Session volume
Token cost per session
Tool calls (e.g. web search)
—tokens / month
—tool calls / month
—total / month
Cost by scale
Runtime cost scales with wall-clock session length regardless of how much the model did; token cost scales with how much it actually read and wrote. Both grow linearly with session count, but a slow agent pays disproportionately more runtime per session.
| Sessions/mo | Runtime | Tokens | Tools | Total |
|---|
Why session-hour billing is a separate line item
Hosted agent platforms increasingly bill the runtime environment itself — the sandbox, browser context, or execution container the agent runs inside — on top of standard model token rates. A representative 2026 rate is roughly $0.08 per active-runtime-hour. This isn't a replacement for token pricing, it's additive: a long agent session pays for both the wall-clock time its environment stays provisioned and every token the underlying model reads or writes while that environment is alive. Two agents doing the identical task can rack up very different bills if one finishes in 90 seconds and the other takes 8 minutes, even with near-identical token counts, because the slower one occupies billed runtime for longer.
This is exactly the gap between this calculator and a pure per-step or per-token agent cost tool: those price the model calls, this adds the dimension that scales with time-in-session rather than work-done. It matters most for agents that wait on slow tool calls, browse pages, or run long tool-use loops — time the model spends idle inside an active session still burns session-hour billing even though it burns very few tokens during that wait.
Layered on top of both is metered tool usage — a common 2026 rate for a built-in web search tool is about $10 per 1,000 searches. An agent that searches liberally can accumulate meaningful tool cost independent of both its runtime and its token bill. Total agent cost in 2026 is really three stacked meters: runtime-hours, tokens, and tool calls — treating any one of them as "the" cost understates the real bill. For the pure per-step/token view without runtime billing, see the AI agent cost calculator or the agent loop budget calculator for P99 runaway-cost planning.
How this calculator works
Free managed agent session cost calculator — models the 2026 session-hour billing dimension (e.g. $0.08/active-runtime-hour) on top of standard input/output token rates and tool costs like web search, so you can see the true cost of a hosted agent run, not just its tokens.
Frequently asked questions
What is session-hour billing for AI agents?
Some managed/hosted agent runtimes now bill a per-hour charge for the active runtime environment itself -- the sandbox, browser, or execution context the agent operates in -- separately from the token cost of the model calls it makes. A representative 2026 rate is roughly $0.08 per session-hour of active runtime. This is additive to standard input/output token pricing, not a replacement for it: a long-running agent session pays for both the wall-clock time its environment stays alive and every token the model reads or writes during that time.
How is this different from a per-step or per-token agent cost calculator?
Step-based and token-based agent calculators price the model calls: how many turns, how many input/output tokens per turn. This calculator adds the dimension those miss -- the runtime environment's own hourly charge, which scales with wall-clock session duration, not with how much the model actually did during that time. A slow agent that takes 20 minutes to finish a task pays for 20 minutes of session-hour billing even if it made very few model calls in that window; a fast agent finishing in 2 minutes on the same task pays far less runtime cost regardless of token count. Both dimensions matter and this tool models them together.
What other per-use costs do managed agents typically add?
Beyond runtime-hour and token cost, hosted agent platforms commonly meter individual tool calls separately -- a common 2026 rate for a built-in web search tool is about $10 per 1,000 searches. If your agent's task involves browsing, searching, or calling other metered tools, those per-call costs stack on top of runtime and token cost the same way a cloud bill stacks compute, storage, and egress as separate line items.