The best LLM gateways in 2026, compared
Nine gateways — managed, open-source and self-hosted — on hosting model, OpenAI compatibility and published price. Every figure was read off the vendor's own documentation, and a cell that says “not stated” means the vendor did not say it.
The short answer
An LLM gateway is one API endpoint that routes requests to many model providers, so routing, failover, spend tracking and access control live outside your application code. Choose a self-hosted gateway if you already run a platform team, a managed one if you do not, and a platform if the gateway is only part of the problem you have.
The options at a glance
| Option | What it is | Hosting model | OpenAI-compatible | Published price |
|---|---|---|---|---|
| 2kw.ai | Managed platform: OpenAI-compatible gateway, schema-driven document extraction from 17 formats, and per-field cost tracking with grounding scores. | Managed, EU-hosted, operated by a German company. BYOK on the Scale and Enterprise tiers. | Yes — point your base URL at 2kw.ai and existing OpenAI SDK code runs unchanged. | €99 / €299 / €499 per month; annual billing costs ten months. 7-day free trial. Custom tier on request. |
| Cloudflare AI Gateway | A control and visibility layer in front of your existing providers: analytics on requests, tokens and cost, caching, rate limiting, and request retry with model fallback. | Managed, on Cloudflare's network. | Yes — an OpenAI-compatible /chat/completions endpoint at gateway.ai.cloudflare.com, with the provider named in the model parameter as provider/model. | “Available on all plans.” No separate gateway pricing is stated on the overview page. |
| Eden AI | Managed aggregation across many AI providers — LLMs plus OCR, documents, speech, vision, translation and embeddings. | Managed. A separate EU endpoint routes exclusively within the European Union and restricts routing to providers compatible with EU residency. | Not stated on the pricing or EU endpoint pages. | 5.5% platform fee at checkout, no subscription, no API call limit. Advanced tier priced on request. |
| Helicone | “Open-source LLM observability and monitoring platform” that also acts as a gateway — caching, rate limits and automatic fallbacks alongside request tracking and alerting. | Managed cloud, open source, with on-prem deployment on the Enterprise tier. | A pass-through proxy rather than one unified endpoint: requests go to gateway.helicone.ai with the provider named in the URL path and in headers. 24+ providers listed. | Hobby free (10,000 requests, 1 GB storage, 1 seat). Pro $79/month. Team $799/month. Enterprise on request. |
| LiteLLM | Open-source gateway you run yourself — an “OpenAI Proxy Server (LLM Gateway) to call 100+ LLMs in a unified interface”. | Self-hosted, in your own infrastructure, with an air-gapped deployment option on Enterprise. | Yes. | Open source, free. Enterprise adds virtual keys, budgets, spend tracking, audit logs, SSO/SCIM and RBAC at custom pricing; a 30-day trial key is offered. |
| OpenRouter | Managed router optimised for breadth of model access, with credits spent across providers. | Managed proxy. Hosting locations are not disclosed in the documentation. | Yes — “implements the OpenAI API specification for /completions and /chat/completions endpoints”, described as “a drop-in replacement for OpenAI”. | No subscription and no markup on inference. Fees apply when buying credits: 5.5% ($0.80 minimum) by card, 5% by cryptocurrency. BYOK is free up to $25,000/month usage ($200,000 on Enterprise), then 5% above that. |
| Portkey | Open-source gateway plus a managed platform — “a unified interface for interacting with over 250 AI models”, with guardrails, routing and observability. | Self-host the open-source gateway, or use the managed platform. Private cloud and VPC hosting are Enterprise features. | Yes — documented as a drop-in OpenAI SDK replacement for Python and Node. | Open source free to self-host. Developer free (10k logs/month, “not suitable for production workloads”). Production $49/month (100k logs, $9 per additional 100k). Enterprise custom. |
| Requesty | Managed gateway — “600+ models behind one OpenAI-compatible endpoint, with routing, caching and governance included”. | Managed. The EU offering runs exclusively in Frankfurt (AWS eu-central-1) with zero data retention. | Yes — change your base_url to the EU endpoint, “zero code changes”. | “Get started free”, with custom invoicing in EUR. No per-tier figures are published on the EU page. |
| Vercel AI Gateway | “A unified API to access hundreds of AI models through a single endpoint” — text, image and video generation plus embeddings, with retries across providers and spend monitoring. | Managed, on Vercel. | Yes — OpenAI Chat Completions and OpenAI Responses, plus Anthropic Messages compatibility. | “No markup on tokens. Tokens cost the same as they would from the provider directly, with zero markup, including with Bring Your Own Key.” No separate subscription figure is stated on the overview page. |
Every cell was verified against the linked vendor page on 31 August 2026. Those pages are the authority — pricing and features change without notice. 2kw.ai is one of the options listed and has an obvious interest in the outcome; its row is marked so you can weigh it accordingly.
