How to Call the Quan 3.5 API

Quick Start

Quan 3.5 uses the same endpoint as every Quan model: https://stockup.cc/v1/query. Create a key in the StockUp Developer Console, add a prepaid balance, pass the key in the x-api-key header, and send OpenAI-style chat messages in JSON. Only the model field changes from a Quan 3.4 request.

curl -X POST https://stockup.cc/v1/query \
  -H "Content-Type: application/json" \
  -H "x-api-key: STOCKUP_API_KEY" \
  -d '{
    "model": "quan-3.5",
    "messages": [
      {
        "role": "user",
        "content": "Compare AAPL and MSFT valuation today. Separate confirmed data from interpretation."
      }
    ],
    "stream": false,
    "webSearch": true,
    "temperature": 0.2
  }'

The response has the same shape as Quan 3.4: read the answer from candidates[0].content.parts[0].text. The X-Quan-Model-Version response header is 3.5 when Quan 3.5 served the request.

Pricing

Rates only, for comparison: a request that uses 2,000 input tokens and 6,000 output tokens costs $0.179 on quan-3.5, $0.095 on quan-3.4 ($0.114 from December 15) and $0.0285 on quan-3.4-l before any free allowance. The same question does not use the same number of tokens on every model, so measure your own traffic before comparing costs.

Which Model Should You Call?

What Changes From Quan 3.4

Quan 3.5 keeps Quan 3.4’s request and response format. The differences below are the ones an integration can observe.

Integration differences

• Streaming is opt-in. Quan 3.5 returns one JSON response unless the request sends "stream": true. Quan 3.4 streams unless the request sends "stream": false. Send stream explicitly in both cases.
• Parameters are validated. temperature must be 0 to 2, topP / top_p 0 to 1, and topK / top_k an integer from 1 to 1000. A value outside its range, or one that is not a number, returns 400 INVALID_PARAMETER with parameter and received in the body, before any model call.
• Smaller response body. thoughtSignature is removed from response parts and responseId from the payload.

Reasoning depth by question type

Everyday questions — a price, key figures, news, earnings, dividends, analyst views, what a company does, comparisons, funds, upcoming events, the economy and the Fed, the market, a sector, or explaining a concept — and “why is this stock moving” questions are answered at a lower reasoning level in a short format. Before the answer is returned, it is checked against the data the server retrieved for the request, and it is rewritten if it fails the check. Opinion, forecast, portfolio and multi-part questions keep the full reasoning path. The X-Quan-Thinking-Level header reports the level used.

Grounding and context

• Latest reported results. For analytical questions about a single company, the most recently reported period is read from the company’s SEC filings and given to the model. X-Quan-Latest-Earnings: attached shows when this happened. If the lookup fails, the request continues without it.
• Conversation history. A ticker named in an earlier user message is used when the current message refers back to it. Limits stated in earlier user messages, such as a maximum position size, are carried forward. They come from user turns only, and a value in the request’s constraints field or in the current message always wins. X-Quan-Carried-Constraints lists any limits carried forward.
• Portfolios in the question. Holdings written in the question text, such as “40% NVDA, 30% MSFT, 30% TLT”, are used as the portfolio when the request has no portfolio field. Each symbol is still checked against price history.
• Technical figures. For tickers in the question, moving averages, the 52-week range and average volume are computed from the verified daily price series.

Calculations and unknowns

With Decision Intelligence enabled for the key:

• A question with several parts is split, and each part is grounded and answered separately. X-Quan-Multi-Intent-Sections gives the number of sections.
• Reported allocation weights never sum past 100%.
• Risk figures use the standard Sortino denominator, Wilder’s RSI and compounded portfolio returns. The risk-free rate used is reported and marked when it is an assumption.
• DCF inputs stated plainly in the question are accepted, including the horizon and cash-flow timing, and a single-stage model is allowed.

A balance the user describes as unknown is never treated as cash. A request to treat it as cash is refused before any search or model call, with an X-Quan-Truth-Guard header and a quanTruthGuard object in the response. When cash is unknown, a sale counts in full against a one-day turnover limit.

Answer styles

Quan 3.5 accepts an optional answerStyle field. balanced (the default) and simple lead with the overall view in everyday words and explain finance terms; simple uses shorter sentences. advanced keeps the institutional write-up. Figures and checks are the same in every style. The X-Quan-Answer-Style header reports the style used.

