Why Token-Based Billing Broke Our AI Budget — And the Guardrails We Put in Place

TL;DR: GitHub Copilot’s silent migration to token-based billing turned a predictable $40/mo line item into $800/mo. We built usage dashboards, hard caps, and local fallbacks — full comparison.

The Context

4-person team, 2,000+ completions/day. GitHub announced token-based pricing in May; our org was migrated in June without opt-in. No email, no dashboard, no budget alerts. First invoice: $800 (27x budget). Support response: “usage is available in organization settings” — but only shows aggregate, not per-seat or per-repo.

What We Tested

Tool	Use Case	Verdict	Why
GitHub Copilot (token)	Daily coding, no controls	❌	Uncapped; no per-seat limits; no real-time usage API
Cursor Pro (token)	Same, with slightly better UI	❌	Shows usage in dashboard but no hard caps; still $0.03/1K tokens
Continue.dev + local models	Unlimited completions, zero marginal cost	✅	Runs on dev machines; no network calls for simple tasks
LiteLLM proxy + budgets	Hard caps, per-key limits, alerts	✅	YAML config: budget: 50 per key; auto-rejects over-limit requests
Custom usage tracker	Per-repo, per-dev, per-model visibility	✅	SQLite logger + Grafana dashboard; 200 LOC; catches anomalies in hours

The Pivot Point

A dev opened a 50-file monorepo in Cursor, triggered “explain codebase” — 2.3M tokens in 10 minutes ($69). Same query on local Ollama: $0. We realized any tool without hard caps is a financial risk.

What We Use Now

Three-layer guardrails:

1. Hard cap: LiteLLM proxy enforces $50/mo per dev key — auto-fails over to local
2. Visibility: Custom tracker logs every call (model, tokens, cost, repo, dev) to Grafana
3. Culture: “Local first” default in Continue.dev config; cloud models require explicit @cloud tag

When You’d Choose Differently

• Enterprise with negotiated contracts: Copilot Enterprise includes usage governance
• Teams without infra for local models: Accept the risk or use Cursor’s slightly better dashboard
• Low-volume usage (<500 completions/day/dev): Token billing may still be cheaper than seat licenses

Tool Crucible Rating

Overall	Ease	Value	Support
3.8/5	3.0/5	4.5/5	2.5/5

---

This is part of our AI cost management series. See full comparison: Token-Based Billing Guardrails