OmniRouteby Cheaper Inference

OmniRoute / Routing and control

Choose how each request moves.

Inspect routing strategies, free-tier accounting, compression, failure isolation and network controls.

Inspect routing Start with Cheaper Inference Back to overview

Routing strategies

Choose an `auto` route or build your own.

Use a ready-made mode or combine 19 strategies around the goal that matters.

Model · fallback enabledautoBalanced default
OpenAI-compatible request
curl http://localhost:20128/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"auto","messages":[{"role":"user","content":"Write a TypeScript retry helper."}]}'

Routing alias auto selected. Objective: Balanced default. Fallback enabled.

Drain subscription quota

Use quota you already have before touching paid APIs.

priority · fill-first

Spread load

Balance traffic across keys and accounts to reduce rate-limit pressure.

weighted · round-robin · p2c · least-used · headroom

Cheapest first

Route to the lowest-cost eligible target.

cost-optimized

Context-aware

Keep long contexts on models that fit, then relay when needed.

context-relay · context-optimized · cache-optimized

Randomized

Distribute eligible traffic across targets.

random · strict-random

Adaptive scoring

Use 15-factor scoring, last-known-good and reset-aware windows.

reset-window · reset-aware · lkgp · auto · fusion · pipeline

Free-tier accounting

Approximately 1.51B published tokens each month.

455 catalog entries are deduplicated into 40 recurring pools; only published positive budgets feed the headline.

1.51B

Pool-deduped accounting keeps shared quotas from being counted more than once. No-cap providers stay visible without inflating the token total. Checked 2026-08-30.

View the calculation (opens in a new tab)
Recurring or keyless free-forever providers56
  • Mistral1B/month
  • LLM7150M/month
  • Nara150M/month
  • Gemini CLI · forever60M/month
  • Kiro50 credits/month
  • Qoder · foreverFree catalog
  • LongCat · forever50M/day
  • Cerebras1M/day
  • NVIDIA NIM~40 RPM
  • Pollinations · foreverNo key
  • Cloudflare Workers AI · foreverDaily neurons
  • OpenCode Free · foreverNo-auth catalog

Context compression

Reduce eligible tokens by 15–95%.

Tool output, repeated context and bloated structured data can be compressed before they reach the model.

Illustrative Ultra profile example from the audited documentation.

Before69 tokens
The function should iterate over the list of items, and for each item, it should check whether the value is greater than zero, and if so, add it to the running total.
After19 tokens
sum all list items where value > 0

Composable pipeline

  1. Session-Dedup
    Drop content already sent earlier in the session.
  2. CCR
    Archive large blocks and restore them on demand.
  3. Lite
    Apply conservative lossless cleanup.
  4. RTK
    Filter and deduplicate command and tool output.
  5. Responses Tool Output
    Compact Responses API tool payloads.
  6. Headroom
    Apply reversible tabular and structured-data compaction.
  7. Relevance
    Keep sentences most relevant to the latest query.
  8. Caveman
    Use rule-based prose condensation.
  9. Aggressive
    Apply stronger configured pruning.
  10. LLMLingua-2
    Use code-safe semantic pruning in an isolated worker.
  11. Ultra
    Combine high-savings stages with preservation guards.
  12. OmniGlyph
    Encode eligible repeated structures compactly.

Profiles

Published profile range~30%x-omniroute-compression: standard

Standard compression selected. Published range ~30%.

Controls and output styles

  • Named profiles and active selector
  • Visual Compression Studio
  • Anthropic Context Editing
  • Adaptive context-budget dial
  • Per-request `x-omniroute-compression` header
  • Offline eval harness
  • Token-threshold auto-trigger
  • Terse prose
  • Less code (YAGNI)
  • Terse CJK 文言

Control precedence: request header → combo → profile → adaptive policy → default.

Failure isolation

Isolate provider, connection and model failures.

A bad key should not kill a provider. A limited model should not kill a connection.

model ⊂ connection ⊂ provider
Provider circuit breakerWhen a provider is unhealthy, combo routing skips it until a half-open probe succeeds.
Connection cooldownCool down one key or account while sibling connections keep serving.
Model lockoutLock one rate-limited or unavailable model while other models on the connection remain available.

Network and local data controls

Network routing and local data controls.

Proxy scopes control network routing while local-first storage keeps operational data on your machine.

Three proxy scopes

Route globally, per provider or per individual connection.

  • Global · per-provider · per-connection
  • SOCKS5 and HTTP(S) proxies
  • Proxy marketplace integration

TLS and client fingerprint controls

Match supported provider clients down to selected transport characteristics.

  • JA3/JA4-capable transport via wreq-js
  • Provider-specific client profiles
  • IPv4 and IPv6 egress controls

Private and local-first

Keys, usage and history stay on the host you control.

  • SQLite on disk with AES-256-GCM encrypted credentials
  • No OmniRoute account required
  • Self-host on infrastructure you choose
RussiaChinaIranCubaTürkiye

OmniRoute by Cheaper Inference

Choose how you run OmniRoute.

Self-host OmniRoute, or use Cheaper Inference for a hosted cost-ranked gateway.

Start with Cheaper Inference Self-host OmniRoute (opens in a new tab) View on GitHub (opens in a new tab)