August 2026 headline
All Copilot activity sits on the tke enterprise. The second enterprise, tkelevator, has zero Copilot seats and is excluded.
Usage is growing, and deepening
Monthly active developers grew from 211 in March to 364 in August. More importantly, the share of users in agent-first phases (2 and 3) went from 0% to 7.9%.
Active developers vs. licensed seats
Monthly active (28d) and weekly active (7d rolling average), March – August 2026.
Adoption phases
453 engaged users, Aug 5 – Sep 1. Phase reflects how deeply a developer uses Copilot, from passive completions to agent-first workflows.
Average lines of code added per developer per day, by phase
Phase 2 developers ship 3.9× more code than Phase 1
438.5 lines/day versus 112.3. Only 33 of 453 engaged developers (7.3%) have reached agent-first working patterns. Moving another 10% of the Phase 1 population into Phase 2 is the single highest-leverage adoption action available.
Where the AI units go
August 2026, 1,874,466 AI units. VS Code chat dominates, but CLI and agentic surfaces now account for 12% of consumption.
By integration
Share of August AI units
By model
Top models, August AI units and usage value
| Model | AI units | Value |
|---|---|---|
| gpt-5.4 | 353,554 | $3,536 |
| claude-opus-4.8 | 281,357 | $2,814 |
| gpt-5.6-terra | 201,478 | $2,015 |
| gpt-5.6-sol | 189,575 | $1,896 |
| gpt-5.5 | 177,985 | $1,780 |
| claude-sonnet-5 | 155,888 | $1,559 |
| gpt-5.3-codex | 153,704 | $1,537 |
| claude-sonnet-4.6 | 144,950 | $1,450 |
| gpt-5.6-luna | 64,254 | $643 |
| claude-sonnet-4.5 | 34,230 | $342 |
Healthy model mix
No single premium model runs away with the budget. The most expensive model is 19% of spend, and mid-tier models (Sonnet, Codex, Terra) carry a large share. This is a well-balanced profile.
What it costs today
Gross usage is the total value of AI consumed. Billable is what remains after the included entitlement pools are applied. The gap between them is the promotional pool that expired on 1 September 2026.
| Month | AI units | Gross usage | Billable | Covered by pools | Avg daily users | Peak |
|---|---|---|---|---|---|---|
| June 2026 | 2,318,830 | $23,188 | $10,071 | $13,117 | 139.9 | 230 |
| July 2026 | 1,824,944 | $18,249 | $2,911 | $15,338 | 139.8 | 230 |
| August 2026 | 1,874,466 | $18,745 | $1,915 | $16,830 | 132.8 | 206 |
Promotional entitlement pool expired 1 September 2026
Through August, TK Elevator had a promotional pool of $16,770/month on top of the standard pool of $10,640/month. That is why August billed only $1,915 against $18,745 of gross usage. From September the promotional pool is gone and only the standard pool (551 seats × 1,900 credits = $10,469/month) applies. At August consumption levels, the metered line moves from $1,915 to roughly $8,276/month — a 4.3× increase. This is a change in what is invoiced, not a change in behaviour.
Money spent vs. money returned
A deliberately simple model. Costs are measured. Benefits are derived from a single lever, developer time saved, so the customer can argue with one number rather than ten.
Expenses · run-rate
Copilot only. GitHub Enterprise licensing is excluded.
- Copilot Business · 551 seats × $19$10,469 / mo
- Metered AI, August billed$1,915 / mo
- August actual$12,384 / mo
- Copilot Business · 551 seats$10,469 / mo
- Metered AI, post-promo estimate$8,276 / mo
- Forward run-rate$18,745 / mo
Value returned · time saved
364 monthly active developers at a fully-loaded cost of $130,000/year each.
| Time saved | Hours / mo | Value / mo | Value / yr |
|---|---|---|---|
| 30 min / dev / week | 788 | $53,357 | $640,284 |
| 1 h / dev / week | 1,576 | $106,714 | $1,280,568 |
| 2 h / dev / week | 3,152 | $213,428 | $2,561,136 |
| 3 h / dev / week | 4,728 | $320,142 | $3,841,704 |
value = hours × $67.71/h (= $130,000 ÷ 12 ÷ 160 h)
How the break-even is calculated
$34.02 ÷ $67.71 per developer-hour = 0.50 hours = 30 minutes / dev / month
As a share of a 160-hour month: 0.31% of engineering capacity
At the forward run-rate of $18,745: $51.50 per dev ÷ $67.71 = 46 minutes / dev / month (0.48%)
Put differently: if Copilot saves each active developer half an hour a month, the platform is free. Every minute beyond that is net return. Independent field studies place realistic savings in the range of 1 to 3 hours per developer per week, which is 8 to 26 times the break-even threshold.
Net position at 1 h/dev/week
- Value returned+$106,714 / mo
- Copilot cost, forward run-rate−$18,745 / mo
- Net monthly benefit$87,969 / mo
- Net annual benefit$1,055,628 / yr
Cost as a share of payroll
GitHub's own ROI framing, taken from the impact API for this enterprise.
Even the heaviest agent-first users cost 1.62% of their own payroll. The lightest cost 0.79%. Both sit far below any plausible productivity gain.
Assumptions, stated plainly
- $130,000 fully-loaded annual developer cost. This is GitHub's default modelling figure, not a TK Elevator figure. Substitute the real number and every value scales linearly.
- 160 productive hours per month, 4.33 weeks per month.
- 364 active developers, the measured 28-day active count. The 187 idle seats are counted as cost but contribute zero benefit, so the model is conservative.
- USD throughout. Consumption data is billed in USD; no EUR conversion is applied.
- Time saved is not measured in this data. It is the one input the customer must supply or accept from benchmark. Everything else on this page is observed.
Three things worth doing
1 Reclaim 187 idle seats
551 seats, 364 active. The 187 seats with no activity in 28 days cost $3,553/month, or $42,636 a year.
Either reassign them to teams on the waiting list or release them at the next true-up. Idle seats also drag down every per-seat metric used to justify the platform internally.
2 Move Phase 1 into Phase 2
315 developers sit in Phase 1 at 112 LoC/day. The 27 in Phase 2 run at 438 LoC/day.
This is an enablement problem, not a licensing one. Agent mode, Copilot CLI and PR review are already available on their seats and barely used — copilot-pr-reviews is 347 AI units out of 1.87 million.
3 Brief finance before the September invoice lands
The metered line moves from $1,915 to roughly $8,276 with no change in developer behaviour, purely because the promotional pool expired. If nobody has flagged this, it will read as a 4× cost overrun. The enterprise budget is $30,000/month with PreventFurtherUsage, so there is no service risk, but there is a credibility risk.
Data gap: pull request telemetry is empty
Every PR metric returns zero — pull requests merged, median time to merge, code-shipped multiplier. Without it there is no cycle-time or throughput story, which is the evidence a CFO will ask for next. Worth confirming whether TK Elevator's repositories sit inside the tke EMU enterprise or whether this is a telemetry gap to raise internally.