How Janus Edge compares
Most AI gateways fall into one of two camps. Some are hosted services that keep your traffic metadata; others are open-source routers whose governance features are only available after a sales call. Janus Edge is a governed, self-hosted gateway that you can buy straight from this site. The tables below come from each vendor's public pricing pages and documentation; if something has changed, please let us know.
| Janus Edge | LiteLLM | Portkey | Kong AI Gateway | TrueFoundry | Helicone | |
|---|---|---|---|---|---|---|
| Self-hosted with SSO/SCIM, purchasable online | Yes — Business, per seat | Enterprise, contact sales | Enterprise, contact sales | Enterprise, contact sales | Enterprise, contact sales | Enterprise, contact sales |
| Pricing metric | Per active user, flat annual | Annual capacity license | Per 100k logs + enterprise | Per gateway + per request + per model | Per user + per request | Per request + GB |
| Cost tied to your AI spend | Never | No | Log volume | Request volume | Request volume | Request volume |
| Free tier for small teams, self-hosted | 25 users, all core features | OSS; SSO ≤5 users | OSS router only | OSS proxy, basic plugins | SaaS only, 3 users | OSS |
| Per-person model grants | Yes | Yes | Workspace level | Consumer level | Yes | Observability focus |
| Metering incl. cached / cache-write tokens | Yes | Yes | Yes | Partial | Yes | Yes |
| Prompts stored by default | No | Configurable | Logged (SaaS) | Configurable | Logged | Logged |
| Air-gap / offline license | Every edition; perpetual keys in Enterprise | Enterprise | — | Enterprise | Enterprise | — |
| Source available | ELv2, full source | MIT + proprietary enterprise dir | Router MIT; control plane closed | Apache core; enterprise closed | Closed | Apache-2.0 |
| Deployment | Linux binary, Docker, Helm chart, Compose; image files for offline registries | Docker, Helm | Docker (router) | Docker, Helm, VM | Managed / VPC | Docker, Helm |
LiteLLM, Portkey, Kong, TrueFoundry and Helicone are trademarks of their respective owners. Pricing reflects public pricing pages reviewed October 2026.
Routing, models and people
Spreading traffic across your own GPU servers, keeping model names stable while the models behind them change, and letting people bring the AI plans they already pay for.
| Janus Edge | LiteLLM | Portkey | Kong AI Gateway | TrueFoundry | Helicone | |
|---|---|---|---|---|---|---|
| Load-balanced pools on live engine load | Reads vLLM queue and KV-cache, llama.cpp slots; Business | Least busy and latency, counted at the gateway | Weighted, with circuit breaker | Least connections and latency; Enterprise | Weight, latency and priority | Cheapest or fastest provider |
| Conversations kept on the server holding their prompt cache | Automatic, no client header needed | Opt-in session affinity | Sticky by header; Redis across instances | Consistent hashing on a header; Enterprise | Sticky routing | — |
| Stable model names with automatic fallback | Managed models; fallback on Business | Aliases and fallbacks | Configs attached to keys | Model alias; Enterprise | Virtual models | Fallback chain written by the client |
| People use their own ChatGPT, Copilot, Grok or Mistral plan | Each person signs in from the portal; every edition | ChatGPT and Copilot, one sign-in for the whole proxy | — | — | Codex with a ChatGPT sign-in | — |
| Per-model speed and load in the console | Tokens per second, concurrency and jobs per day | Prometheus metrics | Latency analytics | Prometheus metrics and audit log | Time to first token and per-token latency | Latency and time to first token |
Auto-routing, usage insights and decision models
Sending each request to the right model, learning what people use AI for, and using small classifier models for guards and decisions. Several gateways added parts of this in 2026; the differences are in what the router can see, how you prove a route is good enough, and whether it all runs on your network.
