Frontier Tuning billing and cost planning

Cost planning for Frontier Tuning connects the work you intend to run with measured usage and applicable billing evidence. As an agent maker, you can use job measurements to compare improvements and discuss spending with your administrator. Those measurements alone don't establish charges.

Distinguish usage from billing

Keep three evidence scopes separate when you review an experiment.

Evidence What it helps you understand Boundary
Job results and diagnostics. Available execution measurements, scores, and grading coverage for a selected job. Not environment-wide totals or an invoice.
Organizational credit reporting for supported services. Consumption by service, user, or spending policy. Credits aren't currency; reporting scope and timing differ from billing.
Applicable commercial terms and billing records. Approved units, rates, included work, and billed amounts. Job token counts alone don't establish these terms.

The Copilot Credits billing overview explains the common framework for eligible services. The dashboard and billing comparison explains operational versus billing views; neither establishes a Frontier Tuning report mapping. Keep GitHub Copilot host usage separate from tuning service usage.

Understand what changes evaluation usage

In the public v1 CLI reference, an unscoped evaluation uses every Evaluation sample saved in the tuning environment when the job starts, regardless of skill. Skill filters, explicit sample IDs, and sample limits can narrow the work. A sample limit applies after skill filtering and can't be combined with explicit sample IDs.

Selected models are candidates within one combined evaluation by default. Separate per-model jobs evaluate the same selected samples for model comparisons. More samples, repeats, or separate jobs add work; they don't establish a fixed price per sample or model.

BestOfN and TreeSearch can use more tokens and time without improving quality. Larger evaluation-criteria sets can take longer to grade. Measure the changed configuration rather than assuming a universal price effect. For improvement choices, see Choose a tuning flow.

Interpret a job's reported results

Inspect actual graded coverage alongside scores. A successful evaluation can contain failed or ungraded submissions. A zero score is measured; an absent score isn't zero. Missing diagnostic counters mean coverage is unknown, not complete.

The following commands show per-sample results and compare two jobs. Use job identifiers from the intended tuning environment:

frontier-tuning evaluate results <job-id> --samples
frontier-tuning evaluate compare <job-a> <job-b>

Compare matched samples and content, instructions, criteria, model/version, strategy settings, and tool/source state, except for the variable you're testing. Check that the same samples were actually graded. If criteria change, re-evaluate the baseline under the revised criteria.

Available token totals can include grading and orchestration, not just target-model usage. Measured totals include failed submissions even when comparison scores exclude failed or incomplete submissions. Missing usage is unmeasured, not free; partially metered jobs aren't eligible for the comparison's lowest-token summary. These commands don't display prices.

Plan repeated and shared work

Separate improvement runs from consumer traffic over your planning period. Training duration and model availability duration, where applicable, are different assumptions. Record repeated evaluations, retries, and shared resources so that you don't count shared training again for every consumer.

Use the admin guide's cost categories and spending assumptions rather than adding presumed charges for every category.

Separate spending controls from job limits

For supported services, administrators use spending policies for monthly limits, optional per-user limits, and alerts. Policies don't reserve credits or grant environment or connected-data access. Another policy can remain applicable after policy deletion; a policy change isn't a universal off switch.

Concurrent-job limits instead concern job submission. The public CLI documents user or tenant job-limit refusals and service recovery guidance. Inspect existing jobs and follow that guidance rather than repeatedly submitting work.

See the admin guide for eligibility and containment boundaries.

Avoid unnecessary work

Agree on scope and possible spend before generation, refinement, evaluation, or training. Synthetic sample generation is asynchronous and can incur cost where enabled. A small initial check can expose operational problems, but a one-sample smoke test doesn't validate quality.

Apply intended instruction, criteria, and sample changes before training. Later edits don't retrain an already-produced model. Check evaluation readiness separately from training-job success, then evaluate before deciding on another run. Inspect recorded jobs after an uncertain submission before retrying.