🧠 O4-Mini Reasoning (V3)

Overview

V3 uses o4-mini, OpenAI's reasoning-optimized model. Before generating the visible content, the model "thinks" internally using reasoning tokens that are billed but not visible. This produces more logically structured articles with better argument coherence than standard completion models.

How Reasoning Works

o4-mini uses a two-phase internal process:

┌──────────────────────────────────────────────┐
│  PHASE 1: REASONING (internal, not visible)  │
│                                              │
│  Model "thinks" about:                       │
│  - Best article structure for this service   │
│  - Which competitor claims are plausible     │
│  - How to structure the cancellation guide   │
│  - Consumer rights applicable in Australia   │
│                                              │
│  Uses N reasoning tokens (billed as output)  │
└──────────────────┬───────────────────────────┘
                   │
┌──────────────────▼───────────────────────────┐
│  PHASE 2: GENERATION (visible HTML output)   │
│                                              │
│  Produces final article based on reasoning   │
│  Uses completion tokens (also billed)        │
└──────────────────────────────────────────────┘

Key implication: Set max_completion_tokens high enough (10,000+) to allow both reasoning AND generation. Setting it too low results in truncated articles.

System Message (Important)

V3 requires a system message to enforce Billoff branding and word count. Without it, the model tends to drop brand mentions:

"You are a senior SEO content writer for Billoff (billoff.com). "
"CRITICAL: The word 'Billoff' MUST appear multiple times in your output. "
"NEVER write 'Postclic'. "
"Target output: 1,600–2,200 words of pure HTML."

Why needed: Reasoning models tend to be very literal — if the system message is absent, they may produce excellent structure but miss branding requirements specified only in the user prompt.

API Call Configuration

# Python (Billoff/scripts/03_generate_v3.py)
resp = client.chat.completions.create(
    model="o4-mini",
    messages=[
        {"role": "system", "content": system_msg},
        {"role": "user",   "content": prompt},
    ],
    max_completion_tokens=10000,  # includes reasoning + completion
)

# Extract reasoning tokens
reasoning_t = getattr(
    resp.usage.completion_tokens_details,
    "reasoning_tokens", 0
)

// Browser (openai.js) — streaming
const resp = await fetch('https://api.openai.com/v1/chat/completions', {
  method: 'POST',
  headers: { 'Authorization': `Bearer ${apiKey}` },
  body: JSON.stringify({
    model: 'o4-mini',
    messages: [systemMsg, ...messages],
    max_completion_tokens: 10000,
    stream: true,
  })
});
// Reasoning tokens visible in usage.completion_tokens_details.reasoning_tokens

Cost Breakdown

Model pricing: Input $1.10/1M, Output $4.40/1M, Reasoning $4.40/1M tokens.

Python Script Reference

# Run V3 batch (all 20 sample services, 3 parallel workers)
python scripts/03_generate_v3.py

# Output: Billoff/data/results_v3.json

# Compare all 3 methods side by side on 1 service
python scripts/test_compare_3methods.py

# Specific service
python scripts/test_compare_3methods.py --service "Spotify"

Reasoning token tracking

# The script tracks reasoning tokens for accurate cost calculation
reasoning_t = 0
if hasattr(usage, "completion_tokens_details") and usage.completion_tokens_details:
    reasoning_t = getattr(usage.completion_tokens_details, "reasoning_tokens", 0)
elif hasattr(usage, "reasoning_tokens"):
    reasoning_t = usage.reasoning_tokens or 0

# Cost = (prompt × input_rate) + (completion × output_rate) + (reasoning × reasoning_rate)
cost_usd = (prompt_t/1e6)*1.10 + (completion_t/1e6)*4.40 + (reasoning_t/1e6)*4.40

Quality Results (Ocado test)

Tip: Increase max_completion_tokens to 12,000–15,000 for longer articles.

Pros & Cons

Method Comparison Summary

Recommendation

• Use V1 for your top 100–500 most important service pages (high traffic, competitive)
• Use V2 for bulk generation of the long tail (1,000–10,000 services)
• Use V3 for services where consumer rights and legal nuance matter most