Comparison of AI gateways by jurisdiction, EU residency, model count, providers, OpenAI compatibility, modalities, deployment, observability and company data.
AI gateway/router project with insufficient independently verifiable corporate and catalogue metadata.
ZDR
Azure AI Foundry
Microsoft enterprise AI platform with a very large model catalogue and Azure-native governance/deployment.
Models
Providers
Employees
Certifications
ISO/IEC 27001SOC 2
Jurisdiction
🇺🇸United States
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
LiteLLM
MIT-licensed open-source AI gateway/proxy with broad provider support and customer-configured model catalogues.
Models
Providers
Employees
Certifications
SOC 2 Type IIISO/IEC 27001
Jurisdiction
🇺🇸United States
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
TrueFoundry
Enterprise AI gateway/control plane with 1,000+ LLMs and customer-VPC/on-prem deployment.
Models
Providers
Employees
Certifications
SOC 2
Jurisdiction
🇺🇸United States
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
Eden AI
EU-native multi-provider AI gateway with routing, failover and broad expert-model coverage.
Models
Providers
Employees
Certifications
SOC 2ISO/IEC 27001
Jurisdiction
🇫🇷France
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
Respan
AI gateway/observability product formerly branded Keywords AI.
Models
Providers
Employees
Certifications
Jurisdiction
🇺🇸United States
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
Opper AI
EU-hosted AI gateway for agents with 700+ models, smart routing, fallbacks and regional controls.
Models
Providers
Employees
Certifications
Jurisdiction
🇸🇪Sweden
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
llmgateway.io
Publicly enumerable multi-provider LLM gateway with transparent model/provider catalogue.
Models
Providers
Employees
Certifications
Jurisdiction
🇺🇸United States
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
AI/ML API
Large multimodal model API with OpenAI/Anthropic-compatible interfaces and a broad catalogue.
Models
Providers
Employees
Certifications
Jurisdiction
🇦🇪United Arab Emirates
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
Atlas Cloud
Multimodal inference platform for language, image, video and audio models.
Models
Providers
Employees
Certifications
Jurisdiction
🇺🇸United States
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
Martian
Routing-first model router focused on automatic model selection and cost/quality trade-offs.
Models
Providers
Employees
Certifications
Jurisdiction
🇺🇸United States
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
Edgee
Agent gateway with model routing, provider-route diversity and air-gapped deployment.
Models
Providers
Employees
Certifications
SOC 2
Jurisdiction
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
Requesty
OpenAI-compatible multi-provider gateway with explicit model-versus-endpoint catalogue views.
Models
Providers
Employees
Certifications
ISO/IEC 27001SOC 2 Type II
Jurisdiction
🇬🇧United Kingdom
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
Google Vertex AI
Google Cloud enterprise AI platform with managed model access, regional deployment and Google-native controls.
Models
Providers
Employees
Certifications
ISO/IEC 27001SOC 2
Jurisdiction
🇺🇸United States
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
Novita AI
Hosted inference platform for open models across text, image, video and audio.
Models
Providers
Employees
Certifications
SOC 2
Jurisdiction
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
Cortecs
European sovereign LLM router using exclusively EU-established inference providers.
Models
Providers
Employees
Certifications
ISO/IEC 27001
Jurisdiction
🇦🇹Austria
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
EUrouter
EU-only AI router emphasizing European infrastructure and provider/data sovereignty.
Models
Providers
Employees
Certifications
Jurisdiction
🇳🇱Netherlands
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
Amazon Bedrock
AWS-managed foundation-model platform with regional infrastructure, marketplace and native AWS governance.
Models
Providers
Employees
Certifications
ISO/IEC 27001SOC 2
Jurisdiction
🇺🇸United States
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
Helicone
Open-source LLM observability platform with a gateway/proxy and managed routing.
Models
Providers
Employees
Certifications
SOC 2
Jurisdiction
🇺🇸United States
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
Anannas
OpenAI-compatible gateway with price/latency/throughput routing, fallback and BYOK.
Models
Providers
Employees
Certifications
Jurisdiction
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
Envoy AI Gateway
Open-source AI gateway built around Envoy Proxy and designed for self-hosted infrastructure.
