Guide
LLM gateway model counts: what “500+” means
Short answer
LLM gateway model counts are not directly comparable because vendors count different objects: model families, versions, aliases, provider routes or deployments. A range in this catalogue preserves recorded differences between sources or scope. Choose using the exact model identifiers, regions and capabilities your application needs, not the largest headline total.
Count the unit your application depends on
| Unit | Meaning |
|---|---|
| Family | A product lineage; it can contain many versions or sizes. |
| Version | A specific release, ideally pinned rather than a moving default. |
| Alias | A name that may resolve to another model identifier over time. |
| Deployment | A hosted instance or region with its own capacity and availability. |
| Provider route | A way to reach a model through a particular upstream account or host. |
| Listing | A catalogue entry; not proof of access on your plan. |
Build an approved inventory with model ID, pinned version, required modality, region, context limit, tool/JSON support, price, plan entitlement and fallback route. Check it against the model directory and a small application acceptance suite. Mark aliases and previews explicitly; record the check date and deprecation policy.
Why recorded ranges differ
Coverage is recorded as low and high bounds, with a source note describing what was counted. A range may reflect conflicting published totals, different dates or an enumerated list with narrow and broad scopes. Only contradictory claims about the same scope support a conclusion of vendor disagreement.
- 1 Two model figures
- The recorded low and high bounds differ. Check whether the source note describes a scope difference, an enumerated range or conflicting published claims.
- 11 Two provider figures
- Same again for upstream providers, and more often. A marketing figure and an enumerated list on a docs page are rarely the same length.
- 8 No model figure
- No usable model total published at all. Several of these are proxies with no catalogue of their own, where the answer genuinely depends on your upstream accounts. Recorded as unpublished, not as zero.
- 11 No provider figure
- No usable provider total. For some of these the unit does not exist, which the next section is about.
Model coverage splits 21 single figure, 1 range, 1 one-sided bounds and 8 unpublished. A single figure is not the same as a reliable one — it means one number, on one reading, checked on one date — but it is at least a number the vendor states consistently.
The widest published spread for models is Portkey, at 250 to 2,300. Both bounds are defensible: the low one counts what you can call through the chat endpoint, the high one counts everything reachable including adapters and non-text models. There is no arithmetic that turns those into one honest figure, which is the point. The widest provider spread is MLflow AI Gateway, at 14 to 100, where an enumerated docs table and an interface hint disagree by a factor of several.
What to do with this: read the low bound as the smallest recorded claim, not a service guarantee, and investigate whether the range reflects different scopes, dates or conflicting documentation. Then stop using either one to choose. The source notes on 10 of the 31 product pages record the vendor's own pages contradicting each other, with links, so a specific question to a salesperson is available to you instead of a general one.
Round numbers need a counting definition
4 products in this catalogue report the same headline figure of 100 models. That pattern alone does not prove rounding or inaccuracy. Of the 23 products that publish a headline model figure at all, 9 publish a multiple of a hundred, and 4 of them share a value with at least two competitors.
- 100 models — claimed by 4 products
- 4 separate vendors arriving at the identical figure is not 4 vendors who counted. It is the nearest round number above whatever the catalogue held on the day the marketing page was written, which is why the same value recurs and why it rarely matches a live model listing.
There is a second reason a single total cannot exist for most of these products. 28 of 31 expose more than one API surface — chat completions alongside embeddings, images, speech, reranking or a management API — and New API documents 13 . A count that spans all of them answers a question nobody asked. A count restricted to chat completions is smaller, more useful, and not what the marketing page says. Being OpenAI-compatibleOpenAI-compatible: Speaks the same API shape as OpenAI’s, so existing client libraries work by changing a base URL. This is what makes switching gateways cheap — and worth insisting on. makes the surfaces look alike without making the catalogues behind them comparable.
“Providers” is not one unit
A provider count sounds like the most solid figure in the category: upstream inference vendors are discrete, nameable things, and you cannot round them. In this catalogue the word is used for several different objects, and the products themselves are the ones using it loosely.
