Real-world cost savings
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.
API-equivalent estimates. Full case study
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 the same model. A switch misses the cache once, then
later calls hit it again, so cache loss does not eat into the savings.
Switching Models Loses Cache, Not the Savings
Familiar Agents, more model choices
Keep working in Codex, Claude Code, Qoder, Pi, OpenCode, Hermes or DeepSeek Harness.
HiRoute runs locally and brings supported subscriptions, API connections and local models
into one place. For example, combine an economical API model with your Codex subscription,
or use a smaller and a stronger model from the same supported source.
| What you want | What HiRoute provides |
|---|
| Spend less on your strongest model | Route routine work to economical models and keep the primary model for complex stages and low-score protection. |
| Keep your workflow | Continue in your familiar Agent while models share the work within a session. |
| Try a new model gradually | Start at a trust level you choose, keep a familiar primary model as backup and adjust with evidence from your own tasks. |
| Understand each decision | Inspect the actual models, routing reasons, stage performance, usage and cost estimates. |
One task, the right model for each stage
Each new message gets a fresh decision. A decision model answers three independent questions
anchored on the task's timeline, and HiRoute combines the answers with your plan:
- Previous: how did the last step go? A competence score, also used for model performance
reports and trust suggestions.
- Current: how hard is this step? A simple probability. Only simple-enough work can go to the
economical model.
- Next: will the last step's problem carry into the next? When a step scored below the
competence floor and its problem still applies to the work ahead, low-score protection hands the
next turn to your primary model. Once it no longer applies, protection is released.
The decision model only answers; thresholds, model choice and execution stay in HiRoute, so you can
retune routing without rewriting prompts. The first turn has no previous step, so only difficulty applies.
Keep asking follow-up questions or let your Agent advance a long task. HiRoute can reassess
at supported new-input and context-reconstruction boundaries; tool continuations normally
reuse the stage's decision. See routing behavior and supported-client limits.
Explore routing modes
Try new models with a primary model ready to help
Start from New models on Home. Instead of switching everything at once, HiRoute rolls a
new model out by task difficulty.
- 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.
- 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.
- 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.
Adjust with evidence from your own tasks
The session timeline shows which model actually ran and why. Model performance reports
bring together stage scores, the share handled by the regular group and available cost
estimates. When there is enough evidence, a report suggests a step up or step down; replay
the setting on your own records, inspect the changes and choose whether to publish.
You can also tailor the judgment criteria to your own tasks and compare prompt changes
using the published evaluation method.
How well does the routing policy work?
Both approaches receive the same task context and complete routing policy. One selects
a model group directly; the other answers three questions and lets HiRoute combine the
answers. The comparison uses engineering workflows and mathematical dependency states.
| Scenario | Decision model | Full-policy direct selection | Three-question routing |
|---|
| Engineering | Jev 1.14 | 67.31% | 96.15% |
| Engineering | OpenAI Decisions / gpt-6-luna | 50.00% | 91.35% |
| Mathematical dependency states | Jev 1.14 | 54.86% | 100.00% |
| Mathematical dependency states | OpenAI Decisions / gpt-6-luna | 65.97% | 88.89% |
Metric: final group-selection accuracy, averaged equally across workflow families.
Engineering has 80 determinate cases across 26 workflows; mathematics has 40 across six.
This is a post-hoc reviewed subset, with the same exclusions applied to every model
and both formats after inspecting outputs. Original labels and full-population results
remain available. Jev calls ran on 28 October 2026 and OpenAI calls on 29 October; inputs
and policy match, but the calls were not contemporaneous.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