Models
Providers
Employees
Certifications
Jurisdiction
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
Kong AI Gateway
Enterprise API and AI gateway with policy, routing, governance and multi-cloud connectivity.
Models
Providers
Employees
Certifications
Jurisdiction
🇺🇸United States
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
Portkey
Enterprise AI gateway with routing, guardrails, governance and an open-source gateway core.
Models
Providers
Employees
Certifications
SOC 2 Type IIISO/IEC 27001
Jurisdiction
🇺🇸United States
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
RouteScope
AI gateway/router project with insufficient independently verifiable corporate and catalogue metadata.
Models
Providers
Employees
Certifications
Jurisdiction
EU residency
OpenAI compatible
Observability
Modalities
Deployment
ZDR
Model counts are dated snapshots; where available, catalogues are measured directly from public model endpoints, and OpenAI compatibility is read from documentation and shown as a label, never a score. Rows are ordered by model count, largest first, with equal counts in alphabetical order. Each sortable column sorts in one direction only, model count breaks ties, and rows without a value stay at the bottom. Every figure carries how it was established, so an absent number reads as a fact about the product rather than a gap in the research.
# Methodology
How to read the table
No overall score is published and no company is called the best OpenRouter alternative. The table opens ordered by model count, largest first, with equal counts in alphabetical order; each sortable column sorts in one declared direction with model count as the tiebreak; and rows without a value always stay at the bottom. Each column is established the same way for every entry, and where a value cannot be established the cell says so.
## Model
A distinct model addressable through the gateway's public API, after removing duplicates, serving-provider prefixes and routing variants. Measured from a public endpoint wherever one exists.
## Provider
A distinct upstream inference provider or model company reachable through the gateway. Not counted twice for being offered in two regions.
## Route
One model served by one provider. The same model offered by four providers is one model and four routes, so route counts are never shown as model counts.
## Endpoint
An individually addressable API entry as the vendor publishes it. Recorded only where a vendor states one, and never computed from routes or modalities.
## Snapshot date
Every measured figure carries the day it was counted. Catalogues move, so a count is an observation, not an attribute, and superseded observations stay listed in the expanded row and in the changelog.
## OpenAI compatible
Whether the vendor documents an API that OpenAI clients can be pointed at. Yes, partial, no or unknown, read from documentation. A comparison attribute, never a score.
## Jurisdiction vs residency
Jurisdiction is where the operating company is incorporated. EU residency is where requests are processed. Gateway location and inference location are recorded separately again. None is inferred from another.
## Zero data retention (ZDR)
Whether the vendor documents that prompt and response content is not stored once the response has been returned. Read from documentation as yes, configurable, enterprise-only or no, and shown beside deployment because both describe where content can go.
## Unknown values
A cell with no number carries its reason instead: not published, configured by the customer, a different metric, conflicting sources, or simply not yet recorded. A blank is never a zero.
Model and provider countingOne counting rule, applied to every catalogue on the same day.
Where a public model endpoint can be enumerated, the count is measured directly and displayed with the date it was taken. Where only a vendor-stated number exists, it is labelled as such so the two are never confused. The current measured baseline was taken on September 8, 2026 and re-counted on September 15, 2026 across 7 gateways.
A count excludes:
duplicate provider routes for the same base model
quantisations of a model already counted
turbo, fast or preview variants that represent the same base model
deprecated or unavailable models still listed in an endpoint
non-model pseudo-entries such as auto-routing aliases
A vendor floor such as “500+” is stored as a floor and rendered with the plus sign. Where a floor and a measured count both exist, both are shown rather than the larger one being chosen. Where a gateway routes only to providers the customer configures, no catalogue count is published for it; where the software documents how many integrations it ships, that number is shown and labelled as integrations.
Measured and vendor-stated figures are never presented as the same kind of evidence, and a route or endpoint count is never presented as a model count.
Snapshots and historyFigures are dated observations. Refreshes add an observation; they never overwrite one.
Model catalogues are re-measured periodically and each refresh is published as a new dated observation. Company and social data are refreshed together so a snapshot stays internally consistent. Every record keeps every figure on record and the expanded row lists them all, including the vendor’s own number alongside a measurement, so the gap between the two stays visible.
Evidence statusesWhat the small line under each number means.
Every quantitative value carries one of these. The same vocabulary is used throughout the table:
MeasuredCounted by this project directly from the provider's public model catalogue or API on the date shown.
