One endpoint in front of 4,500+ models. Alphe picks the cheapest model that still clears your quality bar, and cuts up to 70% of your inference bill.
Drop-in OpenAI-compatible endpoint · No rewrites · Live in under five minutes
Which model, which tool, whether one of your own agents already does the job: a space of answers, not a single choice. Alphe scores the whole space first and calls the cheapest route that still clears your bar. The other routes are what you would have paid for.
Read the last 40 support tickets, group them by theme, and open a Linear issue for each theme.
9 ms to score them. Nothing above has been called yet, so nothing above has been billed.
A real request, scored and priced in the open. Every number below is one the router had to produce before it was allowed to send anything.
Plan Score, price, send cheapest pass
Holds Strict JSON · eu-west-1 only
Tag 12,000 support tickets by intent and escalate anything angry.
Alphe classifying
Scored as bulk classification. Shallow reasoning, 1.2k median context, strict JSON out.
Alphe pricing candidates
Four models clear the quality bar. Haiku 4.5 is the cheapest of them at this context length.
12,000 tagged, 341 escalated. $11.42 spent against $412 on the frontier model you were using.
200 OK · 940 ms
const client = new OpenAI({ baseURL: "https://api.alpheai.com/v1", apiKey: process.env.ALPHE_KEY, });
Change the base URL. Nothing else. Alphe speaks the OpenAI, Anthropic and raw HTTP shapes, so existing SDKs, agent frameworks and eval harnesses keep working.
Scored before anything expensive sees the request
A small resident classifier reads the request before anything expensive sees it: task type, reasoning depth, context length, tool surface, output contract.
312 candidates repriced as providers move
Every candidate model is priced against the live token cost, its measured quality on that task class, and its current latency. The cheapest one that clears your bar wins.
re-run passed · table updated
Outputs are scored against your rubric, not assumed correct. A response that misses is escalated to a stronger model automatically and the routing table learns from it.
1,284 tokens · $0.0009 · 940 ms
The answer comes back with the semantic cache warmed, the trace written, and cost attributed to a team, a feature and a customer. You can answer "what did this feature cost last week" without instrumenting anything.
Models are the part everyone talks about. Alphe routes across agents, tools and workflows on the same call, weighs what each candidate costs and how fast it answers, then sends the request to whichever one clears your bar for the least money.
None of it is a separate integration to wire up. One endpoint, one key, and the decision made per request.
One prompt. Several tools, more than one model, a receipt at the end. Pick a job and watch which route it takes.
Read the 74 vendor contracts in Drive and flag every auto-renewal clause.
Scored as bulk extraction over long context. gemini-2.5-flash clears the bar at a fortieth of the frontier price, so 71 of the 74 go there.
Receipt $1.14 · 3 m 12 s · $47.60 if every page had gone to one frontier model
Every Monday, post last week's spend by team to #eng.
Forty-one rows to summarise and nothing to reason about, so this stays on gpt-4o-mini. The whole digest costs four hundredths of a cent.
Receipt $0.0004 · 1.9 s · on a schedule, so nobody has to remember
Migrate the monorepo off SDK v2. Open one pull request per package.
Mechanical rewrites go to qwen3-coder at $0.30/M. Four packages touch the auth surface. Those are worth claude-opus-5.
Two suites fail on the same fixture. Both packages re-routed to opus, re-run, green.
Receipt $18.90 · 24 m · $214 all-frontier. Same diff, same tests.
Answer docs questions inside the app. Nothing slower than 400 ms.
Latency is the constraint here, not depth. claude-haiku-4.5 answers 94% of these inside the budget.
Receipt $0.0021 median · 210 ms p50 · 380 ms p95
One request rarely means one kind of work. Alphe splits it: the images and video one way, the documents another, the code a third. Each part is charged to the workspace and provider account that owns it, before it picks a model.
Your tools, agents and workflows are connected to Alphe, not to a model. When something new wins at code review or long-context extraction, switch. The whole stack comes with it.
One question, five answers, one grader. Six weeks of a running flood, asked from the first breach to the day of the run, and every clause of it wants a different source. Alphe routes it for $0.0074 and grades 6.5 for accuracy and 7.5 for how it reads. The sharpest answer is two points better on accuracy at 65× the bill; the best-written one is a point and a half better at 95×. The fastest answer lands in 18 seconds and scores 3 out of 10 for being worth reading.
Accuracy points returned per dollar spent on this question. Alphe returns 878 of them; the best single model returns 130. Column height is logarithmic: the spread is 77×.
Alphe's own measurement — one question, one grader, five ways of answering it. Accuracy and quality are 0–10 grades; cost is the whole bill for the run, provider tokens included; wall clock is time to the finished answer.
A gateway gives you every model behind one key but no opinion about which one to call. A framework gives you agents and workflows but leaves the model, the tool and the bill to you. Going direct gives you neither. Alphe is the decision layer over all four, and the price of the answer is part of the decision.
| Capability | Best Alphe One decision layer | Model gateways OpenRouter Together | Agent frameworks LangChain CrewAI | Direct to provider OpenAI Anthropic |
|---|---|---|---|---|
| Route across models Pick the model per call, not per project. | ||||
| Route across agents Hand the job to the agent that handles it. | ||||
| Route across tools One catalogue, called on your behalf. | ||||
| Route across workflows Multi-step runs picked the same way. | ||||
| Picks the tool for the job Selection, not a list you maintain. | ||||
| Price is part of the decision Cheapest route that still clears the bar. | ||||
| Fails over mid-request A provider going down is not your outage. | ||||
| One key for every provider No per-vendor accounts to keep alive. | ||||
| Drops into code you already wrote Same request shape, one base URL changed. |
Scroll the table sideways for the rest →
Marks describe the category, not any one product in it. The named examples are there to say which category is meant.
Three compounding levers: routing each call to the cheapest model that clears your bar, serving semantic cache hits instead of re-billing near-identical prompts, and compressing context that never influenced the answer.
70% reduction: routing, semantic cache hits and prompt compression, measured against your current provider mix.
Models are benchmarked and added to the routing table as they ship, behind the same key, the same bill and the same rate limit. You deploy nothing.
+ 4,500 models on the same routing table
Handing a model the whole catalogue is how it picks the wrong thing. Alphe ranks 7,000+ integrations against the request, passes on the shortlist, and runs the calls in order, the same selection it does for models, one layer down.
Early access is open. Point one endpoint at Alphe, keep your code, and read the difference off your own invoice.