Routing strategies

AI Gateway · all 18 combo strategies animated — with an honest map of what runs here today vs OmniRoute.

All 18 combo routing strategies, animated. Each tile below animates the path its strategy takes — ● request → ▮ targets, green outline = the pick. The grid is OmniRoute's (MIT); the badges underneath say honestly which flows this gateway runs today. From OmniRoute ↗ · try the live ones on Connections.
Animated grid of 18 routing-strategy tiles
What runs here today — no vaporware: live, partial, or roadmap
priority✅ live here
the fallback chain: router-ranked candidates tried in order
auto (12-factor)✅ live here (our flavor)
the router scores fit · cost · speed per task with transparent rationale
cost-optimized✅ live here
model "auto/cheap" — cheapest that clears the quality floor
headroom✅ live here
model "auto/free" — drains the free-tier pool with most remaining allowance first
fill-first◐ effectively
cooldowns drain a model until it 429s, then move on — fill-first by consequence
reset-aware◐ partially
429 cooldowns honor Retry-After — reset-time awareness without the ranking
⚠ A caveat on fusion / panel strategies. Fanning a prompt out to several models and synthesizing the answers buys less diversity than it looks: “Artificial Hivemind” (Jiang et al., arXiv 2510.22954) finds widespread mode collapse on open-ended prompts — both within a model (repetition) and across models (homogeneity) — and that LM judges are poorly calibrated exactly where humans disagree. Route across models for cost, speed and capability with confidence; route for a second opinion on open-ended questions with much less.
Roadmap (OmniRoute-only today): weightedround-robinp2cleast-usedrandomstrict-randomreset-windowcontext-relaycontext-optimizedlkgpfusionpipeline
Feature comparison — OmniRoute's table, plus an honest column for this app
FeatureOmniRouteOther routersLLM Switchboard
🌐 Providers27820–1003 wired + 26-provider directory
🆓 Free providers90+ (40+ free forever)1–53 wired free tiers — see the Free-tier page
🔀 Routing strategies18–19 strategies1–34 live + 2 partial (priority · auto · cost · headroom)
🗜️ Token compressionRTK + Caveman stacked (15–95%)None / 20–40%None — roadmap
🧰 Built-in MCP server104 tools, 3 transports, 31 scopesRareNone — the app is the tool
🤝 A2A agent protocol6 skills, JSON-RPC 2.0NoneNone
🧠 Memory (FTS5 + vector)YesRareNone — stateless by design
🛡️ Guardrails (PII, injection)YesRareAuthlee SSO gate only
☁️ Cloud agentsCodex, Cursor, Devin, JulesNonen/a — bring your own agent, point it at /v1
🥷 TLS fingerprint stealthJA3/JA4 via wreq-jsNonen/a — transparent, nothing to hide
🖥️ Multi-platformWeb · Desktop · Termux · PWAWeb onlyZero-dep Node — self-host anywhere (Railway, laptop, edge box)
🌍 i18n43 locales0–4EN
The honest read: OmniRoute is a Swiss-army proxy platform; this gateway is the routing core — one endpoint, smart fallback, free-tier headroom — inside a product that also does benchmarks, architectures, RSI and more. Different shapes, shared idea.
Sign in to continue

LLM Switchboard is private — sign in with Authlee to access the control room.

Sign in with Authlee
← Back to home