swadhin.ai

Leaderboard · updated 16 September 2026

AI Gateway Leaderboard

The top score is 20 of 20 capabilities documented, held by two gateways, as of September 2026. The table ranks 9 gateways on twenty capabilities each, and lists four more beneath it with under 1,000 GitHub stars. Of the 260 cells, 234 link the vendor page they were read from; the other 26 have no page behind them and earn nothing. Every capability was checked against the gateway's own documentation and links the page it was found on. Last change 16 September 2026.

Open-source , ranked on how many of twenty capabilities their own documentation shows. One point a capability, so the figure is a count. The selects above the table narrow it; the ranks do not move.

We went through every gateway on this table one capability at a time, by hand, from each vendor's official documentation, between 8 September 2026 and 9 September 2026. Of the 260 cells, 234 link the page the answer came from, and every one of those links was checked again on 15 September 2026. For the other 26 we found no page either way.

How we checked
Every mark was checked against the vendor's own page
We worked through the twenty capabilities for each gateway in turn and linked the page each answer came from, so you can open any cell and see what we saw. A tick says the capability is documented. It does not say it was installed and exercised here, and a documented capability can still be a thin one.
A question mark is where we found nothing
It earns nothing, the same as a cross, and it says something different: we looked and found no page either way. There are 26 such cells. A gateway with thin documentation reads lower here than one with a page for everything, which is a fact about its documentation as much as about the product.
Four star counts are the whole project's
Kong, Apache APISIX, Helicone and Higress ship their gateway inside a larger platform, and the repository counted is that platform. The cell says so on each of the four.

Documented, of 20

one point a capability · 9 ranked

One point a capability, twenty in all. A tick is a page in the vendor’s own documentation that shows the capability, a cross is a page that shows it is not there, and a question mark means no page was found. Only a tick counts. Open a mark to read the page.

Columns

Narrow the table with the selects. A rank stays with its gateway.

Click a column heading to sort the ranked rows.

Every gateway, ranked on how many of twenty capabilities its documentation shows. Data as of 15 September 2026.

What the columns mean
Rank
Where a gateway sits, ordered by how many capabilities it documents, among those with at least 1,000 GitHub stars. Two gateways with the same points are in alphabetical order, so the points are the finding and the rank is the row's position. Sorting by another column leaves these numbers where they are.
Gateway
The project, linked to its own site where the review holds one and to its repository otherwise. Beneath it, who publishes it and which kind of thing it is: a gateway in the request path, a hosted aggregator, or a thin relay.
Documented
How many of the 20 capabilities the vendor's own pages document, one point each. One documented as absent counts nothing. One nobody could find a page for counts nothing either, and says so. Every capability counts the same, nothing is a percentage, and nothing was run.
Stars
The repository's GitHub star count on the day beneath the table, read from the GitHub API by a script and never typed. A gateway under 1,000 is listed beneath the table rather than ranked; inside the table, stars move no rank. Four of the repositories are the platform the gateway ships inside, and the cell says so.
Sourced
How many of the twenty cells link a page, tick or cross. A low count here says the documentation was thin or the review did not find it, which is a different thing from the capability being absent.
Providers
The count of model providers the vendor's documentation names, in the review's own words, linking the page it was read from.
Timeouts
A request to a provider can be given a time limit, so a slow provider fails fast instead of holding the caller.
Retries
A failed request is sent again, with a count and a wait the operator sets.
Fallback
When a provider or model fails, the request goes to another one the operator named.
Circuit breaker
A provider that keeps failing is taken out of rotation for a while, rather than being tried on every request.
OpenAI-compatible API
Existing OpenAI client code works by changing the base URL.
Streaming
Tokens are returned as they are generated, not all at once at the end.
Tool calling
The model's requests to call a function pass through the gateway intact.
Structured output
The caller can ask for JSON, or for output matching a schema, and get it.
Provider translation
One request format is translated to each provider's own API, so the caller writes it once.
Routing
A request is sent to a provider or model by a rule: by cost, by latency, by weight, or by what the request contains.
Load balancing
Requests are spread across several copies or providers of the same model.
Rate limiting
A caller, key or team can be held to so many requests or tokens a minute.
Caching
A repeated request is answered from a store rather than sent to the provider again.
Cost tracking
Spend is recorded per key, user or team, and can be capped.
Observability
Requests are logged and traced, and the gateway exports to tools such as Prometheus or OpenTelemetry.
Authentication
Callers present a key the gateway issues and checks, rather than the provider's own.
Guardrails
Prompts and answers can be screened, for personal data, for banned content, or by a policy the operator writes.
MCP
The gateway speaks the Model Context Protocol, the open standard for connecting models to tools and data.
Kubernetes
The project publishes a Helm chart, an operator or a manifest for running it on Kubernetes.
Prompt management
Prompts are stored, versioned and served from the gateway rather than hard-coded in the caller.
AI gateways ranked on how many of twenty capabilities their documentation shows. Which rows are shown is set by the selects above the table; a rank stays with its gateway.
DetailsReliability · 4Protocol · 5Traffic · 5Operations · 6
RankGatewayDocumentedStarsSourcedLicenceDeploymentProviderstimeoutretryfallbackbreakerAPIstreamtoolsJSONtranslateroutingbalanceratecachecostlogsauthguardMCPk8sprompts
1LiteLLMBerriAI · AI gateway20 of 2058,789BerriAI/litellm20 of 20MITself-hosted · managed100+✓✓✓✓✓✓✓✓✓✓✓✓✓✓✓✓✓✓✓✓
2Portkey AI GatewayPortkey (PRISMA AIRS AI Gateway, Palo Alto Networks) · AI gateway20 of 2012,998portkey-ai/gateway20 of 20MITself-hosted · managed250+ models✓✓✓✓✓✓✓✓✓✓✓✓✓✓✓✓✓✓✓✓
3Kong AI GatewayKong Inc. · AI gateway17 of 2044,139Kong/kongcounts the whole project20 of 20Apache-2.0self-hosted · managed12+ listed✓✗✓✗✓✓✓✗✓✓✓✓, enterprise licence✓, enterprise licence✓✓✓✓✓✓✓
4BifrostMaxim HQ · AI gateway16 of 208,086maximhq/bifrost19 of 20—no licence in the recordself-hosted · managed20+✗✓✓✗✓✓✓✗✓✓✓✓✓✓✓✓, enterprise licence✓, enterprise licence✓?✓
5HigressAlibaba / Higress Group (CNCF Sandbox) · AI gateway16 of 209,389alibaba/higresscounts the whole project19 of 20Apache-2.0self-hosted · managedAll LLM providers (count ?)✗✗✓✗✓✓?✓✓✓✓✓✓✓✓✓✓✓✓✓
6Apache APISIX AI GatewayApache Software Foundation · AI gateway15 of 2017,124apache/apisixcounts the whole project20 of 20Apache-2.0self-hosted9 named + OpenAI-compatible✗✗✓✓✓✗✗✗✓✓✓✓✓✓✓✓✓✓✓✓
7Helicone AI GatewayHelicone · Aggregator14 of 206,158helicone/heliconecounts the whole project20 of 20GPL-3.0managed100+✗✗✓✗✓✗✓✓✓✓✓✓✓✓✓✓✓✗✗✓
8Agent Router (Envoy AI Gateway)Agentic AI Foundation / Envoy community · AI gateway11 of 202,094envoyproxy/ai-gateway18 of 20Apache-2.0self-hosted16 named + generic OpenAI-compatible✗✗✓✗✓✗✓✗✓✓✗✓✗✓✓✓?✓✓?
9new-apiQuantumNous (One API derivative) · Proxy or relay8 of 2048,140QuantumNous/new-api15 of 20AGPL-3.0self-hosted56 current channel types✗✓✗✗✓✗✗✗✓✓✗✓✓✓?✓????

