The Hidden Backend of AI Astrology: Thinking Tokens and Kundali Architecture

🌐 Read this in: বাংলা · हिन्दी

"When your astrology app shows an 'AI-powered prediction' within seconds, is there really artificial intelligence working behind the screen — or is it a pre-written template?"

This question came to my mind when I first opened a major astrology app and saw a "complete AI kundali analysis" appear in just 2 seconds. We have worked out how much data and how much cost is required to perform genuine AI-powered astrological analysis. And if the whole kundali were really written by AI, that cost would be very hard for any free app to bear.

Today I will not name any company or make accusations. I will only present mathematical evidence to show why real AI-driven astrological analysis is difficult, and how MyAstrology.in has solved the problem through a different path.

How Much Data Does a Real Kundali Actually Contain?

Suppose you were born on 2 October 2026, at 11:50 AM, in Kolkata. When MyAstrology.in's JavaScript engine (vsop87-planets.js) generates the kundali from this data, the following components are all calculated:

Create your free Janam KundaliPrecise planetary positions with VSOP87

Kundali →

📊 Core kundali components: 9 planets with rashi, degree, nakshatra, and pada; 12 house positions; Lagna-Chandra-Surya; Vimshottari Mahadasha-Antardasha-Pratyantardasha; Shadbala (Saptavargaja, Kala, Chesta, Drig); Ashtakavarga; Navamsa (D-9); Dashamsa (D-10); transits; aspects; yogas and karakas; Gulika-Mandi-Pranapada.

When all this data is expressed as numbers and text, each kundali contains the equivalent of 80 to 90 pages of information. Assuming an average of 250–300 words per page, that totals 20,000 to 27,000 words.

Now, in the language of AI, these words are measured in tokens. In English, 1 word ≈ 1.3 tokens, but in Bengali, ≈ 3–4 tokens. So a complete kundali's raw data equals roughly 60,000 to 80,000 tokens.

Thinking Tokens: The Hidden Cost of AI

Here is something most users do not know — thinking tokens. Modern AI models like Gemini 2.5 Pro and OpenAI's o1 do not just write an answer; they first think internally. This thinking process also consumes tokens.

🔍 What are thinking tokens? Suppose you ask AI — "If someone was born on 25 June 2017 at 11:50 AM in Kolkata, which rashi was the Moon in?"

The AI will not answer immediately. It will first think internally:

1. Convert the birth date to a Julian Day
2. Get Kolkata's latitude and longitude
3. Extract the Moon's position from the lunar theory (ELP-2000/82)
4. Subtract the Lahiri Ayanamsa
5. Determine the nakshatra
6. Only then produce the answer

This entire thinking process is counted as "thinking tokens." For a precise astrological question, it can be 2–3 times the input tokens.

So, to give a correct answer to just one question, the AI's cost becomes:

🧮 Token cost breakdown (for one question):

• Input (kundali data): ~60,000 tokens
• Thinking tokens: ~30,000 tokens
• Output (answer): ~5,000 tokens
• Total: ~95,000 tokens

The Real Cost of the Gemini API: Mathematical Proof

The approximate Gemini API rates in 2026:

💰 Gemini 2.5 Pro rates:
• Input: $1.25 per 1 million tokens
• Output: $10 per 1 million tokens
• Thinking tokens: charged at the output rate

Now suppose a complete 80–90 page kundali needs to be written by AI. If it cannot be done in a single call (which is technically impossible given the size), multiple API calls are required. Assume 10 calls:

📈 AI cost per user:

• Input: 60,000 × 10 = 600,000 tokens
• Thinking: 30,000 × 10 = 300,000 tokens
• Output: 100,000 × 10 = 1,000,000 tokens
• Total ≈ 1.9 million tokens per user

Cost ≈ $0.75 (input) + $13 (thinking and output) ≈ $13.75 ≈ ₹1,150 per user

Now a simple question — can any free app serve 100,000 users a month at ₹1,150 each? Answer: Impossible. That would be over ₹10 crore per month.

"If a free app spends ₹1,000 on AI per user, the company would shut down within weeks. Yet these apps keep running year after year — so what does that mean?"

So What Do Big Apps Actually Do?

Now we come to the hard question — how does a "AI prediction" appear the moment a user opens a major astrology app? From the technical side, a few explanations are plausible (we have not seen inside any particular app — these are inferences):

🔸 Possibility 1: Template-based system
50–100 pre-written text blocks are stored in a database. One is displayed according to the user's rashi. There is no real-time AI involved.