- The unit does not apply — 6
- First-party platforms that serve models on their own infrastructure rather than routing to third-party APIs: Amazon Bedrock, Azure AI Foundry, Fireworks AI, Google Vertex AI, Groq and Together AI. For these, a provider count is a category error rather than a low score. What they publish instead is a count of model families or partner labs in a marketplace, which is a different measurement wearing the same label.
- Model-family prefixes counted as providers — 1
- Merge Gateway publishes no provider total, and the documentation uses the word in two senses: the prefixes in the model identifiers, which name the lab that trained the model, and the execution host that actually serves the request. The first number is much larger than the second and is the one a reader would assume. The source note on the product page records both, with the pages they came from.
- Integrations counted as providers — 1
- Eden AI reaches its upper bound only by including integrations that have nothing to do with routing a chat completion — document parsing, translation, speech and similar expert services. The lower bound counts the providers you can send a language-model request to. If you are choosing on language-model breadth, the lower bound is your number.
- Upstream inference vendors, as advertised
- The remainder, where the figure does mean what it appears to mean, and where a docs table you can count is usually shorter than the headline. This is the only group where comparing two provider counts tells you something, and even there it does not tell you whether the one provider you depend on is among them.
The practical consequence is narrow and worth acting on: never compare two provider counts without reading what each one counts. Every product page here carries the source note for both figures, which is the shortest route to finding out.
What to count instead
Three questions replace the total. All three are answerable, all three change a migration plan, and none of them is a single number you can put on a homepage.
- Are the models you call today on the list, at the versions you pin?
- Not the families — the identifiers. A catalogue that carries a model family but not the dated snapshot your evaluations are pinned to will fail your regression suite on the first day. Ask for the specific identifiers in writing, and ask what happens to a pinned version when the upstream deprecates it.
- Can you add an endpoint the vendor has never heard of?
- This is the answer that makes a catalogue total irrelevant, because a product that takes an arbitrary base URL has no ceiling. 15 of 31 document a way to register your own endpoint, 5 document something partial, and 5 publish nothing about it either way. A further 4 offer the adjacent thing rather than this one — AI Gateway HQ, Fireworks AI, Respan and Together AI will host weights you upload, which is not the same as calling a URL you already run. One product documents an explicit no, which is more useful to a reader than silence.
- When a new model ships, do you wait for the vendor?
- A gateway with a curated catalogue adds models on its own schedule, and that schedule is rarely published. If you cannot register the endpoint yourself, the honest form of the coverage question is not how many models exist today but how many days behind the upstream you will be. Where a vendor states a target — some do, in days — it is recorded in the model count note on its product page.
One more field changes how much any of this matters. 20 of 31 advertise rule-based routingConditional routing: Rules that pick the model per request — cheap model for simple work, expensive one for hard work, a specific model for one customer., which lets you pin a model per request or per rule rather than accepting a default. The remaining 11 publish nothing about it, which is recorded as unpublished rather than as a no. If you can pin a model, the catalogue total is a footnote: you are calling three models and want to know that those three are healthy. If you cannot, you inherit whatever the default is on the day, and the total starts to matter for the wrong reason.
