An LLM gateway is a service that sits between your application and model providers, giving your code one API for many models while it handles keys, routing and failover, request logs and spend limits. This comparison covers six gateways teams shortlist in 2026, using the same checklist for each, with every pricing and feature claim linked to the vendor's own documentation as of October 2026.
What an LLM gateway does
Five jobs, in the order a request meets them:
- One API for many models. Your code calls one base URL in a familiar format (usually OpenAI's) and switches models by changing the
modelfield. - Authentication and keys. The gateway issues its own keys and keeps provider credentials out of your apps.
- Access rules. It decides whether this key may call this model.
- Routing and failover. It picks a provider for the model and retries elsewhere if one fails.
- Metering, logs and limits. It records tokens and cost per request and stops traffic when a budget is reached.
On ATP Token those stages map to status codes you can test: 401 for a bad key, 403 when the model isn't allowed for the project, 502 or 503 when providers fail, and 402 when the project's credits run out (how it works, errors).
The six gateways at a glance
"Not compared here" means we did not find the detail in the vendor's public docs; check with the vendor before you rely on it.
How to choose: six questions
1. Where must the gateway run?
If traffic cannot pass through a third party, the shortlist is self-hosted: the LiteLLM proxy, Helicone's open-source gateway, or Portkey's open-source and VPC options. Everything else on this page is hosted.
2. How do you want to pay?
Three models exist:
- Provider price plus a fee on top-ups. OpenRouter: 5.5% on card purchases on Standard ($0.80 minimum), 8% on Business (FAQ).
- Provider list price, no token markup. Vercel AI Gateway; payment processing fees may apply (pricing).
- Prepaid credits at per-model list rates. ATP Token: 1 credit = USD 0.01, pay-as-you-go credits don't expire, top-ups from USD 5 (credits). Rates per model are on the pricing page.
Self-hosted gateways add no fee, but your provider bills and engineering time are the cost.
3. What is a budget attached to, and is it hard?
This is where gateways differ most.
- OpenRouter: budgets in guardrails per member or key, resetting daily, weekly or monthly; workspace budgets on Enterprise (guardrails).
- Vercel: team, project, key and member budgets. Vercel calls them a soft cap: the request that crosses the limit still completes, then requests get 402 (budgets).
- LiteLLM: budgets per key, user, team or customer with reset durations like
30d. - ATP Token: credits are allocated down organization → workspace → project, and a project can only spend its allocation. There is no reset period; the project runs until its balance is used, then gets 402 until someone allocates more (budget caps recipe).
A reset-period budget suits "no more than $500 a month on this key". A prepaid allocation suits "this product line has $5,000 for the quarter".
4. Which SDK formats do your apps already use?
Most gateways speak the OpenAI format. If some services use the Anthropic SDK (Claude Code does) or the Google GenAI SDK, check native support: OpenRouter and Vercel document Anthropic Messages; ATP Token accepts the OpenAI, Anthropic and Google GenAI SDKs unchanged (OpenAI SDK, Anthropic SDK, Google GenAI SDK).
5. Which models and modalities?
OpenRouter lists 500+ models from 80+ providers. ATP Token has 70+ models from 11 vendors, including video models such as Seedance 2.0 and Kling v3 Pro and image models such as Nano Banana Pro (media models). If you only need a handful of frontier LLMs, catalog size matters less than the other questions.
6. What does the log keep, and for how long?
Portkey publishes retention per plan: 3 days of logs on the free tier, 30 days on Production. ATP Token keeps request logs for 7 days as a debugging view, with billing events as the billing record (request logs API). If you review spend monthly, export or summarise weekly.
Each gateway in more detail
OpenRouter
- Fits: the widest hosted catalog and fast model evaluation.
- Controls: organizations, workspaces, guardrails with budgets, model and provider allowlists, ZDR rules; per-key credit limits on all plans.
- Watch out for: the fee on credit purchases, possible credit expiry after one year, invoicing only on Enterprise. Full breakdown in OpenRouter alternatives for teams.
Vercel AI Gateway
- Fits: teams on Vercel, or anyone wanting a hosted gateway priced at provider list price.
- Controls: four budget scopes, spend alerts at 50/75/100%, request logs, provider and model fallbacks.
- Watch out for: soft-cap budgets; spend on your own provider keys isn't counted against budgets; some controls are paid add-ons.
LiteLLM
- Fits: platform teams who want full control and already run infrastructure.
- Controls: virtual keys, budgets at several levels, access groups for models.
- Watch out for: you operate it. Per-model budgets at user and key level need an Enterprise license.
Portkey
- Fits: teams that want a gateway with observability and guardrails in one product.
- Controls: budget limits on providers or virtual keys with alert thresholds and weekly or monthly reset, on Enterprise and select Pro plans (budget limits). RBAC from the $49 Production plan.
- Watch out for: log quotas per plan (10k logs a month free, 100k on Production).
Helicone
- Fits: observability-first teams; the gateway is open source and written in Rust.
- Controls: request logging with tokens and cost, routing, fallbacks, caching across 100+ models.
- Watch out for: check budget and access-control features against your requirements; we did not find them in the gateway overview.
ATP Token
- Fits: companies where several teams share one AI budget and each project should have a funded allocation and a reviewed model list.
- Controls: organization → workspace → project hierarchy, project-scoped keys, per-project allowed models, credit allocation as the ceiling, per-request logs, owner/admin/member roles (team and roles).
- Watch out for: a smaller catalog than OpenRouter; credits are non-refundable; request logs are kept 7 days.
Running two gateways
Many teams end up with two: a large-catalog gateway for research and evaluation, and a governed one for production. The rule that keeps this sane is that production keys never live in the research account, and each production service has its own key and budget. See one project, one key.
Trying the comparison on ATP Token
- Create a project and enable two or three models you want to compare, for example claude-sonnet-4-6, gemini-3-5-flash and deepseek-v4-flash.
- Allocate a small amount of credits, say 500 (USD 5), as the ceiling for the test.
- Point your existing OpenAI, Anthropic or Google GenAI client at the gateway and run the same prompts against each model.
- Compare tokens and credits per request in Request logs (usage and logs). Side-by-side model pages are at /compare.
