Bing AI Answers Without API Key: 2026 Guide

To get Bing AI answers without an API key, you have two working options: scrape Bing Copilot using browser cookies (the _U auth cookie) via an unofficial client like sydney.py, or call a hosted actor that already handles the auth, rotation, and parsing. Microsoft does not offer a free public API for Copilot answers, so cookie-based access or a managed scraper are the only realistic paths in 2026.

Quick Answer

You can get Bing AI answers without an API key by authenticating with your Bing _U cookie through an unofficial Python client like sydney.py, or by using a pre-built Apify actor that abstracts the cookie handling. The cookie method takes ~15 minutes to set up but breaks when Microsoft rotates session formats (roughly every 2–4 months in 2024–2025). Hosted actors cost a few cents per query but stay maintained when Bing changes its endpoints. For one-off experiments, use sydney.py; for production pipelines or anything over ~500 queries/day, use a managed solution like Bing AI Search (Copilot) Answers.

Why doesn't Bing offer a free Copilot API?

Microsoft retired the free Bing Search API tier in August 2025 and folded generative answers into paid Azure AI services. The current "Grounding with Bing Search" SKU starts at $35 per 1,000 transactions and requires an Azure subscription, billing setup, and tenant approval. There is no public REST endpoint that returns Copilot's generative summaries the way the chat UI shows them.

That gap is why projects like sydney.py, EdgeGPT (now archived), and BingImageCreator exist — developers reverse-engineered the WebSocket protocol Copilot uses internally so they could call it with the same auth cookie a browser sends.

How does cookie-based Bing Copilot access work?

When you log into bing.com, Microsoft sets a session cookie named _U. Copilot's chat backend trusts this cookie for WebSocket connections to sydney.bing.com. The flow looks like this:

1. Browser sends _U cookie to bing.com/turing/conversation/create to get a conversation signature.
2. Client opens a WebSocket to sydney.bing.com/sydney/ChatHub with the conversation ID.
3. Messages are sent as JSON frames; responses stream back token-by-token with citations attached.

sydney.py wraps all of this. The minimal example:

import asyncio
from sydney import SydneyClient

async def main():
    async with SydneyClient() as sydney:
        response = await sydney.ask("What is the latest Llama model?", citations=True)
        print(response)

asyncio.run(main())

You set the cookie via the BING_COOKIES environment variable. The library handles the WebSocket handshake, message framing, and citation parsing.

How do I extract the Bing _U cookie?

Three steps that take about 60 seconds in Chrome or Edge:

1. Log into bing.com and open Copilot chat at bing.com/chat.
2. Open DevTools → Application → Cookies → https://www.bing.com.
3. Copy the Value of the _U cookie (it's ~200 characters, starts with 1 or similar).

Paste it into your environment:

export BING_COOKIES="_U=<paste_value_here>"

Caveats: the cookie expires when you log out, when Microsoft forces a session refresh (often weekly), or when your IP geolocates differently from the original login. If you hit "User needs to solve CAPTCHA" errors, open Bing in a normal browser, complete the challenge, and re-export the cookie.

What are the limits of unofficial Bing AI scraping?

Cookie-based access works, but it has hard ceilings:

• Rate limits. Empirically, ~30 messages per conversation and a few hundred conversations per day per cookie before throttling. Microsoft does not publish official numbers.
• CAPTCHAs. Triggered by datacenter IPs, fast request patterns, or unusual geos. Residential or mobile IPs cut these by ~80% in our testing.
• Schema drift. The JSON envelope changes silently. In Q1 2024 alone, sydney.py shipped 4 patches for response format changes.
• Account risk. Heavy automated use can flag the underlying Microsoft account. Use a throwaway account, not your work one.
• Concurrency. A single cookie maps to one logical user. Running 10 parallel queries through one cookie returns errors after the second or third request.

For a side project or research, these limits are fine. For a product that needs predictable uptime, they are not.

Can I use a hosted actor instead of managing cookies?