✓ documented✗ documented as absent? no source foundent. enterprise licence · mgd. managed planEvery mark links the page it was read from.

Under 1,000 stars (4)

The table ranks gateways with at least 1,000 GitHub stars. These have fewer, or no count to test. Their points are computed the same way and shown here rather than ranked.

How the rank is computed

The rule, in full

One table, one ranking. A row is a gateway. There are 20 capabilities, and a gateway scores one point for each its own documentation shows. The figure is the count. A capability documented as absent counts nothing, and one nobody could find a page for counts nothing and is marked as such. There are no weights, no percentage, and no denominator that shrinks when a page is missing.

The groups. Reliability 4: timeouts, retries, fallback, circuit breaker. Protocol 5: openai-compatible api, streaming, tool calling, structured output, provider translation. Traffic 5: routing, load balancing, rate limiting, caching, cost tracking. Operations 6: observability, authentication, guardrails, mcp, kubernetes, prompt management. The groups are headings to read the twenty by. They carry no weight of their own: a capability counts one wherever it sits.

Rank. Most points first, then the name, and every row gets its own number, the way every board here does it. Two gateways on the same total are a tie the rank does not show, so the points are the finding and the rank is the row’s position.

A documented capability that needs an enterprise licence or a managed plan counts and is marked, kept as the review scored it. There are six such cells, listed in the decision record.

The star floor. A gateway with fewer than 1,000 GitHub stars on the day the counts were read is listed beneath the table with its points and never ranked. Inside the table, stars move no rank. A count that could not be fetched is treated the same way, because a floor nobody can test is not a floor.

Of the 260 cells, 234 were read from each gateway's own documentation, one page per cell, checked by us, between 8 September 2026 and 9 September 2026, and every one of those links was HTTP-checked on 15 September 2026. The other 26 have no page behind them. The star counts were read from the GitHub API on 15 September 2026 and never rank.

