Local seats or a fixed-run demo.
Local runners cost $15 USD per contributor per month, or $180 annually. Your own compute has no Tbench per-minute fee. The AI runner demo is $0.10 per started run, including approved Linux x64 CPU compute up to 2 CPU and 4096 MiB. Each run is bounded to ten minutes including 120 seconds reserved for preparation and cleanup, leaving up to eight minutes for policy steps.
An operator applies the fixed-run policy; your owner or billing manager then accepts it by setting budgets and enabling work. Existing metered accounts retain their approved rate unless explicitly changed. A price change pauses new work for renewed budget acceptance and never reprices existing reservations.
The 15-day approved trial covers local Actions and local verification only. Demo approval does not authorize paid model APIs. Model profiles and execution capacity need separate qualification; larger workloads require a different approved allowance.
Reserve first. Execute second.
An owner or billing manager sets a monthly USD budget and a per-job limit in Usage & budgets. Tbench separately approves hosted access and qualifies devices. New accounts default to disabled with zero budget. Every submission reserves the maximum policy runtime plus 120 seconds for preparation and cleanup. Concurrent submissions share the same organization write fence.
A budget cannot be reduced below spent plus reserved usage. Pausing blocks new claims; running work retains its reservation. There is no automatic credit reset for an unresolved job, including across UTC month boundaries.
A fixed fee. Durable evidence.
The demo fee is reserved before enqueueing. Cancellation before leasing releases it. Once leased, successful, failed and cancelled runs incur the same fixed fee after confirmed cleanup. Execution time is recorded, but CPU minutes are not added to the demo charge. Repeating a submission or completion does not create a second charge for the same job.
Existing minute-based jobs retain their original metered policy. Each usage row records its pricing model and original fixed fee, if applicable. The ledger is usage evidence, not a paid invoice. Billing details and Stripe collection are managed from your organization’s billing workspace.
Expired leases or missing cleanup are marked uncertain and retain the full reservation. Operator reconciliation requires execution and cleanup evidence; a retry never overwrites the failed attempt. Customer spending limits are not an AWS account spending cap; infrastructure approval and shutdown controls remain separate.
Your own pipeline, on approved capacity.
Create an immutable policy with named steps, a pinned image digest, an exact repository and bounded resources. Submit a commit with compute_class: "hosted" through the independent runner API, or choose Hosted in the pipeline console. Ordinary customer devices cannot claim hosted jobs, and hosted devices cannot claim local jobs.
The hosted independent-runner path supports custom, GitLab and Jenkins invocation adapters. Native GitHub runner scheduling remains the existing local-device integration; it is not silently migrated into a hosted fleet. Source preparation and workload isolation must be qualified for each hosted device before approval.
Open the pipeline console