Routing and observability for coding agents
Every turn goes to the model that fits it, and every token is tied to the work it produced. Your engineers keep using Claude Code or Codex.
10×lower cost per task than Fable 5.1 on Terminal-Bench 4
Sound familiar?
Coding agents made your team faster. They also made the bill unreadable.
One bill, no breakdown
The invoice says how much. It does not say which team, which tool or which pull request it paid for.
Frontier prices for git status
Every turn goes to the most expensive model, including the many that a model at a tenth of the price handles fine.
Overruns you hear about on day 31
By the time finance flags the month, the budget is already gone and nobody can say where it went.
Product
See where the tokens go, then send each turn to the right model.
- Saved
- $29,428
- Spend
- $6,310
- Requests
- 91.6K
Savings and actual spend
SavedSpend- M. Okaforplatform · 31 PRs$2,140$69 per PR
- R. Lindqvistpayments · 12 PRs$1,870$156 per PR
- S. Tanakaplatform · 27 PRs$1,320$49 per PR
- D. Alvarezmobile · 22 PRs$980$45 per PR
Every token, next to the work it produced
See what each engineer, team and tool spends, and what it shipped.
Attribution
Spend by engineer, team and tool, matched to merged PRs.
Finance reporting
By cost center and quarter, including capitalizable spend.
Benchmarks
Near frontier results at a tenth of the cost.
Pass@1 on Terminal-Bench 4 · higher is better
54.0%
Porridge router, 3.9 points below Fable 5.1
Cost per task on Terminal-Bench 4 · lower is better
$1.89
Porridge router, 10× lower than Fable 5.1
How it works
Running by lunch.
- 01
Run one command
The installer points Claude Code or Codex at your router, for one machine or for a repository.
curl -fsSL $ROUTER/install.sh | bash - 02
Keep working
Each turn is read and sent to the model that fits it. Hard turns still get a frontier model.
- 03
Read the dashboard
Spend by engineer, team and PR, with budgets that warn you while there is still time to act.
Savings
What would your team’s bill look like?
Move the slider to your monthly spend on coding agents and pick the model most of it goes to today.
The rates come from our production pilot: the same requests, priced at list rates, routed through Porridge and sent straight to one model. Your mix of work will differ.
- You keep, per year
- $189,600
- Your bill, per year
- $50,400
- $240,000 today
Production pilot: 79% less than sending every turn to Fable 5.1.
FAQ
Questions teams ask first.
Which coding tools does it work with?
Claude Code and Codex. One install command writes their config for you, for a whole machine or for a single repository your team commits.
Do engineers have to change how they work?
No. They run the same tools with the same commands. The tools talk to the router instead of the provider, and the router picks the model for each turn.
Which models can it route to?
Frontier models from Anthropic and OpenAI, plus lower cost models such as Kimi and GLM. The routing feed above shows a typical mix.
What about turns that really need a frontier model?
They get one. Hard turns like planning a migration or a large refactor still go to a frontier model. The savings come from the many small turns that never needed it.
Do you store our prompts and code?
Not by default. Content capture is off unless you turn it on. The router records model choice, token counts and cost, which is what the dashboard needs to attribute spend.
Can we run it in our own cloud?
Yes. The router is one service with the dashboard built in, and the install command points your tools at whichever deployment serves it.
How is it priced?
It depends on team size and current spend. Book a demo and we will walk through the numbers using your own usage.
Spend frontier money only on frontier work.
We will run your team’s real usage through the router and show you the bill before and after.