Skip to content

Model Selection & Pricing

Choose the least expensive available model that completes the task reliably. Include retries and review time, not just the rate for one request.

What determines cost

Factor What to check
Model Input, cached-input, and output rates for the actual surface
Reasoning effort Extra reasoning work where the model exposes this control
Context Fresh input versus eligible cached input
Workflow Agent steps, retries, and repeated history
Plan and policy Model availability, discounts, budgets, and organization restrictions

Use the GitHub model comparison, supported models, and billing documentation for current availability and charges. Direct API list prices from Anthropic or OpenAI are not a Copilot bill estimate.

Workflow Optimization covers modes; Context Management covers cache stability and handoffs.

Good default heuristics

Task Starting point Escalate when
Syntax lookup, short explanation, small edit Auto or a fast, inexpensive model The answer is wrong or lacks needed context
Routine implementation with clear acceptance checks A model that has handled comparable work Repeated failures cost more than a stronger model
Cross-file diagnosis or ambiguous planning A stronger reasoning model Verification still exposes unresolved assumptions
Architecture or high-stakes review A capable model plus explicit human review Expert review is needed; model choice is not a substitute
Execution of an approved, concrete plan Try a cheaper model in a fresh session The implementation requires new design decisions

These are starting points, not a guarantee that a model tier fits every task. Measure accepted outcomes on your own work.

Auto selection

Auto selects within the supported pool for the surface. Where supported, task-optimized Auto considers task complexity as well as health and availability. Check its current pool, policy restrictions, and any eligible discount rather than treating it as access to every premium model. See GitHub's Auto-selection documentation.

Reasoning effort

Where effort is configurable, choose it with the task. Use lower effort for straightforward work; increase it when deeper reasoning improves acceptance results enough to justify the extra work. A concise final answer does not show how much reasoning was billed.

Pick model and effort before a long task. If the task changes enough to justify a different configuration, use a short handoff rather than automatically carrying the entire history. Follow the cache workflow.

Policy and data handling

Availability varies by plan, preview status, geography, organization policy, and client. A model in the picker is not automatically approved for regulated work.

Check legacy annual-plan access, regional restrictions, compliance surcharges, and extended-context/effort availability against the current model and pricing rules. Do not apply chat cost assumptions to code completions or next edit suggestions; see Enterprise Governance.

Before routing sensitive work, confirm the applicable provider terms, data-retention controls, and organization policy. GitHub documents model hosting and policy controls. Do not infer the same retention terms for every model, particularly previews. Fable models, for example, retain prompts and outputs by default; eligible enterprises need separate approval and configuration for the time-bound zero-retention exemption.

Judge routing by completed work

Compare cost, elapsed time, acceptance checks, and retries for similar tasks. A cheap request that needs several repair attempts can cost more than one successful stronger-model run.

Keep billing mechanisms separate: model rates, discounts, eligible caching, and session overhead act differently. See Comparisons & Data before combining savings estimates, and Enterprise Governance for organization-level levers.


Next: Enterprise Governance →