Gartner says your AI coding bill will pass a developer's salary by 2028

The meter is always running. That's the line from Gartner's June 24 prediction, and it's the one that should change how you build. By 2028, Gartner expects the per-developer cost of AI coding agents to exceed the average developer's salary, driven by consumption-based token pricing and agents that burn tokens on every single action.

Most enterprises underestimate this badly. Gartner notes many vendors lack transparency into how tokens are even metered, and recommends governance frameworks, context engineering, and routine token-usage reviews. This isn't abstract. A widely-debated HN essay, "Why Current LLM Costs Are Not Sustainable," hit 107 points and 189 comments arguing today's prices are propped up by VC and hyperscaler cross-subsidies, with providers absorbing 90%+ of the true cost of every token. Per-token prices have fallen ~10x annually since 2021 even as some workloads grew ~100x in consumption. That's the LLM cost paradox: unit prices drop, your bill explodes anyway.

The counter already exists. Weave Router hit 130 points on HN by embedding each request with a small in-process ONNX model, scoring it against frozen intent-cluster centroids, and routing to the cheapest model that historically matched the frontier on that cluster. The team claims 60-70% of Claude Code requests are short, structurally simple completions an open-source model handles at parity, at roughly one-fortieth the cost. Token-compression middleware is becoming a standard install-once layer (rtk, headroom, LeanCTX all claim 60-95% reductions).

Here's the reframe: token discipline is not a billing footnote, it's a first-class engineering constraint, same as memory or latency. Treat context like a budget. Route cheap work to cheap models. Compress tool output before it hits the window. The teams that bake this in now won't get the 2028 invoice that makes their CFO question the whole program.