Learn · Comparison
AI gateway vs AI control plane.
Vidai, LiteLLM Proxy, Portkey and Kong AI Gateway side by side, on governance, drop-in compatibility and performance.
How the products compare
Cost control, policy and audit on every call.
Stronger answers: inside the perimeter, built in, acts rather than alerts.
| Capability | Vidai | LiteLLM Proxy | Portkey | Kong AI Gateway |
|---|---|---|---|---|
| Category | Sovereign AI control plane | Self-hosted open-source gateway | SaaS gateway | API gateway plugin |
| Deployment | Containers with PostgreSQL, inside your network | Self-hosted (FastAPI, Docker, Kubernetes) | SaaS, hosted by Portkey | Plugin on the Kong gateway |
| Where requests travel | Inside your perimeter only | Inside your perimeter | Across Portkey's perimeter on every call | Inside (self-hosted); cloud (Konnect) |
| Where the audit trail lives | Your storage, your retention, recorded in the request path | Your storage | Portkey storage | Your storage |
| Cost attribution | Per team, key, app and model, in real time, on a dashboard | Per virtual key, with a database | Per key, per workspace | Per consumer |
| Spend control (acts, not just alerts) | Hard budget breakers stop the next call when a cap is hit; automatic fallback to a cheaper model | Per-key budget limits with alerts | Spend limits with alerts | Rate limits (calls, not £ or $) |
| Routing for cost | Route by rule, and switch to a cheaper model when a budget is reached. What-if shows which models would have cost less on your real calls. You decide where calls go; nothing reroutes on its own | Configurable fallback chains | Provider failover, not cost-led | Plugin chain |
| Where prices live | A rate card kept in step: our optional public rate-card server (best effort), your own rate-card server (Enterprise), or prices you enter | Prices in config files; updates need a rebuild | Catalogue managed by the vendor | Per plugin |
| Rate history (priced correctly after a vendor price change) | Built in (SCD2): every call priced at the rate in force at that moment | Not in their documentation | Not in their documentation | Not in their documentation |
| Policy enforcement (PII redaction, region lock, model allow-list) | Built into the request path | Guardrail integrations | Add-on guardrails module | Plugin ecosystem |
| Guardrail models | Pattern checks plus VidaiGuard's own models; your own terms as patterns. We can help integrate your own models | Set integrations (Lakera, custom hooks) | The vendor's guardrail module | Per plugin |
| Regulatory mapping (EU AI Act, DORA, GDPR Art. 44, UK CTPR, FCA, ISO/IEC 42001) | Mapped explicitly: the audit trail is the answer for the regulator | Not in their documentation | Not in their documentation | Per plugin |
| Providers | OpenAI, Anthropic, Vertex, Bedrock, Azure OpenAI, plus self-hosted Llama and Mistral | The same set plus the OpenAI-compatible long tail | The same set | Via plugins |
| Install footprint | A 25 MB engine; containers with PostgreSQL, inside your network | Python runtime, Postgres and Redis for full functionality (hundreds of MB once dependencies are resolved) | SaaS by default; self-hosted is a Node.js app with Postgres and Redis | Kong gateway, Nginx, Postgres and the AI plugin |
The cost question
Attribution in real time, hard budget breakers that act rather than alert, and prices that stay right after a vendor price change.
Cost governance →The sovereignty question
Where requests travel, where the audit trail lives, and whether regulatory mapping comes with the product. With Vidai, all of these answers stay inside your perimeter.
Regulatory governance →Performance · the head-to-head test
Governed traffic at proxy speed.
Throughput, requests per second Higher is better
| Load profile | Vidai with governance | LiteLLM pass-through | Portkey pass-through | vs LiteLLM | vs Portkey |
|---|---|---|---|---|---|
| Low load | 1,713 | 131 | 495 | 13× | 3.5× |
| Modest load | 1,959 | 151 | 557 | 13× | 3.5× |
| Sustained production load | 2,526 | 152 | 656 | 17× | 3.9× |
p95 latency, milliseconds Lower is better
| Load profile | Vidai with governance | LiteLLM pass-through | Portkey pass-through | LiteLLM, times slower | Portkey, times slower |
|---|---|---|---|---|---|
| Low load | 8.5 ms | 646 ms | 92 ms | 76× | 11× |
| Modest load | 7.8 ms | 303 ms | 68 ms | 39× | 9× |
| Sustained production load | 12.8 ms | 5,009 ms | 1,200 ms | 391× | 94× |
Even doing more work, Vidai delivered 13 to 17 times the throughput of LiteLLM and 3.5 to 3.9 times that of Portkey across the three load profiles. Its own overhead, measured on one machine with no network in between, is 0.58 ms median; see the three tests.