🔸 Possibility 2: Light AI + minimal input
Only Chandra rashi and Lagna (2–3 fields) are sent to AI. The AI writes a small paragraph, which is then copied across 5–6 pages.

🔸 Possibility 3: Caching
The same answer is reused for thousands of users with the same rashi. This reduces AI cost, but personalization is lost.

MyAstrology.in's Hybrid Solution

We have solved this problem in three layers:

🟢 Layer 1: Computation in JavaScript, not AI

Our vsop87-planets.js engine uses the VSOP87 model. Its results are compared automatically with Swiss Ephemeris 2.10.03 on 401 dates, and the panjika's planetary table for 2025–2032 is built from NASA JPL's DE441. All computation happens in JavaScript — no AI involved.

🟢 Layer 2: AI only shapes language, not computation

We send Gemini only the relevant data (planets, houses, dashas) in JSON format. We instruct it — "Write the classical significance of the name and answers to the customer's 5 questions — do not calculate or guess anything beyond the calculations given."

Result: the whole kundali does not have to be rewritten, so the cost stays low and the same points do not repeat.

🟢 Layer 3: Human final verification

Every premium report is read by Dr. Prodyut Acharya (PhD, Vedic Astrology), and corrected where needed, before it reaches the customer.

Comparative Analysis: Templates vs Real Computation

Aspect	Generic App (probable)	MyAstrology.in
Calculation source	Usually not published	VSOP87 (checked against Swiss Ephemeris); NASA DE441 table for the panjika
Input data	2–3 fields	80+ page kundali
Thinking tokens	Limited	Not for computation, only language
Cost per user	Unknown	Defined, measured
Human verification	Generally absent	By Dr. Acharya himself
Limitation disclosure	Generally absent	Openly published

Our Evidence: Swiss Ephemeris Audit

We do not only claim — we show proof. Below is our engine compared with Swiss Ephemeris 2.10.03 — 2 October 2026, 11:50 AM, Kolkata (October 2026 engine):

Planet	Swiss Ephemeris	MyAstrology.in	Difference
Sun	Virgo 14°50'53"	Virgo 14°50'51"	2″
Moon	Taurus 27°43'02"	Taurus 27°42'50"	12″
Mars	Cancer 08°11'25"	Cancer 08°11'24"	1″
Mercury	Libra 08°06'13"	Libra 08°06'34"	21″
Jupiter	Cancer 25°35'00"	Cancer 25°35'00"	0″

⚠️ Honest disclosure: The earlier engine used a shortened series for Venus — on some dates across a century it was off by up to 9 arcminutes. In October 2026 Venus, Mars, Jupiter and Saturn were moved to the full VSOP87D series; they now agree with Swiss Ephemeris within 6 arcseconds on 401 dates. Mercury (21″ here) still uses a shortened series, and we say so openly.

"Let technology be honest. Let astrology be precise. Let trust be built on evidence."

❓ Frequently Asked Questions (FAQ)

Q1: What are thinking tokens, and how do they increase AI cost?

A: Thinking tokens are the tokens used for an AI's internal reasoning. When you ask an astrological question, the AI does not answer immediately — it first internally thinks through Julian Day conversion, planetary positions, ayanamsa subtraction, and nakshatra determination. This reasoning consumes tokens, roughly 2–3 times the input tokens.

Q2: Why does AI prediction appear instantly on big apps?

A: Genuine AI-driven deep astrological analysis costs more than ₹1,000 per user. No free app can bear that cost. So they likely use pre-written templates, lightweight AI, or caching — none of which is true personalized analysis.

Q3: What role does AI play in MyAstrology.in?

A: AI does not calculate anything on MyAstrology.in. Planetary positions, dashas and yogas are computed by our JavaScript engine with fixed formulas. In premium reports, AI (Gemini) writes the interpretation from those results — it does not make new calculations. Every premium report is read by Dr. Prodyut Acharya before it reaches the customer.

Q4: How much data is in one kundali, and why is it hard for AI?

A: A full kundali contains 80–90 pages of information — 20,000–27,000 words, or 60,000–80,000 tokens. This volume cannot be processed in a single API call. Repeated calls cause repetition and drive costs up exponentially. This is why MyAstrology.in operates in separate layers.

Q5: How can I verify if an app is genuinely using AI?

A: Ask five questions: (1) What calculation source do you use? (2) Do you publish independent audits? (3) What does AI do — computation or just language? (4) Does a human verify the output? (5) What are your limitations? If clear answers are missing, be cautious.