Frontier models for decisions. Cheap models for everything else.
Coding agents spend most of their tokens on work that isn't reasoning: pulling whole files into context, scanning the repo, and writing predictable code. We put a routing layer between your engineers' AI tools and the models behind them, so that routine work runs on fast, inexpensive models and only the real decisions reach the frontier model. Your engineers keep the tools they already use.
Reported by Spotify
~90%
fewer Claude Code tokens with this pattern, frontier model kept for decisions.
Spotify Engineering (September 2026) reported about 90% fewer Claude Code tokens in tests of a two-model routing setup on a Java monorepo. Spotify is not a TokenShunt customer. Your savings depend on how much of your token use is reading and boilerplate, which is why we measure your own sessions first.
Who it's for
Built for teams paying real money for AI coding tools.
- Engineering organizations spending real money on Claude Code, Cursor, or usage-based GitHub Copilot.
- Platform and developer-productivity teams asked to cut AI tooling spend without slowing anyone down.
- Companies whose AI coding bill grew faster than headcount after agents went mainstream.
The problems we solve
- 01
Frontier prices for reading files
An agent that pulls a 2,000-line file into a frontier model's context to find ten relevant lines pays reasoning prices for what is really a search.
- 02
Guidance that agents ignore
Asking an agent in its prompt to “use a smaller model for reads” is a suggestion it can drop. Savings that hold come from hard rules at the routing layer.
- 03
Blanket downgrades hurt quality
Switching every engineer to a cheaper model saves money until the hard problems start failing. Only routine work should move.
How routing works
Stop paying frontier prices for routine work.
Most of what a coding agent does isn't reasoning. We route file reads, repo scans, and boilerplate to fast, inexpensive worker models — self-hosted where code can't leave your network — and keep the frontier model for the decisions that need it.
Reported by Spotify
~90%fewer Claude Code tokens1 with this pattern. We measure yours before promising anything.
- ≈ free
Cache & repo index
Repeated lookups, symbol and file search
- Lowest
Cheap worker models
Reading large files, repo scans, boilerplate, tests
- Moderate
Mid-size models
Edits from a clear spec, small refactors
- Highest
Frontier model
Architecture, debugging, novel problems
Illustrative. Line weight shows relative token volume; anything that fails a quality gate goes back to the frontier model.
1 Spotify Engineering (September 2026) reported about 90% fewer Claude Code tokens in tests of a two-model routing setup on a Java monorepo. Spotify is not a TokenShunt customer. Your savings depend on how much of your token use is reading and boilerplate, which is why we measure your own sessions first.
Process
How the engagement runs.
Clear stages, explicit decision gates, and an artifact at every step — so progress is visible and nothing depends on trust alone.
Step 01
Token map
Classify your agents' real sessions by job (file reads, repo scans, boilerplate, edits, decisions) and price each one.
Step 02
Routing policy
Decide per job which model handles it, with hard limits such as files over a set size never reaching the frontier model.
Step 03
Worker models
Stand up cheap worker models for bulk reading and predictable code, via API or on your own hardware where code can't leave.
Step 04
Roll out behind a flag
Start with one team, compare routed and unrouted sessions on the same tasks, then widen.
Step 05
Tune
Adjust thresholds and routes from real usage until the cost curve flattens without complaints from engineers.
Deliverables
What you own at the end.
- Token map of your current AI coding spend
- Routing policy for Claude Code, Cursor, and Copilot workflows
- Worker-model setup (API or self-hosted)
- Routing layer deployed in your environment
- Before/after cost and quality report
- Runbook your platform team can own
Typical outcomes
What changes for your business.
- A lower AI coding bill, measured against a baseline of your own sessions rather than a vendor benchmark.
- No change to how engineers work: same editors, same agents, same review process.
- Frontier-model quality kept for architecture, debugging, and novel problems.
FAQ
Questions leaders ask us.
How much can this save?
It depends on how much of your token use is reading and boilerplate rather than reasoning. Spotify Engineering reported about 90% fewer Claude Code tokens in its tests of this pattern. We measure your own sessions first and report savings against that baseline.
Do our engineers have to change tools?
No. Routing sits between the tools your engineers already use and the models behind them.
Which models do the routine work?
Whatever is cheapest and good enough for each job on your code: small commercial models via API, or open-weight models on your own hardware when source code must not leave your network.
What if a cheap model gets something wrong?
Routine jobs are chosen because they're easy to check, and anything that fails your tests or quality gates goes back to the frontier model. See Quality Verification.
Other services
Start the conversation
Stop paying frontier prices for routine work.
Tell us which coding tools your engineers use and roughly what you spend. We'll show you where the tokens go and what routing would change.
We respond within 1 business day. Mutual NDA available before any data discussion.