Crossroad

A decision model for software

Get a free key →Read the docs →

  1. 1Abstract
  2. 2Why Crossroad
  3. 3Try it out
  4. 4Results
  5. 5Interface
  6. 6Pricing
  7. 7Free tier
↓

Crossroad answers a question about a piece of context with a value your code can use: one option from a list, a boolean, or a number on a scale. Send the context and the questions in one request; receive the answers, checked against the questions you asked, each with the probability of every option, and the token count. The model scores the options rather than writing about them, in one pass.

Why Crossroad

Answers your code can use. Route, classify, tag or score: each answer is one of your options, a boolean or a number in your range, with the probability of each. Crossroad reads text; Crossroad M also reads one image or a recording of up to 30 seconds.

Fast. The model scores every option in one pass instead of writing an answer, so a short request usually comes back in under a second.

Free, within fair limits. Requests aren’t billed for now, and every account starts with the same limits, so no one account takes over the GPUs.

Try it out

Options

Results

BTZSC scores zero-shot text classifiers on 22 datasets. Trained without any of them, our recipe places above all 35 systems on the benchmark's leaderboard from 4B parameters on, at half the size of the best of them, Qwen3-Reranker-8B. At 8B, the size of the default Crossroad model, it scores 0.745 macro-F1 against that system's 0.722.

0.6B1.7B4B8B14B
Figure 1. Macro-F1 over the 22 BTZSC datasets against parameters. In grey, the 35 leaderboard systems and their best score at or below each size. Hover over our points for their scores.

Interface

One endpoint. The context is state, as text or JSON, the questions are keyed by name, and the answers come back under the same names with the usage. Keys stay on your server; each request carries an idempotency key so a retry cannot be billed twice.

Every question names its answer type, and the type fixes the shape of the reply. A request carries up to sixteen questions about the same context.

Three question types, one request.
TypeFieldsAnswer
choiceinstruction, options (2 or more)one of the options, with each option’s probability
booleaninstructiontrue or false, with the probability of true
scoreinstruction, min, maxthe expected score in its range, with each level’s probability
Listing 1. POST /v1/decisions
{
  "state": "Hi! We were charged twice for our team's plan this month. Could you help us sort this out?",
  "questions": {
    "route": {
      "type": "choice",
      "instruction": "Which team should handle this request?",
      "options": [
        "billing",
        "engineering",
        "sales"
      ]
    },
    "needs_reply": {
      "type": "boolean",
      "instruction": "Does this message need a reply?"
    },
    "urgency": {
      "type": "score",
      "instruction": "How urgent is this request?",
      "min": 1,
      "max": 5
    }
  }
}

Full documentation →

Pricing

Crossroad is free for now: requests aren’t billed.

USD per million tokens
ModelInputOutput
CrossroadFree$0.000
Crossroad MFree$0.000
Jev 1.14$0.042$0.000
GPT-6 Luna$0.100$0.500
Gemini 3.5 Flash-Lite$0.300$2.500

Standard text rates, checked October 31, 2026. Model names link to provider pricing. Excludes caching, batch discounts and free tiers. Token counts and capabilities vary by model.

Get a free key →

Free tier

While Crossroad is free, any account can create keys and make requests without a payment method. Requests made while it is free stay free after billing starts.

By default each account can have four requests in progress, running or waiting for a GPU, and send 120 a minute on each model, and a request waits up to 10 seconds for a GPU. A filter counts one request for every batch of up to 16 items. Every account starts with the same limits so the GPUs stay shared fairly; an account whose traffic crowds out others may get lower limits.

A 429 means a limit was reached and the refused request did not run; a filter keeps the batches it finished. Retrying with the same idempotency key is safe. Its code says which limit:

CodeWhat to do
concurrency_limitYour account has as many requests in progress on this model as its limit allows, four by default. Send the next when one finishes.
rate_limitYour account used its requests for the minute, 120 by default. Retry after the seconds in Retry-After.
Any otherEvery GPU is busy. Retry after the seconds in Retry-After.