Yes — this is what the Bing AI Search (Copilot) Answers actor on Apify handles. You send a JSON input with your query, the actor runs against Bing Copilot, parses the generative answer plus citations, and returns structured output. No cookie management, no WebSocket plumbing, no CAPTCHA debugging.

A typical input:

{
  "queries": [
    "What are the best open-source vector databases in 2026?",
    "Compare Postgres pgvector vs Qdrant performance"
  ]
}

Output per query includes:

• answer — the generative summary text
• citations — array of {title, url, snippet} objects Copilot referenced
• relatedQueries — Copilot's follow-up suggestions
• timestamp

You call it via Apify's REST API, the JS/Python SDK, or trigger it from n8n, Make, or Zapier. Pricing is pay-per-use — typically a few cents per query depending on response length — which is meaningfully cheaper than Azure's $35/1k Grounding SKU for low-to-mid volume use cases.

When should I pick sydney.py vs a hosted actor?

A decision table based on what I've shipped with both:

If you're building an internal Slack bot that answers ~50 queries/day for your team, sydney.py is fine. If you're enriching a dataset of 50,000 product queries or powering a user-facing feature, the hosted route is the only path that won't wake you up at 2 a.m.

Is scraping Bing Copilot legal?

Bing's Terms of Use prohibit automated access without permission. That said, scraping public search results has been litigated repeatedly — the hiQ v. LinkedIn ruling and subsequent cases generally protect scraping of publicly accessible data, though the legal picture varies by jurisdiction and the data involved.

Practical guidance: don't redistribute Copilot's verbatim generated answers as your own product output, respect robots.txt where it applies to the endpoints you hit, and don't use it to violate Microsoft's content policies (e.g., generating disallowed content at scale). If you're operating a commercial product at meaningful scale, talk to a lawyer and consider the Azure Grounding API as the licensed path.

How do I integrate Bing AI answers into a pipeline?

A common pattern: enrich a list of questions with AI-generated answers plus citation URLs for fact-checking.

Using the Apify Python client:

from apify_client import ApifyClient

client = ApifyClient("<APIFY_TOKEN>")

run_input = {
    "queries": ["best CRM for solo founders 2026"]
}

run = client.actor("practical-tools/bing-ai-copilot").call(run_input=run_input)

for item in client.dataset(run["defaultDatasetId"]).iterate_items():
    print(item["answer"])
    for c in item["citations"]:
        print(" -", c["url"])

Plug this into Airflow, a cron job, or trigger it from a webhook. Citations are particularly useful — you get a list of URLs Copilot considered authoritative, which doubles as a curated seed list for downstream scraping or RAG ingestion.

FAQ

Q: Does Bing Copilot have an official free API? No. Microsoft retired the free Bing Search API tier in August 2025. The current licensed option is Azure AI "Grounding with Bing Search" at $35 per 1,000 transactions, which requires an Azure account and approved tenant.

Q: Will my Microsoft account get banned for using sydney.py? Light personal use rarely triggers action, but high-volume automation can flag the account. Use a dedicated throwaway Microsoft account for any scripted access, and never use your primary work or personal account.

Q: How long does a Bing _U cookie stay valid? Typically a few days to a few weeks before Microsoft forces a refresh. CAPTCHA challenges, IP changes, or detected automation can invalidate it sooner. Production systems need either rotation logic or a managed solution that handles this.

Q: Can I get citations and source URLs, not just the AI summary? Yes. Both sydney.py (with citations=True) and the hosted Bing Copilot actor return citations as structured arrays containing the title, URL, and snippet of each source Copilot referenced. This is the main advantage over scraping plain search results.

Q: What's the cheapest way to get 1,000 Bing AI answers per day? For 1,000/day, a hosted actor like Bing AI Search (Copilot) Answers at pay-per-use pricing typically runs $10–30/day with zero maintenance — cheaper than the $35 Azure Grounding rate and far less work than running and rotating cookies for sydney.py yourself.