| Product | Models | Providers | Custom endpoints |
|---|---|---|---|
| agentgateway | ~1,002 Low: Check date not recorded · High: Check date not recorded | 20–44 Low: Check date not recorded · High: Check date not recorded | Documented |
| AI Gateway HQ | Not published Low: Check date not recorded · High: Check date not recorded | 11 Low: Checked 2026-09-17 · High: Checked 2026-09-17 | Weights, not endpoints |
| Amazon Bedrock | ~100 Low: Check date not recorded · High: Check date not recorded | Not published Low: Check date not recorded · High: Check date not recorded | Partly documented |
| Apache APISIX AI Gateway | Not published Low: Check date not recorded · High: Check date not recorded | 10–20 Low: Check date not recorded · High: Check date not recorded | Documented |
| Azure AI Foundry | ~10,000 Low: Check date not recorded · High: Check date not recorded | Not published Low: Check date not recorded · High: Check date not recorded | Partly documented |
| Bifrost | Not published Low: Check date not recorded · High: Check date not recorded | 20–23 Low: Check date not recorded · High: Check date not recorded | Documented |
| Braintrust Gateway | ~100 Low: Check date not recorded · High: Check date not recorded | 9 Low: Check date not recorded · High: Check date not recorded | Documented |
| Cloudflare AI Gateway | Not published Low: Check date not recorded · High: Check date not recorded | 24 Low: Check date not recorded · High: Check date not recorded | Documented |
| Eden AI | 1,104 Low: Checked 2026-09-25 · High: Checked 2026-09-25 | 33–68 Low: Check date not recorded · High: Check date not recorded | Documented as no |
| Envoy AI Gateway | Not published Low: Check date not recorded · High: Check date not recorded | 16–19 Low: Checked 2026-09-02 · High: Checked 2026-09-02 | Documented |
| Fireworks AI | ~27 Low: Checked 2026-09-26 · High: Checked 2026-09-26 | Not published Low: Check date not recorded · High: Check date not recorded | Weights, not endpoints |
| Google Vertex AI | ~200 Low: Check date not recorded · High: Check date not recorded | Not published Low: Check date not recorded · High: Check date not recorded | Documented |
| Groq | 11 Low: Checked 2026-09-26 · High: Checked 2026-09-26 | Not published Low: Check date not recorded · High: Check date not recorded | Not documented |
| Helicone | ~100 Low: Check date not recorded · High: Check date not recorded | 20–100 Low: Check date not recorded · High: Check date not recorded | Not documented |
| Higress | ~100 Low: Check date not recorded · High: Check date not recorded | 26–31 Low: Check date not recorded · High: Check date not recorded | Documented |
| Hugging Face Inference Providers | 136 Low: Checked 2026-09-23 · High: Checked 2026-09-23 | 14–18 Low: Check date not recorded · High: Check date not recorded | Not documented |
| Kong AI Gateway | Not published Low: Check date not recorded · High: Check date not recorded | 17 Low: Check date not recorded · High: Check date not recorded | Documented |
| LiteLLM | Not published Low: Check date not recorded · High: Check date not recorded | Not published Low: Check date not recorded · High: Check date not recorded | Documented |
| LLM Gateway | ~200 Low: Check date not recorded · High: Check date not recorded | ~40 Low: Check date not recorded · High: Check date not recorded | Partly documented |
| Merge Gateway | 3 Low: Checked 2026-09-26 · High: Checked 2026-09-26 | 4–24 Low: Checked 2026-09-02 · High: Checked 2026-09-02 | Not documented |
| MLflow AI Gateway | Not published Low: Check date not recorded · High: Check date not recorded | 14–100 Low: Check date not recorded · High: Check date not recorded | Documented |
| New API | 100+ Low: Check date not recorded · High: Check date not recorded | 30+ Low: Check date not recorded · High: Check date not recorded | Documented |
| OpenRouter | 458 Low: Checked 2026-09-25 · High: Checked 2026-09-25 | ~83 Low: Check date not recorded · High: Check date not recorded | Partly documented |
| Orq.ai Router | ~500 Low: Check date not recorded · High: Check date not recorded | Not published Low: Check date not recorded · High: Check date not recorded | Documented |
| Portkey | 250–2,300 Low: Check date not recorded · High: Check date not recorded | 40–48 Low: Check date not recorded · High: Check date not recorded | Documented |
| Requesty | 706 Low: Checked 2026-09-14 · High: Checked 2026-09-14 | Not published Low: Check date not recorded · High: Check date not recorded | Partly documented |
| Respan | 1,525 Low: Checked 2026-09-23 · High: Checked 2026-09-23 | Not published Low: Checked 2026-09-15 · High: Checked 2026-09-15 | Weights, not endpoints |
| Together AI | 272 Low: Checked 2026-09-26 · High: Checked 2026-09-26 | Not published Low: Check date not recorded · High: Check date not recorded | Weights, not endpoints |
| TrueFoundry AI Gateway | ~1,000 Low: Check date not recorded · High: Check date not recorded | 27 Low: Check date not recorded · High: Check date not recorded | Documented |
| Velokey | 94 Low: Checked 2026-09-26 · High: Checked 2026-09-26 | 20+ Low: Checked 2026-09-19 · High: Checked 2026-09-19 | Not published Check date not recorded |
| Vercel AI Gateway | 386 Low: Checked 2026-09-23 · High: Checked 2026-09-23 | Not published Low: Check date not recorded · High: Check date not recorded | Not documented |
A blank reads as not published rather than as no, and a range means the vendor publishes both bounds itself. The custom-endpoint column summarises the vendor’s own leading verdict in its documentation and nothing more; the full wording, with the pages it was read from, is on each product page, because the difference between “yes, for OpenAI-compatible URLs” and “yes” is usually where a migration stalls.
