New release: try new models safely, hand tasks to the right Agent

Pay for your strongest modelonly where the work needs it.

HiRoute is an open-source model router for agents. Routine work goes to economical models and complex work to your primary model, which steps in when an economical model falls short. Keep your Agent, combine subscriptions, APIs and local models, and make your subscriptions go further.

Works with Codex, Claude Code, Qoder, OpenCode and more · macOS / Omarchy / Windows / Linux Server

ONE LONG TASK · THE RIGHT MODEL FOR EACH STAGEMost of the work is routine. Only a few decisions need your strongest model.
ROUTINE WORK

Economical model keeps going

Read code → Collect facts → Apply known fixes

At economical prices
↻
New message · Decide againLook back · size up · look ahead
↗
HARD PARTS AND RECOVERY

Primary model, only where needed

Key decisions → Step in after a low score → Check the result

Primary prices only for this part

Illustration

~90%lower cost on a real task at the same acceptance standard Low-score protectionprimary takes the next turn when an economical model falls short Many AgentsCodex, Claude Code, Qoder, OpenCode and more

Try a new model on a real task

A recorded coding session: a new model handles routine turns, your primary model steps in after a low score, and trust adjusts from your own records.

Recorded real session; trust-level scenes are illustrations · Costs are API-equivalent estimates.

Same acceptance standard.
About 90% lower cost.

In a real research and architecture-review task, Qwen 3.8 Flash did the research and GPT-6 Astra made the critical decision. At the same acceptance standard, the task cost about 90% less than running everything on Astra.

91% lower whole-task cost at the same acceptance standard: HiRoute with Qwen research and Astra decisions $0.150–0.162 versus all GPT-6 Astra $1.953–1.987

Switching models keeps the savings

HiRoute changes models only at clear boundaries: a new user message, or context compaction and rebuild. Tool calls within a turn stay on one model, so a switch misses the cache only once.

Cache and whole-task cost analysis →

Critical decisions stay with the stronger model

Qwen alone avoided major errors in only one of three critical decisions; the mixed setup kept those judgments on Astra and passed all three reviews. Costs are API-equivalent estimates.

Full case study →

No need to hand over everything.
Roll out by task difficulty.

Start from New models on Home and pair the model to try with a familiar primary model. Every change is listed before you publish.
Trial steps ↗

Trust slider from Cautious to Entrusted with Trial, Familiar, Trusted, Hands-off and Entrusted presets; Trusted is selected at a simple-probability threshold of 0.65
01—

Trust sets a difficulty bar

System One estimates how simple each turn is. Turns that clear the bar go to the new model; harder work stays with your primary model.

02—

Open up gradually

Move from Trial to Familiar, Trusted or Hands-off as the new model proves itself, and it takes on more of your work.

03—

Primary still backs it up

If a new-model stage scores below the competence floor and the problem still matters, low-score protection hands the next turn to your primary model.

Which model ran, why,
and how well it went.

The session timeline shows which model ran and why. Model performance reports bring together stage scores, the regular share and cost estimates; with enough evidence they suggest a step up or down, and you decide whether to publish.

Real session timeline: after a low Flash score, the primary model takes the next turn under low-score protection, then work returns to Flash
A real session: primary steps in after a low score, then work returns to the economical model.
Model performance report: a step-up suggestion to Trusted, 75% of stages on Regular at 90% competence, with a cost estimate
Model performance report: a step-up suggestion once the regular model clears the bar.
System One · Decision models

The decision model answers.
HiRoute decides.

For each new message, a decision model answers three independent questions along the task’s timeline. HiRoute applies your plan’s thresholds, picks the group and runs it, so you can retune routing without rewriting prompts. Connect Bailian, OpenAI, OpenRouter Jev, TypeSafe or a compatible endpoint in Models → Decision models.

Decision models and routing mechanism ↗
Three questions: how did the last step go (competence score), how hard is this step (simple probability), and will the last step’s problem carry into the next (earlier problem relevance); a decision model answers them independently and HiRoute applies your plan to pick the economical or primary model
PREVIOUS · LOOK BACK

How did the last step go?

A competence score that also feeds model reports and trust suggestions.

CURRENT · SIZE UP

How hard is this step?

Only simple-enough work can go to the economical model.

NEXT · LOOK AHEAD

Will the last step’s problem carry into the next?

If it does, low-score protection hands the turn to primary; if not, protection is released and difficulty decides.

Keep your Agent.
Not limited to coding.

HiRoute runs locally and brings supported subscriptions, API connections and local models into one place, for coding, research, analysis and automation.

Your decision policy.
One HiRoute interface.

Deploy a custom extension for your own decision policy. HiRoute supplies published categories, available groups and safe history. The extension returns a category and regular or primary group; optional prior-stage scores are observation only. HiRoute retains model selection and execution.

Custom extension API →
Download OpenAPI ↓
CUSTOM EXTENSION API / FIRST RESPONSE
POST /v1/decisions

{
  "group": "regular"
}
First difficulty-routing decision · No history target, so no assessment

Get started in your environment.

Choose the desktop app, or use the Web UI and CLI on a server. Connect your models and keep working with your favorite Agent.

Read the Quickstart →

Stable 0.6.3 · Changelog →

Frequently asked questions

How does model switching affect KV Cache?

Yes. During normal consecutive tool calls in the same turn, HiRoute reuses the routing decision and prefers the current model within the selected group, preserving reuse of the growing prefix. A new user message triggers reassessment; a switch may lose the old cache on the first request, while later calls can reuse the new cache.

After compaction or a history rebuild, some cache entries already need rebuilding, so reassessment at that boundary may incur a smaller additional loss. HiRoute preserves continuous execution to reduce unnecessary cache overhead. Whole-task costs include rebuilding, routing decisions, and retries; the article below explains the effects and Astra × Qwen calculations.

Read the cache and cost analysis: Astra × Qwen 3.8 Flash →

Does context compaction always trigger a new model choice?

It depends on whether HiRoute can still recognize the same-turn continuation and reuse the frozen decision. A reusable continuation inherits that decision. If a history rebuild breaks continuity or the decision cannot be reused, HiRoute judges again. The new choice can keep the same model; clients need no separate compaction event.

What execution history can a decision model see?

HiRoute supplies simplified Agent-turn history: user input, accepted output, tool names, and coarse outcomes. HiRoute binds internal task categories, groups, and execution identities itself. Trimming, eviction, or restart may leave history incomplete; completeness of the assessed stage is tracked separately. Missing or partial assessments are not zero scores.

Must I deploy Jev?

No. Connect Bailian, OpenAI, OpenRouter Jev, TypeSafe, or a compatible endpoint in Models → Decision models to use built-in judgment. Difficulty routing also supports heuristic rules. Choose a custom extension when you need your own processing logic.

Is competence a success rate?

No. It represents competence for a specific execution stage and must be interpreted with plan, revision, assessment coverage, and partial-evidence state.