NewResearch: the same pull request data supports opposite conclusions

Optimization

Not every request needs the expensive model.

Most requests do not need the expensive model. The hard part is not saving money, it is proving that quality held while you did, which requires measuring the work rather than just the latency.

~/acme/webbar 80

Rename the billing props

4 files changed, 61 lines

~/acme/ledgerbar 85

Why is the payments test flaky?

bisected 9 runs, found it

~/acme/platformbar 88

Design the caching layer

write-through, 90s TTL

~/acme/webbar 82

Add a /changelog page from MDX

2 routes, RSS 2.0

~/acme/apibar 86

Write the migration for orders

reversible, 1 new index

~/acme/infrabar 78

Bump the Terraform providers

no plan diff

~/acme/authbar 87

Refactor the session module

11 call sites updated

~/acme/webbar 84

Add tests for the cart reducer

18 cases, 3 edge

~/acme/searchbar 86

Explain why recall dropped

analyzer change, week 31

~/acme/webbar 80

Rename the billing props

4 files changed, 61 lines

~/acme/ledgerbar 85

Why is the payments test flaky?

bisected 9 runs, found it

~/acme/platformbar 88

Design the caching layer

write-through, 90s TTL

~/acme/webbar 82

Add a /changelog page from MDX

2 routes, RSS 2.0

~/acme/apibar 86

Write the migration for orders

reversible, 1 new index

~/acme/infrabar 78

Bump the Terraform providers

no plan diff

~/acme/authbar 87

Refactor the session module

11 call sites updated

~/acme/webbar 84

Add tests for the cart reducer

18 cases, 3 edge

~/acme/searchbar 86

Explain why recall dropped

analyzer change, week 31

Haiku

4.5

scores42$0.80routed

GLM-5.2

Z.ai

scores55$4.40routed

Gemini

3.1 Pro

scores71$12routed

Sonnet

4.6

scores83$15routed

GPT-5.5

Codex

scores87$30routed

Opus

4.8

scores94$75routed
How it decides

Four questions, asked per request.

A router that only asks how hard the prompt looks will send everything expensive the moment it is unsure. Ours prices the decision: the cost of switching, the odds the cheap model finishes without a retry, and which of your subscriptions has quota left before it spends anything on the API.

  • How hard is this turn, given the repository it is acting on?
  • Would switching cost more in retries than it saves in tokens?
  • Is the cheaper model actually finishing this class of task?
  • Which subscription has quota before we pay API rates?
Routing decisionlive sample
  • Difficulty0.21

    low · single-file edit

  • Switch cost$0.004

    below retry threshold

  • Completion odds97.2%

    haiku, this task class

  • Quotaavailable

    team subscription

Routed to haikusaved $0.31
Proof

Cheaper is easy. Cheaper at the same bar is what we do.

Any router can cut your bill by sending everything to a small model. The reason to trust this one is that the quality bar is measured on your work rather than on a public benchmark, using the same machinery that computes your Slop Index, so a saving that quietly raised your rework rate shows up as a loss rather than a win.

  • Quality measured on your repositories, not a leaderboard
  • Savings reported net of retries, rework and review drag
  • Per task class, so you can route aggressively where it is safe
  • Roll back a routing policy the moment the bar moves
What we watch
  • Spenddown
  • Pass rateheld
  • Retrieswatch
  • Reworkwatch

The fourth bar is the one that matters. A router that cuts the bill and raises rework has not saved anything.

Setup

Your tools stay exactly as they are.

The router is a base URL. Point your existing agents and IDEs at it, keep your own provider keys, and nothing else in your workflow changes. Self-hosting is available if your requests cannot leave your network, which for several of the organizations we work with is not negotiable.

  • Drop-in base URL, OpenAI-compatible
  • Bring your own keys, we never hold provider credentials
  • Self-hosted deployment for regulated environments
  • Per-policy rollback, and a kill switch that reverts to frontier
Three stepsabout five minutes
  1. 1

    Get a router key

    From the console, scoped per team.

  2. 2

    Change one line

    Point the base URL at Behold.

  3. 3

    Work as usual

    Readings appear within a day.

base_url = "https://router.beholdlabs.dev/v1"

Put your traffic through it and see.

The rest of the platform

Behold Labs

Tell us a little more.