Every user_initiated_interaction_count in your export is one Copilot
request, billed in Premium Request Units at the model's PRU multiplier
× $0.04. The same interaction also burns tokens — a quantity that
never appears in your bill but is exactly what you'd pay if you called the model's
provider API directly. This page reconciles the two, model by model, workflow by
workflow, day by day, using your data and the published token rates.
This page reads from a usage export you've already uploaded. Head to the upload page first.
For every chat / agent / CLI interaction in your file we look up the model's PRU multiplier and its provider list-price token rate, multiply by the assumed token profile for the workflow it ran in, and total it. PRU charges are exact. Token cost is a directional estimate — adjust the workflow profile assumptions below to refine.
Token cost depends entirely on how big each prompt is. The Copilot export tells us which feature ran every interaction (chat, agent mode, plan mode, CLI…) but not how many tokens each one used. Pick a token profile for each feature found in your file. Defaults follow the analysis: chat = 4K in / 2K out, agent mode = 60K / 15K, CLI = 100K / 25K. Changes recompute every chart and KPI on this page in real time.
Each row is a model that appears in your totals_by_model_feature. PRU
cost is the multiplier × interactions × $0.04. Token cost uses the workflow profiles
above and the model's list-price input/output rates. Sorted by total interactions.
| Model | Provider | Mult | Tier | Interactions | PRU charged | PRU $ | Est. tokens (M) | Token $ | Δ (PRU − Tok) |
|---|
Same data, sliced by the Copilot feature instead of the model. Reveals where the PRU↔token gap actually opens up: a chat-mode interaction looks fairly priced; agent-mode and CLI start to subsidise heavily on premium models.
Two stacked area charts on the same x-axis. Top: what you actually got billed in PRU each day. Bottom: what those exact same interactions would have cost in tokens. Same days, same models, two pricing models.
Where would your token spend actually go? This section assumes you skipped Copilot entirely and called every model's provider API directly at list price. It is upper-bound — real direct-API customers negotiate rates and Anthropic / OpenAI both offer prompt caching that knocks ~30–50% off input cost on agent workflows. Use it as "the worst case if you went DIY".
| Provider | Models you used | Interactions | Input tokens | Output tokens | Input $ | Output $ | Total direct $ | vs PRU paid |
|---|
On June 1, 2026 Copilot Business and Enterprise pool every seat's monthly AI Credits allowance into one shared bucket. This panel projects your uploaded month's token cost against that pool. Seat count defaults to the number of distinct active users in your upload; orgs running a hybrid of Business and Enterprise can rebalance the plan mix below — the pool size, subscription cost, and per-seat averages update live. Pro / Pro+ are excluded — those plans are not pooled.
A short read of the data above — generated from your numbers, not generic advice.
You now know what your fleet would cost in tokens. The optimisation page takes the same upload and answers the strategic questions: what's the right Business / Enterprise mix, which model swaps recover the most credits, how many dormant seats are quietly funding the pool, and what does the next 12 months look like under conservative vs aggressive plays.
Open UBB optimisation insights →Personas · plan-mix sweep · substitution waterfall · scenario sliders · 12-month forecast.