Officially publishedA current figure the provider publishes in its own documentation or product pages. Not a direct measurement, and it may use a different counting rule.
Official catalogueDerived from an official public catalogue that the provider does not market as a headline number.
ReportedFrom a reputable third party, used only because no primary source publishes the figure. Weaker evidence than a measurement or an official statement.
Not publicly listedThe provider publishes no comparable public figure. This is a gap in what the vendor discloses, not a gap in this dataset.
VariableAvailability depends on the providers and endpoints the customer configures. This is a property of the product's architecture, not missing data.
Documented integrationsThe number of provider or model integrations the software documents. The set actually reachable depends on what the operator configures, so the figure is shown but not ranked against hosted catalogues.
Not directly comparableThe provider publishes a figure, but it measures something else — routes, endpoints or provider/model combinations — so it is not comparable with a model catalogue count.
Multiple figuresCredible sources currently report different figures. The disagreement is preserved rather than resolved by picking one without sufficient evidence.
Qualitative fields carry one of these instead:
VerifiedChecked against a primary source: the vendor's own site, documentation, a public API response or a company registry.
Vendor-statedPublished by the vendor but not independently confirmed against a public model catalogue or registry.
EstimatedDerived from a documented range or band rather than an exact published figure.
Not recordedNo value is recorded for this field in the current dataset revision. It is not a judgement about the product.
Not publicly listedThe company publishes no comparable public value for this field.
UnresolvedCredible sources disagree. The disagreement is preserved rather than resolved by picking the more convenient side.
Not disclosedThe company does not publish this information.
Not applicableThis attribute does not apply to this kind of product.
52 fields in the current revision are marked as not recorded rather than estimated. These are open across most of the dataset:
Pricing model detail
Inference regions for managed gateways
Founding years
Route counts, where a vendor publishes one
Registry confirmation for several operating entities
Source hierarchyWhere sources disagree, the higher-ranked source wins.
The product's own API responseA public model endpoint enumerated directly. This is the only evidence that produces a measured catalogue count.
Company registry filingsNational business registers establish the operating legal entity and country of incorporation.
The vendor's own legal pagesTerms, data processing agreements and subprocessor lists, which are contractual rather than promotional.
Product documentationTechnical documentation for regions, deployment options and capabilities.
The vendor's marketing pagesAccepted for product-surface facts and recorded as vendor-stated, never as measured.
LinkedIn company profileUsed for company-size bands and follower snapshots only. Never used to derive a precise headcount.
Competitor blog posts and comparison pagesNot used as evidence for any value in this dataset, in either direction.
Company jurisdiction is established from a registry or the vendor’s own legal pages, never from a website domain, a team page or an office address. Certifications are recorded from the vendor’s compliance pages and marked vendor-stated unless a certificate has been checked; the certified scope matters more than the acronym.
EU jurisdiction, EU residency and locationFour separate fields, populated from separate sources.
EU incorporation is never treated as evidence of EU data residency. The two are separate columns, populated from separate sources.
Jurisdiction is the country in which the operating legal entity is incorporated. Gateway location is where the gateway receives, authenticates and logs a request before forwarding it. Inference location is where the underlying model runs, which is usually set by the upstream provider and often differs per model. EU residency is a documented statement about where requests are processed, recorded with one of these labels. They describe different arrangements, not a scale:
EU by defaultThe gateway processes requests in the EU without the customer having to select a region.
EU availableEU processing is offered as a configurable option on standard plans.
EU routesEU processing exists for a subset of routes or models rather than the whole catalogue.
Enterprise onlyEU processing is available, but only under an enterprise agreement.
Customer controlledThe customer runs the gateway, so residency follows wherever they deploy it.
Not statedThe vendor publishes no claim about where requests are processed.
Not establishedNo residency claim has been read from the vendor's own documentation for this dataset. It is not a statement that the vendor lacks EU processing.
OpenAI compatibilityRead from documentation, shown as a label, never ranked.
Compatibility with the OpenAI API decides how much client code moves unchanged, so it is recorded as its own attribute with one of four labels. It is established from the vendor’s own documentation or from an endpoint this project exercised, and it is never inferred from the product category.