{
  "model": "quan-3.5",
  "messages": [
    { "role": "user", "content": "Is my portfolio of 40% NVDA, 30% MSFT, 30% TLT too concentrated?" }
  ],
  "stream": false,
  "answerStyle": "simple"
}

JavaScript Example

This is the basic pattern for a Node.js backend, serverless function, or agent tool wrapper. Keep the API key on the server side rather than shipping it to a browser client.

const response = await fetch("https://stockup.cc/v1/query", {
  method: "POST",
  headers: {
    "Content-Type": "application/json",
    "x-api-key": process.env.STOCKUP_API_KEY
  },
  body: JSON.stringify({
    model: "quan-3.5",
    messages: [
      {
        role: "user",
        content: "Summarize the most recent reported quarter for Microsoft."
      }
    ],
    stream: false,
    webSearch: true,
    temperature: 0.2
  })
});

if (!response.ok) {
  const error = await response.json().catch(() => null);
  throw new Error(`${response.status} ${error?.code || ""} ${error?.error || ""}`);
}

const data = await response.json();
const answer = data?.candidates?.[0]?.content?.parts?.[0]?.text;
console.log(answer);
console.log("charge (USD):", response.headers.get("X-StockUp-Charge"));

Python Example

Python works well for research notebooks, batch checks, Slack bots, and internal analyst tools.

import os
import requests

payload = {
    "model": "quan-3.5",
    "messages": [
        {
            "role": "user",
            "content": "Explain the difference between P/E and forward P/E."
        }
    ],
    "stream": False,
    "temperature": 0.2
}

res = requests.post(
    "https://stockup.cc/v1/query",
    headers={
        "Content-Type": "application/json",
        "x-api-key": os.environ["STOCKUP_API_KEY"]
    },
    json=payload,
    timeout=120
)

res.raise_for_status()
data = res.json()
print(data["candidates"][0]["content"]["parts"][0]["text"])
print("charge (USD):", res.headers.get("X-StockUp-Charge"))

Streaming Responses

Set stream to true for a chat UI or agent that should show partial output. The API returns server-sent events in the same format as other models: data: lines carrying JSON chunks, then data: [DONE].

curl -N -X POST https://stockup.cc/v1/query \
  -H "Content-Type: application/json" \
  -H "x-api-key: STOCKUP_API_KEY" \
  -d '{
    "model": "quan-3.5",
    "messages": [
      {
        "role": "user",
        "content": "Build a concise bullish and bearish thesis for AMZN using current market context."
      }
    ],
    "stream": true,
    "webSearch": true
  }'

• Billing on a stream. The X-StockUp- billing headers are not sent on a stream, because headers go out before the charge is known. Read the event: billing frame instead. It arrives before data: [DONE] and carries charge, searchCharge, searchQueries, remainingBalance, billingMode and requestId. If settlement fails, the stream sends event: error with BILLING_SETTLEMENT_FAILED and the requestId.
• Checked answers. Everyday and “why is this stock moving” answers are held back until the answer check finishes, so their text can arrive in a single chunk near the end of the stream.

The end of a stream looks like this (values are illustrative):

event: billing
data: {"requestId":"...","billingMode":"cash","charge":0.179,"searchCharge":0,"searchQueries":0,"remainingBalance":24.821,...}

data: [DONE]

Response and Billing Headers

Errors

A verified response-cache hit returns the stored answer without reserving or charging tokens or searches.

Launch Credit

The first 100 developers to redeem code QUAN35 in the developer console receive $5.00 of API credit that can be used only on quan-3.5.

• One redemption per account. The account needs a verified email and must have accepted the current Developer API Terms.
• The credit is spent before your cash balance, and only on quan-3.5 requests. Other models draw from your cash balance as usual.
• Requests paid with the credit are charged at the normal Quan 3.5 rates.
• The code stops working after 100 redemptions, or earlier if the offer is withdrawn.

Moving an Integration From Quan 3.4

1. Change model to quan-3.5.
2. Send stream explicitly. If your code reads server-sent events, send "stream": true.
3. Check that temperature, topP and topK are inside their ranges, and handle 400 INVALID_PARAMETER.
4. If you read billing on streams, read the event: billing frame.
5. Confirm the switch with the X-Quan-Model-Version header.
6. Compare X-StockUp-Charge on your own traffic before moving all of it.

OpenAPI and MCP for AI Agents

The public OpenAPI YAML and OpenAPI JSON list quan-3.5 in the model enum. In the StockUp MCP server, pass "model": "quan-3.5" as a tool argument; the server sends stream: false, so its requests work unchanged.

Production Checklist

1. Store STOCKUP_API_KEY in server-side environment variables or a secret manager.
2. Fund a prepaid balance before switching production traffic; Quan 3.5 has no free allowance.
3. Set maxOutputTokens to the smallest budget your answers need, so the balance check and your costs stay predictable.
4. Log X-Request-Id, X-Quan-Model-Version and X-StockUp-Charge (or the stream’s event: billing) for each request.
5. Treat any response whose Content-Type is not application/json as a server error and retry it with backoff.