Left out4
  • Pricing. Every one of these is free to self-host and most publish no enterprise price, so a column would be thirteen dashes.
  • Speed, latency, CPU and memory. No independent benchmark covers these gateways the same way, so there is nothing every one of them could be scored against, and nothing is estimated to fill the gap.
  • Vendor benchmark claims. Every gateway here publishes none, so a column would be thirteen blanks.
  • The percentage score the source computes, and its weights. It drops every cell it could not source from the denominator, which ties six gateways at 100, and weighs a fallback at seven and a load balancer at two for reasons a reader cannot check. Capabilities are counted instead, one each, and a cell nobody could source counts nothing.

Changes

16 September 2026 · One point a capability, a floor of 1,000 stars, and the table set right3
  • Every capability counts one point now, twenty in all, and the figure is how many a gateway's documentation shows. The board launched with the review's weights, from one to seven, and a reader could not tell why a fallback was worth seven and a load balancer two. The nine ranked gateways keep the order they had.
  • A gateway with fewer than 1,000 GitHub stars on the day the counts were read is listed beneath the table with its points and no longer ranked. Four moved there: Hyperstrate Server, Ferro Labs AI Gateway, Freeport and AI Proxy. Nine rank. Among them, stars still move nothing.
  • The columns had collapsed into one another on launch: a fixed-layout table takes its widths from its first header row, and that row spanned groups. Every header row now carries one cell per column, and the twenty capability headings stand on end so each stays inside its own column.
15 September 2026 · AI gateway board launched5
  • Thirteen open-source AI API gateways, ranked on twenty capabilities their own documentation shows. Each capability carries a weight, a documented one earns it, and the total of seventy is the figure. Nothing was run: every cell links the vendor page it was read from.
  • The review this board is built from was made by swadhin.ai by hand between 8 and 9 September 2026, and every link was HTTP-checked on 15 September. Of 260 cells, 234 link a page and 26 have no source. A cell with no source earns nothing and is marked, which is the one place this board departs from the review's own score: that score was a percentage of the cells it could check, and it tied six gateways at 100.
  • Fourteen weights are the review's own, in three groups. Six capabilities the review collected without weighing them, observability, authentication, guardrails, MCP, Kubernetes and prompt management, are weighed here as a fourth group worth fifteen points.
  • A GitHub star count sits beside the points, read from the GitHub API by a script on the day shown, never typed, and never ranked on. Four of the repositories are the platform the gateway ships inside, and those cells say the count is the whole project's.
  • Left out: pricing, because every gateway here is free to self-host and most publish no enterprise price; speed, latency and resource use, because nothing was measured; and vendor benchmark claims, because the review holds none.

Questions

What is a point?

One capability the gateway's own documentation shows. There are 20 capabilities and each counts once, so the figure is a count: a gateway on 16 documents sixteen of the twenty. A capability documented as absent counts nothing, and so does one nobody could find a page for. The rank is the order of the counts, and two gateways on the same count are in alphabetical order.

Why points and not a percentage?

The review this board is built from scores a percentage of the points it could check, and drops every cell it could not source from the denominator. That ties six gateways at 100, including one with eight cells nobody could source. Counting points instead makes a missing page cost what it should: nothing earned, and the cell says so.

What does the question mark mean?

No page was found that shows the capability either way. It is not a cross: a cross links a page that shows the capability is absent. Of 260 cells, 26 carry the question mark today, and a later pass that finds their pages will change the order and be recorded under Changes.

Why does every capability count the same?

Because a weight is an argument, and a count is a fact. The review this board is built from weighed fourteen capabilities from one to seven and left six unweighted, and the board carried that scheme for a day. A reader could not tell why a fallback was worth seven and a load balancer two, and a score of 57 of 70 said less than it seemed to. One point each says exactly what it means. The four group headings are there to read the twenty by, not to score them: reliability 4, protocol 5, traffic 5, operations 6.

What does the small word under a tick mean?

That the documented capability needs an enterprise licence or a managed plan rather than shipping in the open-source build. There are six such cells, kept as the review scored them and counted in full, with the page linked so the reader can see the condition.

Why is there no price?

Every gateway here is free to self-host and most publish no enterprise price. A column of thirteen dashes would say nothing, so pricing is left out.

What do the stars count, and do they rank?

The repository's GitHub star count on the day beneath the table, read from the GitHub API by a script and never typed. They decide who is in the table and never the order inside it: a gateway under 1,000 stars is listed beneath the table with its points, and among those above, a star is a bookmark, not a capability. Four of the repositories are the platform the gateway ships inside, so those counts are the whole project's, and the cell says so.

Why are some gateways beneath the table?

They have fewer than 1,000 GitHub stars: Hyperstrate Server, Ferro Labs AI Gateway, Freeport, and AI Proxy (aiproxy). The owner set that floor on 16 September 2026 so the ranking is over projects enough people have looked at. Their counts are made the same way and shown with them, and a project that crosses the floor joins the table on the next refresh.

Where do the ticks come from?

From each gateway's own documentation, one page per cell, checked by us. We went through the twenty capabilities for each gateway in turn, and every tick and every cross links the page the answer came from, so you can check any cell yourself.

How often does this update?

There is no fixed schedule. The date at the top is the last time the data changed, not the last time the page was built.

Every ranked number links to the official row it came from.