GitHub Copilot Now Reports Tokens by Model: Finally Explain Input, Output, Cache, and AI Credits

GitHub Copilot administrators can now see which models generated their AI credits and how many input, output, cache-read, and cache-write tokens sit underneath the charge. That turns an opaque credit total into data a FinOps or engineering team can actually diagnose.

GitHub announced the per-model token breakdown on August 11. The data is available in the downloadable AI usage report to Business and Enterprise admins and to individual Copilot users. Before the change, teams could see AI credits without the token detail needed to explain a charge or find a specific optimization.

Four Token Columns, Four Different Stories

Input tokens show the context sent to a model: prompts, files, repository instructions, and conversation history. A surge often points to long sessions, oversized context, or broad repository selection. Output tokens show what the model generated. High output can come from verbose explanations, full-file rewrites, or agents narrating every step.

Cache-read tokens represent reused context and can signal healthy repetition when priced below uncached input. Cache-write tokens create the reusable state and may carry a different rate. A high write number with low reads suggests prompts are changing too often, sessions end before reuse, or model routing breaks cache affinity.

Start with Attribution, Not Blame

Export the report on a regular cadence and join it to the fields GitHub provides for user, organization, product, or cost center. First establish a baseline by model and workflow. Do not rank developers by raw spend: one engineer may be migrating a legacy subsystem while another answers short questions. Cost without outcome invites harmful incentives, such as avoiding valuable agent work or splitting activity across untracked tools.

Instead, investigate changes. If model A's input tokens rose 70% week over week while completed pull requests stayed flat, sample sessions and check context size. If cache reads fall immediately after a client rollout, the update may have changed prompt prefixes. If a premium model dominates routine tasks, adjust the team default or routing policy.

Useful Ratios to Calculate

• Output-to-input ratio: flags unusually verbose generation or extremely context-heavy tasks.
• Cache-read share: cache reads divided by all input-side tokens indicates how much context is being reused.
• Credits per accepted PR: connects billing to a reviewed outcome rather than activity.
• Credits per active user: helps forecast allowance needs, but should be segmented by role and workload.
• Failure spend: credits attached to abandoned sessions or reverted changes exposes retry waste.

Three Optimizations the Report Can Validate

First, narrow context. Exclude generated files, large logs, and irrelevant packages, then confirm input tokens fall without reducing acceptance. Second, route tasks. Put explanation, formatting, and simple edits on an efficient model while preserving a stronger tier for complex debugging. Third, stabilize reusable prefixes and instructions so cache reads rise. Each change needs an outcome guardrail; cheaper tokens are a loss if review time or defect rates increase.

Run changes as controlled trials over comparable repositories or task types. Record the model, token mix, credits, completion, and reviewer time. A token report can show that consumption changed, but only engineering outcomes show whether the change was beneficial.

Build a Weekly Review That Leads to Action

Store each export in a versioned analytics location and create a weekly view by model and workflow. Highlight changes outside a normal band, then assign an owner to inspect a sample rather than sending a generic cost warning to every developer. Record the resulting action—context exclusion, model routing, prompt change, or no change—and check the next report for the expected effect.

Reconcile the report with billing totals before trusting automation. Time zones, late usage, included credits, and adjustments can produce apparent mismatches. Keep the raw export and transformation version so finance can reproduce a number. Apply access controls as well: user-level usage data supports operations, but broad ranking can create privacy and management problems unrelated to optimization.

Watch for Billing-Model Confusion

GitHub has moved much of Copilot toward usage-based AI credits, while some legacy annual arrangements use request multipliers. Do not apply an old multiplier table to a usage-based invoice. Use the report and documentation for the plan attached to the account. The same model name can have different economic meaning under an included allowance, metered token pricing, or a legacy request quota.

The new breakdown is not an optimization by itself. It is the missing measurement layer that lets teams distinguish big prompts, verbose outputs, weak cache reuse, and expensive model selection. Export it before changing defaults, preserve a baseline, and compare with direct API scenarios using our AI Cost Calculator.