Listed alphabetically. The order is not a ranking — these products solve overlapping but genuinely different problems, and the right answer depends on which one you have.
They are not all the same kind of product
How to choose without a spreadsheet
Start with the shape of your cost, not the list price. Usage-fee models — Eden AI at 5.5%, OpenRouter's credit-purchase fees, Vercel's zero markup — cost close to nothing at low volume and track your spend upward. Flat subscriptions like Helicone Pro at $79 a month, Portkey Production at $49, or 2kw.ai from €99 are worse at low volume and better once traffic is steady and predictable.
Then ask what happens when a provider has an outage, because that is when a gateway earns its place. Retry and fallback across providers is offered by Cloudflare, Vercel, OpenRouter, Portkey, LiteLLM and Helicone; the differences are in how much configuration it takes and whether failover is per-request or per-route.
Finally, check where the gateway runs and what it keeps, before you check features. That question is cheap to answer early and expensive to answer during a procurement review. Our EU-hosted gateway comparison covers it in detail.
Common questions
An LLM gateway is a single API endpoint that sits between your application and multiple model providers. Your code calls the gateway; the gateway routes each request to OpenAI, Anthropic, Google or a self-hosted model, and returns the response in one consistent format. It is the layer where routing, failover, spend tracking and access control live, so those concerns stay out of your application code.
It means the gateway accepts requests in the shape the OpenAI API defines — typically /v1/chat/completions — so an application already written against the OpenAI SDK can switch to it by changing the base URL and the API key, with no other code change. It does not imply the gateway routes to OpenAI models, and it does not guarantee that every provider-specific parameter survives the translation.
Self-hosted gateways (LiteLLM, Portkey's open-source gateway) cost nothing to license and put the gateway on infrastructure you control, at the price of deploying, upgrading, monitoring and scaling it yourself. Managed gateways remove that work and add a vendor to your supply chain. The honest test is whether you already run a platform team: if you do, self-hosting is cheap; if you do not, it is the most expensive option on this page.
Every proxy adds a network hop. How much depends on where the gateway runs relative to you and to the provider, and it can be offset — Cloudflare and Helicone both cache responses, which removes the provider round trip entirely for a repeat request. Measure it against your own traffic rather than trusting any vendor's published figure, this page's included.
They overlap and are not the same. A gateway sits in the request path and can route, retry and fail over. An observability tool records what happened. Helicone and Portkey do both; Cloudflare AI Gateway leans toward visibility and control over existing providers rather than unified routing. Decide which of the two you actually need before comparing prices, because the pricing models are not comparable.
They are the gateways whose claims could be verified from their own documentation on 31 August 2026. Bifrost, Kong AI Gateway, TrueFoundry and LLM Gateway are credible options left out for space rather than on quality. 2kw.ai is our product; its row is marked, the ordering is alphabetical rather than ranked, and every competitor figure links to the page it came from.
How this comparison was made
2kw.ai is our product and appears in the table, so read this as a vendor-published comparison. Every competitor figure was read off that vendor's own documentation on 31 August 2026 and is quoted or closely paraphrased. Nothing was estimated, averaged, or taken from a third-party article. Where a vendor page did not state something, the cell says so rather than filling the gap.
Pricing and capabilities change without notice. If an entry is wrong or has gone stale, tell us and it gets corrected.
- 2kw.ai pricing — 2kw.ai/#pricing
- Cloudflare AI Gateway — developers.cloudflare.com/ai-gateway/
- Cloudflare OpenAI-compatible endpoint — developers.cloudflare.com/ai-gateway/usage/chat-completion/
- Eden AI pricing — www.edenai.co/pricing
- Eden AI EU endpoint — www.edenai.co/eu-endpoint
- Helicone pricing — www.helicone.ai/pricing
- Helicone gateway docs — docs.helicone.ai/getting-started/integration-method/gateway
- LiteLLM Enterprise — www.litellm.ai/enterprise
- LiteLLM proxy docs — docs.litellm.ai/docs/simple_proxy
- OpenRouter FAQ — openrouter.ai/docs/faq
- Portkey pricing — portkey.ai/pricing
- Portkey introduction — portkey.ai/docs/introduction/what-is-portkey
- Requesty EU — www.requesty.ai/eu
- Vercel AI Gateway — vercel.com/docs/ai-gateway