Frugal Tokens: the coding-agent bill finally gets a dashboard

A Show HN tool for comparing token cost and cache-miss impact across coding agents lands the same week Meta's Mosseri floated per-engineer token caps - the token-budget-caps thread just got its measurement layer.

In plain terms. A developer published a tool that lets you see how much your coding-agent sessions actually cost, session by session, and how much of that spend is cache misses. Boring? Absolutely - and that's the point. This is what the AI-coding market looks like when the CFO gets involved.

Context

The token-budget-caps thread started with Uber and Microsoft cutting AI coding licenses when Q2 budgets ran dry, then extended to Meta-side proposals of per-engineer token caps (see fil narratif token-budget-caps - Mosseri/Meta reference). Coinciding with DeepSeek raising API prices by up to 12× and introducing peak-hour pricing (#1822, #40247670), the industry has moved from "how do we get more capacity" to "how do we account for what we're already spending." Frugal Tokens is a symptom of that shift.

Under the hood

The tool ingests session logs from major coding agents. It exposes: cost per session, cache-hit ratio, tokens per accepted edit, and cost distribution across sessions. The author's stated motivation was noticing wide spend variance between users on similar tasks. That variance is almost entirely explained by two things: prompt structure (how often are we invalidating the cache?) and context bloat (how many tokens do we spray in for how much useful output?).

In plain terms

If your team runs coding agents at scale, you have two levers. First, cache-friendly prompting - put stable context at the top, moving parts at the bottom. Second, context discipline - the InfoQ "Right 300 tokens" talk (#1939) argued that 300 well-chosen tokens beat 100k noisy ones. Frugal Tokens gives you a metric to prove or disprove that on your own workflows.

Analysis

The interesting move is not the tool itself - anyone with API keys could build it. It's that a measurement layer is appearing at all. The AI-coding market spent 2024-2025 obsessing over model choice ("is Opus better than Sol?"). 2026 is turning into a fight over harness efficiency, and efficiency requires measurement. Expect three or four of these to consolidate into a "Datadog for coding agents" category by mid-2027.

Risks

Session-level accounting misses org-level effects: junior engineers over-prompting, senior engineers using agents as glorified autocomplete. Per-user dashboards can also drift into surveillance-of-productivity if not scoped carefully.

So what

If you manage an engineering org: pick a session-cost tool, get baseline cost/accepted-edit numbers now, and use them to argue for training on prompt structure - not for cutting seats. The variance between users is wide enough (the tool's own author noticed it before building the dashboard) that it's a coaching problem, not a licensing problem.