使用レポートにモデル別トークン内訳が追加Per-model token breakdown in the usage report

GitHub has expanded the usage report for its Copilot and AI-powered features to include a per-model breakdown of the tokens that sit behind AI credit consumption. According to the company's changelog, the report now shows, for each model, how many input, output, and cache tokens were used. For teams and individuals trying to understand where their AI spending goes, this added granularity makes it easier to track consumption and manage costs across the different models available in the platform.

The change matters because modern AI billing is increasingly tied to token usage rather than a simple flat fee. A token is a unit of text that language models process, roughly corresponding to a fragment of a word, and the number of tokens consumed depends on both the length of a prompt and the length of the generated response. Until now, users could see aggregate AI credit usage, but attributing that consumption to specific models required guesswork. By splitting the totals across each model an organization uses, the updated report gives a clearer picture of which models are driving costs and how heavily each is being exercised.

The breakdown into input, output, and cache categories is a meaningful detail. Input tokens represent the text sent to the model, including prompts, code context, and any supporting instructions. Output tokens represent the text the model generates in response. Cache tokens typically refer to cached input, a mechanism where repeated or shared context can be stored and reused to reduce processing overhead. Because providers often price these categories differently, with output tokens frequently costing more than input tokens and cached tokens usually costing less, separating them helps users reason about the actual economics of their usage rather than looking at a single blended figure.

This update fits into GitHub's broader shift toward a credits-based and request-based model for its AI offerings. Copilot plans have moved to structures that include allowances of what GitHub calls premium requests, with usage beyond those allowances drawing on paid credits. As GitHub has broadened the range of models accessible through Copilot, including offerings associated with providers such as OpenAI, Anthropic, and Google, the ability to attribute spending to a specific model has become more relevant. Different models carry different capabilities and different costs, so an organization standardizing on one model for routine tasks and another for more demanding work benefits from seeing those costs itemized.

For administrators, the practical value is likely to show up in budgeting and governance. Engineering leaders can use the per-model data to identify whether expensive models are being used for tasks that a cheaper model could handle, to set internal guidance on model selection, or to forecast future spending based on observed patterns. The visibility also supports chargeback and showback scenarios, where costs are allocated back to individual teams or projects. This mirrors a wider industry trend, as cloud and AI vendors respond to customer demand for cost transparency by adding finer-grained reporting to their consumption dashboards.

The feature is consistent with tooling that already exists elsewhere in the ecosystem. Cloud providers have long offered detailed cost breakdowns by service and resource, and AI platform vendors have added token-level metering to their own consoles and APIs. GitHub's move brings comparable detail into the usage report that Copilot customers already consult, reducing the need to reconcile data across separate systems. It appears to be an incremental but useful refinement rather than a new product, focused on reporting rather than on changing how credits are calculated or priced.

For readers evaluating the change, a few points are worth keeping in mind. The breakdown reflects how the underlying models are metered, so understanding token behavior, including how cache tokens are counted, will help in interpreting the numbers accurately. Organizations that have not yet examined their model mix may find the report surfaces usage they had not fully accounted for. As with any usage reporting, the data is most valuable when reviewed regularly and paired with policies on model selection and credit limits. GitHub has published the details on The GitHub Blog, and the per-model token breakdown is presented as generally available within the existing usage report rather than as a preview or opt-in capability.