GitLab's 2026 Developer Survey found that sprint velocity baselines recalibrated upward by 40% within two quarters of AI tool adoption. Not velocity — the baseline. Management measured the new speed, absorbed it as the expected throughput, and reset the denominator. The old pace became underperformance.
This recalibration happened between March and August 2026 — the exact period when every major AI coding vendor was running promotional pricing.
The promotions expire in thirty-eight days.
The Staircase Down
The subsidy era doesn't end in a cliff. It ends in three steps, each one lowering the floor beneath a baseline that was already ratcheted up.
This is the structural problem: the baseline was set during the cushion. It will not be reset when the cushion is removed.
Already Breaking
The billing shock is not a September prediction. It is a June fact.
When GitHub moved Copilot to usage-based billing on June 1, agentic users reported 10x to 50x cost surges over their old flat-rate plans. One developer on the $39 Pro+ plan burned 8% of monthly credits in two hours — projecting depletion in under two days. Individual reports of bills jumping from $29 to $750 and $50 to $3,000 appeared within the first billing cycle. A journalist documented a projected $180/month bill from a broken editorial workflow on day one.
For the first thirty days of usage-based billing, there was no spending cap. GitHub added opt-in spending limits on July 2 — opt-in, not default. Overage charges accrued silently for a month before the guardrails existed.
That was with the promotional credits still active. September removes 37–44% of the cushion.
The Ratchet
A ratchet turns one direction. When a sprint team delivers 42 story points instead of 32, the sprint retrospective doesn't celebrate — it recalibrates. Next quarter's target becomes 42. The promotional credits made 42 sustainable by subsidizing the agentic workflows that produced the speed. September 1 removes the subsidy. The target stays.
The median all-in cost of an AI coding stack is now $28,000 per developer per year — Copilot or Cursor licenses, ChatGPT or Claude subscriptions, observability tooling, security scanning, vector databases. For a 100-developer organization, that is $400,000–$600,000 annually before governance infrastructure, which runs an additional $50,000–$250,000. One company with 80 developers calculated its monthly AI spend equaled one developer's annual salary.
Thirty percent of engineers surveyed hit usage limits every month. Eighty-two percent of Pro users exhausted their allowance on day one. These numbers are from the promotional period — the period with the highest credit allotments these plans will ever see.
The Escape Valve
The credit cliff has a pressure release, and it is arriving on the same timeline.
DeepSeek V4 reached general availability on July 20. V4-Flash costs $0.14 per million input tokens — roughly 100x cheaper than frontier closed-source models. Kimi K2.7 is already integrated inside Copilot at $0.95 per million tokens. Poolside shipped Laguna S 2.1, a 118-billion-parameter open-weight model beating some closed-source benchmarks. The open-weight tier is no longer a fallback — it is a parallel track running at 10–35x lower cost.
This doesn't resolve the ratchet. An organization that migrates from Copilot's promotional tier to DeepSeek V4 saves money but still faces the baseline problem: management expects 42-point sprints regardless of which model powers them. The tool is cheaper. The expectation is not.
What the escape valve does is change who absorbs the cost. Under promotional credits, the vendor absorbs it. Under open-weight models, the infrastructure team absorbs it. Under neither, the developer absorbs it — in longer hours, in skipped verification, in the cognitive overload BCG calls "AI brain fry" where 14% of workers report acute mental fog from supervising multiple AI systems.
"Speed without control is a liability, not an advantage."
— GitLab, June 2026
The velocity baseline was built on credit. The credit has a maturity date. The baseline does not.