A short story about the price of intelligence

Your agents are burning money on easy work.

Five minutes of scrolling. One idea: score every request by difficulty, then pay only what the work deserves.

01 · The squeeze

Tokens per dollar keeps falling.

Frontier models get smarter every quarter — and your dollar buys fewer of their tokens. Agent harnesses multiply the effect: one instruction can fan out into dozens of model calls.

1,000K tokens / $
2023 2024 2025 2026
directional, not a benchmark — the point is the slope
02 · The pile-up

So you dipped your toes into every provider.

The opinion wars never stop — whose model is best this week, whether you should switch. So you tried them all.

ClaudeMax plan$100/mo
OpenAI CodexChatGPT Plus$20/mo
CopilotPro+$39/mo
GeminiAI Pro$20/mo
OpenRouterusage credits$25/mo
Free tiersscattered quotas$0/mo
≈ $0 / month — that's the problem
03 · The waste

And every request — trivial or hard — goes to the same frontier model.

Renaming a variable costs the same as designing a migration. The model doesn't mind. Your bill does.

fix a typo rename a variable summarize a diff design a migration strategy
frontier-model$$$$every single time
04 · The irony

Meanwhile, the GPU you already paid for does nothing.

Local models are now genuinely good at small and medium coding work. Yours is rendering a desktop wallpaper.

Ollama · local runtime idle
  • qwen3.6:27b ready
  • gemma4:12b ready
  • qwen2.5-coder ready
GPU utilization
0%
05 · Enter ThinRouter

Connect everything you already pay for.

Frontier subscriptions, usage plans, free tiers, your own GPU. ThinRouter catalogs every model behind them.

06 · The score

Every request is scored 0–100 — then routed to the cheapest model that can do the job.

Your agent harness sends the request. ThinRouter measures its complexity and picks the lane.

07 · The lanes

Every lane has a strategy. When a model hits its cap, the request falls through.

Cheapest, cost-effective, top-down, round robin, or random — per lane. Caps and limits never stop the work. Your agents run 24/7.

Watch it happen

You control the replay.

A real request being typed into the playground and routed — scrub it with your scroll.

localhost:4100 · playground — live capture
No mockups

That diagram is a real screen.

Lanes, thresholds, and strategies are all configurable — or hit Smart fill and ThinRouter pre-populates them for you.

localhost:4100 · routing
ThinRouter routing screen — complexity thresholds at 27, 54 and 82, the complex lane selected with its top-down strategy and an ordered list of frontier models

Paste any prompt into the playground and watch it get scored and routed, live.

localhost:4100 · playground
ThinRouter playground — a refactoring prompt scored 59 out of 100 into the complex lane, with the candidate table and the routed model's effective cost
08 · The receipts

Built-in compression. Full visibility.

Prompt compression, out of the box

Repeated history and tool output get compacted before they're forwarded — code, JSON, and errors stay protected, and it fails open if the compressor isn't available.

localhost:4100 · instructions
ThinRouter standing instructions screen — the prompt efficiency panel with hybrid compression mode, token threshold, and target ratio

Insights shows where every request went

The local-versus-cloud split, spend after quota and cache, how requests were classified into lanes, and what provider caching saved you.

localhost:4100 · insights
ThinRouter insights screen — 15 requests across 11 models, local versus cloud split, per-lane classification bars, and cache savings tiles

And every single decision is legible: the score, the signals that fired, the candidates considered, and why the winner won.

localhost:4100 · live log
ThinRouter live log with one request expanded — complexity 59, two signals fired, the complex-lane candidate table, and the routing verdict
The end · your move

Run it tonight.

One Node process. One SQLite file. Point your harness at localhost:4100/v1, set the model to "auto", and stop overpaying for easy work.