What changes in your application
A base URL change, not an SDK rewrite.
Stronger answers: fewer code changes.
| Question | Vidai | LiteLLM Proxy | Portkey | Kong AI Gateway |
|---|---|---|---|---|
| What changes in your application | Base URL and API key. Nothing else. | Base URL and key; non-OpenAI SDKs may need rewriting to the OpenAI shape | Base URL and key; mainly the OpenAI shape | Per-plugin configuration |
| OpenAI SDK against any upstream | Yes | Yes | Yes | Via plugins |
| Anthropic SDK against any upstream | Yes, including non-Anthropic upstreams (OpenAI, Vertex, Bedrock, Azure) | Limited | Limited | Limited |
| Google GenAI SDK against any upstream | Yes, including non-Google upstreams | Limited | Limited | Limited |
| Self-hosted open-weight models (Llama, Mistral, your fine-tunes) | First-class, inside the VPC | Via OpenAI-compatible adapters | Routes to self-hosted endpoints; data still crosses Portkey | First-class |
The full SDK and upstream matrix, in the docs →For developers →
Which one is right for you
Pick the boundary that matches your constraint.
Vidai
You face data-residency or audit obligations, run agent-pace traffic, and want one boundary doing cost, policy and audit on every call.
LiteLLM Proxy
Strong for an early-stage, open-source-only setup with low throughput, where the Python runtime is not yet the limit.
Portkey
Strong when you have no perimeter constraints and want the fastest SaaS time to first request.
Kong AI Gateway
Strong for an existing Kong shop adding AI traffic to a gateway it already runs.
If your main constraint is cost attribution and enforced spend at agent pace, read cost governance. For the whole picture, see the platform and what a control plane is.
Frequently asked
Common questions on this comparison.
| What is the difference between an AI gateway and an AI control plane? | An AI gateway is an application-layer proxy: it routes a model call and emits a log line. An AI control plane is infrastructure: every request also passes through policy enforcement, real-time cost attribution and an audit trail recorded inside your own perimeter. The control plane is what regulated organisations reach for once gateway-shaped tools stop answering the audit, finance and sovereignty questions on their own. |
|---|---|
| Do I need an AI gateway or an AI control plane? | If a single team uses one provider, throughput is low, and nobody is asking the compliance question, an AI gateway is the right choice. You need a control plane when more than one team is using the same infrastructure with different budgets, a regulator or auditor is asking what the AI traffic did, or agent-pace traffic has broken the cost dashboard you set up for chat. |
| Are there self-hosted alternatives to LiteLLM Proxy? | Yes. LiteLLM Proxy itself is self-hosted (Python). Vidai is self-hosted (containers with PostgreSQL, inside your network) and adds in-path policy enforcement, per-team cost attribution and an audit trail mapped to regulatory frameworks. Kong AI Gateway is self-hosted via Kong's plugin model. Portkey is SaaS by default; the self-hosted option still crosses Portkey's perimeter when telemetry is collected. |
| What is the best AI gateway for enterprise? | The honest answer is that the best AI gateway for enterprise is usually not a gateway; it is a control plane. Enterprises that start with a gateway typically rebuild it as something stricter once the audit, residency and multi-team cost questions arrive. The comparison on this page lays out the trade-offs. |
| Can I use Vidai with my existing Anthropic SDK or Google GenAI SDK? | Yes. Vidai is drop-in with the native Anthropic SDK, the native Google GenAI SDK and the OpenAI SDK as-is. Base URL and API key change, nothing else. The full SDK × upstream support matrix is in the documentation at docs.vidai.uk/integrations/. |
Competitor entries summarise each product's public documentation and announcements at the time of writing; check their current documentation before you decide. Product names are trademarks of their owners.
How we start together
Start with one team, on your own AI traffic.
- 01
Pick one app or team
Point it at the Control room. One configuration change; nothing else changes.
- 02
Watch for two weeks
Every call priced, checked and recorded. Rules start in observe mode.
- 03
See the evidence
What it cost, what it caught and what it would have stopped, from your own traffic.