This site publishes these numbers too
GatewayScore publishes these counts with the same limitations. The range and source notes preserve uncertainty rather than resolving it silently. Use the catalogue snapshot for the recorded fields and each provider profile for the scope note; do not turn the range width into a quality score.
The snapshot retains legacy single-count fields alongside the range fields. Differences can reflect source dates, scope or earlier research; they do not by themselves establish a vendor contradiction. Public comparison cells use the range, and their low/high evidence must be inspected separately.
What follows from that is a limit on what this site will do with these fields. Coverage counts are used for the low bound only when something has to be sorted, and never as a score, because a metric this soft cannot carry a ranking. The methodology sets out how each value is verified, what counts as a source, and which fields are deliberately excluded from scoring. If you find a figure here that its vendor contradicts, the corrections page is the fastest way to have it changed.
Reading the number
The number matters if
- You are exploring, and have not decided which models you need.
- You want one model reachable through several upstreams for redundancy.
- Your workload spans text, images, speech and embeddings at once.
- You cannot register an endpoint yourself, so the catalogue is the ceiling.
Ignore it if
- You call three models and pin their versions.
- You can add your own endpoint, which removes the ceiling.
- Your models are self-hosted and the gateway only fronts them.
- You are comparing two figures that count different objects.
Ask instead
- Are these specific identifiers, at these versions, available today.
- What exactly does your provider figure count.
- How many days after an upstream release does a model appear here.
- Can I register a URL you do not carry, and on which plan.
Common questions
Why do two vendors with the same models publish very different model counts?
Because they are counting different objects. One counts every routable identifier, so the same model reached through three upstream hosts counts three times. Another counts distinct model names. A third counts everything in a catalogue table, including embedding, image, speech and reranking models, and a fourth counts only what is available on its serverless tier. None of these is dishonest on its own, and no two of them are comparable.
Is a bigger catalogue better?
Only up to the handful of models you will actually call. Past that, breadth is insurance against a model being deprecated or a provider having a bad week, and the useful form of that insurance is whether you can reach the same model through more than one upstream, not whether the total is four figures. A catalogue of thousands that does not carry the one model your evaluation suite is pinned to is worth less than a catalogue of twenty that does.
What does a range in the models column mean?
That the vendor publishes more than one number for itself, or that the number depends on what is being counted, and this catalogue records both bounds rather than picking one. The low bound is the smallest recorded claim under the cited scopes, not a contractual availability guarantee. Each product page carries the source note that says which pages the two bounds came from, so you can see the disagreement rather than inherit a resolution of it.
A vendor publishes no model count at all. Is that worse?
It is not inherently better or worse. Several products here are proxies with no catalogue of their own: the models available are whatever the upstream accounts you configure expose, so a total would be a figure about your account rather than about the product. That is recorded as unpublished, not as zero, and the question to ask instead is whether you can register an endpoint the vendor has never heard of.
How should I check coverage before committing?
Write down the model identifiers you call today, with versions, and ask for each one in writing rather than asking for a total. Then ask two follow-ups: how long after a model ships upstream does it appear here, and can I add an endpoint you do not carry. Those three answers decide whether coverage will hurt you. The headline number decides nothing.