload that exhausts those windows fastest, and parallel work is the reason to use a control plane at all. The tool that makes you productive is the tool that makes you upgrade.

Three realistic monthly configurations

Nobody runs all five providers on their top plan. These are the three shapes people actually land on, priced from the vendor pages.

The daily-driver row is the one most parallel-agent users converge on, and it is worth pausing there. At $120 a month, the free control plane is 0 percent of the cost and the two subscriptions are 100 percent of it. Any conversation about the price of an agent workbench that stops at the workbench has skipped the entire bill.

Two more line items rarely make it into the spreadsheet. If you run npx t3 serve on a machine that exists to host agents, that box has a monthly cost, and a cloud instance sized for a real build is not cheap. And if you use OpenCode or Grok against raw API keys rather than a subscription, that spend is metered with no ceiling, which is a different risk shape from a plan that simply stops.

What T3 Code can and cannot tell you about that money

This is the part that changed recently and is worth stating precisely, because an older account of T3 Code will tell you it has no cost surface at all. It does now. The Usage page aggregates activity from connected environments and shows API-equivalent token cost, processed tokens, cache savings, provider shares, and model breakdowns, across the past 24 hours at hourly granularity plus 7, 30, and 90 day ranges at daily resolution.

What it cannot do is the part its own documentation is careful about. The figures are reconstructed by scanning local session history and pricing it at API rates. The docs say directly that subscription billing is separate from the raw token cost shown. So the number answers "what would this have cost on the API" and not "what did I spend", and for a subscription user those two questions have very different answers: the marginal turn on a plan you already bought costs nothing until the window closes, at which point it costs your entire afternoon.

None of this is a criticism of the implementation. It is the same architectural consequence as everything else on this page: a control plane that never touches a token cannot see a bill, and one with no account model cannot attribute spend to a person. The estimate is a genuinely useful thing to have added, and it is honest about being an estimate.

The one thing that might not stay free

The clients are MIT and cost nothing to run. T3 Connect, the hosted tunnel that makes a machine reachable from the mobile and web clients, is different: it is somebody else’s infrastructure carrying your traffic, funded by nothing.

The project has been unusually straight about this. T3 Connect is free, every user can link up to three devices, and the stated position at launch was that they do not want to charge for it but may have to change course if the bill becomes unacceptable. The escape hatches were named in the same breath: fork it, self-host it, or use Tailscale instead.

The alternative to buying five subscriptions

There is a second way to pay for agent work that the bring-your-own-key model rules out by design, and it is worth understanding even if you stay on T3 Code: buying inference from the workbench itself, with no provider subscription underneath.

Continuum sells this as Plus. You pay one price, get a key, and run Claude Code or any compatible client against it without holding a Claude or ChatGPT plan at all. That is a materially different commercial shape from the T3 Code stack: one vendor, one invoice, one cap, and no five-hour window belonging to somebody else. The workbench itself stays free either way, and bring-your-own-key still works, so this is an addition rather than a toll.

The comparison that matters is not $0 against $25. It is $25 against the $120 the daily-driver row above costs, for the case where you would otherwise buy two provider subscriptions purely so an agent has something to call. If you already hold those subscriptions for other reasons, bring-your-own-key remains cheaper and T3 Code serves it well. If you are buying them only to feed an agent, a hosted plan removes a layer of the bill and gives you one cap to reason about.

Which arrangement is cheaper for you

Run the arithmetic on your own last month before changing anything. Count the subscriptions you hold, mark which ones exist only to feed an agent, and check how often you actually hit a weekly cap. Most people discover one of two things: either they are paying twice for capacity they never exhaust, or they are on an entry plan that stops them every Thursday. Those two findings point in opposite directions, and no pricing page can tell you which one you are. The team cost calculator does the multi-seat version of the same sum.

Questions people ask

Keep reading

Continue through Claude Code reference with T3 Code on iOS and Android: what the mobile app does.