| Janus Edge | LiteLLM | Portkey | Kong AI Gateway | TrueFoundry | Helicone | |
|---|---|---|---|---|---|---|
| Routing on what the request is about | Category and difficulty; shadow mode; every edition | Complexity tiers from a classifier; beta | Rules on metadata and parameters only | Similarity routing or an external router; Enterprise | Complexity tiers, no shadow mode | Cheapest provider first |
| Different instructions for each route | Router, route and default instructions, added on every turn | Prompt per model deployment, not per router tier | Prompt template per routing target | Prompt decorator per gateway route, not per target | Prompt version per target in a virtual model | Prompt template the client picks |
| Proof a cheaper route is good enough | Blind judge on sampled live traffic, order swapped, with a confidence range | Blind judge in shadow | — | LLM-as-judge policy on prompt-response pairs | — | — |
| Reports on what people use AI for | Category labels, never prompts; by team, model and day | — | — | — | — | Tags your app adds itself |
| Choose the classifier behind guards | Prompt Guard, Llama Guard or a decision model; checks run in parallel | Any configured model, or your own guardrail code | Webhook to your service; Enterprise for custom | HTTP guardrail service; Enterprise | Guardrail servers you deploy | Fixed, hosted, OpenAI models only |
| Decision models served to apps | /v1/decisions and /v1/systemone, with grants, quotas and pricing | /v1/decisions and /v1/systemone | /v1/decisions, one provider | — | /v1/decisions and /v1/systemone | — |
| One reasoning setting works on every provider | Refused levels and settings fitted and retried automatically | Mapped per provider; unsupported settings fail unless dropped | Mapped per provider; drops set by hand | Provider-native settings by hand | Mapped per provider | Mapped per provider |
| Container images as files for offline registries | Every release, per architecture | Registry only | Air-gapped no longer offered | — | Copy script to your registry | — |
Security at the edge
Janus checks prompts and answers at your network edge, before anything leaves it. The question for each product is whether a check is built in, runs on your own hardware, and covers streamed answers.
| Janus Edge | LiteLLM | Portkey | Kong AI Gateway | TrueFoundry | Helicone | |
|---|---|---|---|---|---|---|
| Secret detection in the gateway | Built in; redact or block on Business | Enterprise license | Your own regex rules or a partner service | PII service container; Enterprise | Built in | — |
| PII redaction without a separate service | Cards, IBANs, SSNs, email, phone; redact on Business | Requires Presidio containers | Hosted check on Pro, or regex | PII service container; Enterprise | Built in only on their cloud | — |
| Codename and term lists | Encrypted lists with fuzzy matching (Pr0j3ct) | Keywords and regex, no fuzzy matching | Word lists on output, regex on input | Regex allow and deny lists | Regex patterns | — |
| Prompt-injection classifier on your own hardware | Prompt Guard 2 on your TEI server, or a decision model | Similarity list or external service | Partner services | External services; Enterprise | Built in only on their cloud | Prompt Guard, hosted, OpenAI models only |
| Content safety on your hardware | A decision model or Llama Guard 4 on vLLM; 14 Safety categories | Enterprise license | Hosted or partner moderation | Custom guardrail; Enterprise | Built in only on their cloud | Llama Guard, hosted, OpenAI models only |
| Your own content safety categories | Written in plain language, each with its own minimum confidence | Judge model with your criteria; one overall threshold | Fixed moderation categories; custom checks by webhook | Azure categories with thresholds; Enterprise | Four fixed categories; more with your own code | OpenAI moderation only |
| Ready-made parental controls | Child Safety: age ranges, a tutor that never gives the answer, kind support replies; every edition | — | — | — | — | — |
| Streamed answers checked before they reach the user | Short hold-back window, then released | Stream hooks with hold-back | Only after the stream ends, no action taken | Lakera plugin; Enterprise | Not applied to streams | — |
| Organization-wide policy that teams can tighten but not relax | Mandatory floor, with "Explain effective policy" | Global defaults; per team and key on Enterprise | Organization guardrails in the hosted control plane | Global plugins can be overridden | Baseline rules; a request header can bypass them | Turned on by client headers |
| Where the checks run | Inside the gateway, on your network | Gateway, plus any services you add | Some checks call Portkey's cloud | Gateway, plus Kong's PII service | Several only on their cloud | Helicone's cloud |
A dash means we found no such feature in the vendor's public documentation. "Enterprise" means the vendor documents it as part of a paid enterprise plan. Feature tables reflect each vendor's documentation as reviewed in October 2026.
When to pick something else
We'd rather you end up with the right tool. If you want a hosted service with managed dashboards and you're comfortable with request metadata leaving your network, Portkey and Helicone are both polished. If you already run Kong for your APIs, its AI plugins let you keep a single gateway. And if you just need a free router without governance for a team of fewer than five, LiteLLM's open-source proxy is excellent.
Janus Edge is the better fit when identity, cost attribution and keeping data on your network are requirements, when you run models on your own GPU servers, and when you'd like to buy all of that without going through a long procurement process.