The reverse-gas model

The documentation groups the charges into categories. Compute covers executing instructions, including update calls, timers and heartbeats. Storage covers both WebAssembly heap memory and stable memory, charged per byte per second. Messaging covers ingress messages from users, inter-canister calls and their responses. Threshold cryptography covers ECDSA and Schnorr signing and key derivation. External integrations cover HTTPS outcalls and the chain-specific services such as the Bitcoin integration.

Computation is charged as instructions execute. The network counts the WebAssembly instructions processed while handling a message, and there is an upper bound on instructions per consensus round. A message that exceeds the limit is paused and resumes in the next round, with the cycles consumed each round charged at round end. This is the mechanism behind deterministic time slicing, and it is why a long-running computation does not fail but instead spreads across rounds.

Query calls are the exception. A query is answered by a single replica without going through consensus, and the documentation lists query calls as free. That is a meaningful asymmetry: reading data costs the canister nothing, while writing it costs cycles. The cycles page covers the accounting unit itself, including how it is obtained and what happens when a balance runs low.

By default a canister is scheduled best-effort: the subnet runs it when capacity is available. That is adequate for most applications and costs nothing when the canister is idle. A canister that needs a guarantee can set a compute allocation, expressed as a percentage of one execution core, which reserves execution slots on a fixed schedule.

The documentation gives the schedule in terms of rounds: a one percent allocation is scheduled at least every hundred rounds, two percent at least every fifty, and a hundred percent every round. The total allocatable compute capacity per subnet is bounded, so allocation is a scarce resource rather than a free setting.

Allocation is a rental, not a usage charge. The fee accrues with time and allocation percentage regardless of whether the canister actually executes, which increases idle consumption. That is the trade-off: a canister that reserves compute pays for the reservation even when it has nothing to do, and a canister that does not reserve compute may wait under load.

An ingress message is a call from outside the network — a browser, a command-line tool, another service. The canister that receives it pays for its induction and execution. An inter-canister call is a message from one canister to another, and it is charged to the sender: the request, the response, and the instructions on both sides. A canister that fans out to many others therefore pays for the fan-out, which makes call topology a cost decision as well as an architectural one.

Some operations require cycles to be attached explicitly rather than deducted from a balance. The documentation lists canister creation, threshold signing, key derivation and HTTPS outcalls among the methods that fail if insufficient cycles are provided. In Rust the CDK attaches the required amount automatically; in Motoko the developer attaches it. The HTTP outcalls page covers the outcall case, where the cost depends on the request size and the maximum response size.

The honest summary is that the reverse-gas model moves the funding question from the user to the developer, and it does not make computation free. A canister that is popular costs its operator more than one that is not, and a canister that runs out of cycles stops serving update calls. The model changes who pays and when, not whether anything is paid.