YesThe vendor documents an OpenAI-compatible API, so clients and SDKs written against the OpenAI API can be pointed at the gateway by changing the base URL and key.
PartialCompatibility is documented for a subset of the OpenAI API only — for example chat completions but not the responses, embeddings or audio endpoints — or requires a compatibility layer.
NoThe gateway exposes its own API and documents no OpenAI-compatible endpoint. Existing OpenAI clients need an adapter or a rewrite.
UnknownThe vendor's documentation was checked and does not settle whether the API is OpenAI-compatible.
A gateway whose documentation has not been checked yet shows “Unknown” with a not-recorded status, which is a different statement from a documented unknown.
ObservabilityFive levels of built-in observability, read from documentation and never scored.
How much a gateway lets a team see of its own traffic decides how quickly cost, latency and failure questions can be answered without a second tool. It is recorded as one of five levels, established from the vendor’s own documentation and product material rather than by exercising the dashboards. The levels describe how deep the built-in tooling goes; they are not a quality score, and nothing on this site is ranked by them.
NoneNo meaningful built-in observability found.
BasicBasic usage and billing information.
LimitedUseful logs, metrics or usage, but relatively narrow.
DetailedRequest-level logs plus cost, token and latency analytics, with useful filtering and breakdowns.
AdvancedDetailed observability plus traces, custom dimensions, dashboards, alerts, evaluation or debugging tooling.
A self-hosted gateway that exposes metrics and traces to the operator’s own stack is rated on what it exposes, even where there is no hosted dashboard. Where the public material does not establish a richer layer, the level is the most it supports and is marked estimated rather than raised on assumption.
Company, employee and social dataSize bands, not headcounts. Snapshots, not rankings.
Company size uses LinkedIn company-size bands such as 11–50. Precise headcounts are not reconstructed from third-party databases, because those figures are themselves estimates presented as facts. A band describes the company, not the product.
LinkedIn and X follower figures are point-in-time snapshots, captured on the same day so the pair stays comparable, and displayed rounded by scale. They live in a Traction block in each expanded row, not in the table.
Follower count is not a measure of product quality and is never used to order any list on this site.
Ownership, parent company and acquisition status are recorded where a filing or a vendor announcement establishes them, because an acquisition can change a product’s roadmap, pricing and data handling without changing its website. Two product names belonging to one company are recorded as a single entry.
Funding is recorded as the number of disclosed financing rounds and the investors named in them, from the company’s own announcements or, where it publishes none, from a named company database. Grants, strategic investments outside a disclosed round and acquisitions are described rather than counted. Funding that could not be publicly verified is shown as not publicly listed, never as zero, and a cloud provider’s product or a community project records the field as not applicable, because a corporation’s history is not the funding of one product.
Corrections and biasEvery vendor is measured with the same rule. Corrections are dated and reviewable.
Vendors and readers can request a correction through the public repository. Send the gateway, the field, the corrected value and a primary source: a public endpoint, a registry entry, a legal page or product documentation. A competitor’s comparison article is never accepted as evidence for a rival’s figures, in either direction.
Historical figures are never silently overwritten. A superseded measurement stays listed in the expanded row next to the one that replaced it, and the change is logged.
The protections against bias are structural rather than promised: no composite score exists and every order on the site is a single stated column, the table opens on model count, every sortable column sorts in one declared direction with unrecorded values at the bottom, measured and provider-stated figures are always labelled apart, and no gateway is excluded from a filter it qualifies for.
# Open project
How to contribute
The dataset, the methodology and the site are maintained in the open. Anyone can propose a correction or an addition, and every change is reviewed against a primary source before it becomes canonical.
A competitor’s comparison page is never accepted as evidence, in either direction. Registry filings, legal pages, public endpoints and product documentation are.
Correct inaccurate dataOpen an issue or a pull request against the record, with the field, the corrected value and a primary source.
Add a missing gatewayA multi-provider gateway, router or AI API that is not tracked yet. Only the fields you can source need to be filled.
Improve a sourceReplace a vendor marketing claim with a registry filing, legal page, public endpoint or documentation link.
Report outdated informationCatalogues, ownership and residency options change. Point at what moved and when you observed it.
Improve the siteAccessibility, performance, table behaviour, or the data-collection scripts. The stack is Next.js and TypeScript.