Over the last two years, AI coding assistants stopped being a nice-to-have and became the way our teams work. Then the vendors moved to usage-based billing, and a previously predictable cost started behaving like a rollercoaster. Projected across several hundred developers, we were looking at a 4-5x increase in our AI tooling bill.

Both obvious alternatives were bad. Absorb the cost, and finance would own the engineering AI roadmap. Restrict usage, and we would throw away the momentum our teams so meticulously built. So we looked further for what was actually driving the token usage, expecting a billing problem.

What we found was a working-habits issue. On an identical feature, an undisciplined session burns most of its budget on reasoning tokens the developer never sees. The most expensive sessions were the ones using the largest model on the smallest questions, with contexts growing without restrictions, and developers prompting their way toward a solution instead of driving one. Fixing these habits cut measured token usage by roughly 70% on the same work. Teams doing it reported better output and better work satisfaction, with velocity unchanged.

We'll cover: (1) the three levers that actually drive spend: model choice, conversation size, and interaction count; (2) two changes you can make the same afternoon: matching model capability to task and rewriting your agent instructions; (3) spec-driven development, the workflow change behind most of the savings.

Our developers resisted at first, and they were right to: this asks them to carefully choose the model, keep context tight, and verify output at every step. Nobody got slower, because a disciplined session needs less back and forth than an undisciplined one. The friction was the point.

You'll leave able to estimate your own token spend, identify which habits burn the most for the least return, and run a spec-driven session the next morning.

OUTLINE (45 minutes, including Q&A)

0-3 — The rug pull. Usage-based billing arrives with no meaningful notice. Projected 4-5x across several hundred developers. Two bad options.
3-6 — Shared vocabulary and the three cost dimensions. Context, token, agent, reasoning, output. Establishing that reasoning tokens are billable and invisible.
6-13 — Lever 1: model choice. The 0.33x to 15x spread. What large models are genuinely needed for versus what small models handle fine.
13-23 — Lever 2: conversation size. Input and output both bill. Instruction files as prevention; output compression; the instruction rewrite from 2,500 tokens to 300.
23-33 — Lever 3: interaction count. Vibe coding versus spec-driven on an identical feature, with token comparison and live demo (5-6 min, recorded fallback).
33-37 — What it cost us and what we got. Developer resistance, the 70% result, honest scope of what was and wasn't measured. Frameworks to start with. Close on the friction being the point.
37-45 — Q&A.

TAKEAWAYS (3)

Estimate what your organization's AI assistant usage costs per developer per month, and identify which of the three levers is driving the largest share.
Choose between large and small models per task type, using capability boundaries that hold in practice rather than defaulting to the most capable option.
Run a spec-driven session on a real feature and compare its token consumption against the equivalent unplanned session.