# Alphe

> Alphe is an OpenAI-compatible API gateway that chooses the model for each
> request instead of leaving it pinned in your code. Every inbound call is
> classified, priced against the candidates, routed to the cheapest model that
> still clears the quality bar you set for that endpoint, checked against your
> rubric, and logged with the reason it went where it went.

Alphe AI, built in India, pre-launch as of August 2026. Two founders: Ankit
Kumar Verma, who owns classification, routing policy and the model catalogue,
and Anant Gupta, who owns the data plane and caching. Contact:
hello@alpheai.com.

What it does, in the order a request meets it: classification (task type,
reasoning depth, context length, tool surface — roughly 4 ms added), routing
against live per-provider token pricing and measured per-task-class quality,
verification of the answer against a per-endpoint rubric with automatic
escalation to a stronger model on a miss, semantic caching, failover when a
provider degrades, and cost attributed to a team, feature and customer rather
than to one line on an invoice.

Compatibility: the OpenAI and Anthropic SDKs work by changing the base URL, so
anything accepting a custom base URL — LangChain, LlamaIndex, Vercel AI SDK —
already works. Bring your own provider keys or use Alphe's. Run the proxy
hosted, in your own VPC, or self-hosted.

Pricing: $0.20 per 1,000 routed queries, metered, no seats and no minimum.
Provider token costs pass through at cost on every plan. A custom plan is
priced as a share of measured savings or as a flat annual fee.

## Pages

- [Home](https://alpheai.com/): the routing argument, the live model coverage
  table, the savings calculator, and the security posture.
- [Platform](https://alpheai.com/platform/): the five stages — classification,
  routing, verification, caching, observability — and what each one costs in
  latency.
- [Pricing](https://alpheai.com/pricing/): both plans, the calculator, and six
  answers about how billing actually works.
- [Docs](https://alpheai.com/docs/): how to call Alphe — the base URL swap,
  what `model: "auto"` means, quality bars and rubrics, the decision record,
  caching, failover, deployment modes, limits, security, billing mechanics and
  eight answers for integrators.
- [About](https://alpheai.com/about/): why the company exists, the six
  principles it is built on, and the two people who build it.
- [Contact](https://alpheai.com/contact/): shadow-mode signup and what the
  first four steps look like.

## When to use Alphe

Use it when the answer to "which model should this call go to?" is different
per request, or when it stopped being obvious:

- An inference bill is growing faster than usage and nobody can say which
  feature spends what.
- One model is pinned across every call, so the cheap requests pay frontier
  prices and the hard ones occasionally get a model that is too small.
- A provider outage becomes your outage.
- You need to know why a request went where it went, after the fact.
- Prompts cannot leave your network, so the gateway has to run inside it.
- You want a measured number before changing anything: shadow mode mirrors
  traffic and reports what routing would have cost.

Do not use it as a model host (Alphe calls providers, it does not serve
weights), for training or fine-tuning, as a vector database, as an agent
framework, or on a path where a few milliseconds of added latency is the
binding constraint.

How to call it: point an OpenAI-compatible client at
`https://api.alpheai.com/v1`, send `model: "auto"`, keep the rest of the
request as it is. Keys are issued during early access — there is no self-serve
signup, so a key comes from hello@alpheai.com rather than a dashboard. Full
instructions for agents: https://alpheai.com/agents.md

## Machine-readable

- [/agents.md](https://alpheai.com/agents.md): when to use Alphe, when not to,
  and how an agent should call it.
- [/llms-full.txt](https://alpheai.com/llms-full.txt): the full text of every
  page in one file.
- [/sitemap.xml](https://alpheai.com/sitemap.xml): every indexable URL.
- [/robots.txt](https://alpheai.com/robots.txt): crawl policy. Nothing is
  disallowed.
- Every page is also served as Markdown at its own URL: send
  `Accept: text/markdown`, or append `index.md` to the path
  (`/docs/index.md`). The site sets `Vary: Accept`, honours q-values, and
  returns 406 for an Accept it cannot satisfy. A missing path returns a real
  404 whose body is a Markdown site map.

## Notes

- There is no endpoint-by-endpoint API reference yet. Alphe is OpenAI
  wire-compatible, so the request and response shapes are the ones your SDK
  already documents; /docs/ carries everything else. The full reference lands
  at general availability and belongs in this file when it does.
- The one board on the home page — a single research question answered five
  ways — is Alphe's own measurement, graded by one grader against one task, and
  it is labelled as such on the page. There is no third-party leaderboard on the
  site at the moment.