AWS Certified AI Practitioner(12) - Pricing & Model Improvement

📘 Amazon Bedrock – Pricing & Model Improvement

1️⃣ Pricing Options

🔹 On-Demand (Pay-as-you-go)

• How it works: Pay only for what you use, like an electricity bill.
• Pricing basis
  • Text Models → Input/Output token count
  • Embedding Models → Input token count
  • Image Models → Number of images generated
• Available Models: Base Models only
• ✅ Pros: Flexible, good for unpredictable workloads
• ❌ Cons: Can become expensive if used continuously over time

---

🔹 Batch Mode (Bulk processing, up to 50% discount)

• How it works: Group multiple requests together → results stored as a single file in Amazon S3
• Discount: Up to 50% cheaper
• ✅ Pros: Great for large-scale processing, strong cost savings
• ❌ Cons: No real-time response, results are delayed
• Best use case: Large batch jobs where immediate results are not required

---

🔹 Provisioned Throughput (Reserved capacity, guaranteed performance)

• How it works: Like a gym membership — reserve processing capacity for a set period (e.g., 1–6 months)
• Guaranteed performance: Ensures a maximum number of input/output tokens per minute
• Available Models: Base, Fine-tuned, and Custom Models
• ✅ Pros: Stable performance and capacity, supports custom models
• ❌ Cons: Not a cost-saving option, purpose is performance guarantee

---

📊 Pricing Options Comparison Table

Option	Billing Method	Pricing Basis	Available Models	Pros	Cons	Best Use Case
On-Demand	Pay-as-you-go	- Text: Input/Output tokens - Embedding: Input tokens - Image: Generated images	Base Models only	High flexibility Great for unpredictable workloads	Expensive for long-term use	Occasional use / Unpredictable demand
Batch Mode	Bulk processing	Results stored in Amazon S3	Base Models only	Up to 50% discount Efficient for large-scale jobs	No real-time response Delayed results	Large requests / No need for instant results
Provisioned Throughput	Reserved capacity (1–6 months)	Guaranteed tokens per minute	Base, Fine-tuned, Custom Models	Guaranteed stable performance Supports custom models	Almost no cost savings	When using custom models / Need guaranteed performance

---

2️⃣ Model Improvement Techniques (Low → High Cost)

1. Prompt Engineering

• Improve results simply by optimizing prompts
• No extra computation → Lowest cost

2. Retrieval Augmented Generation (RAG)

• Uses an external knowledge database (Vector DB)
• No model retraining → relatively low cost
• Additional cost for building and maintaining the database

RAG = “Model + Search function” → lets the model find external knowledge it doesn’t already know.

3. Instruction-based Fine-tuning

• Fine-tune the model with labeled data and specific instructions
• Requires extra computation → Higher cost

4. Domain Adaptation Fine-tuning

• Retrain the model with a large domain-specific dataset
• Requires extensive data preparation + heavy computation → Highest cost

---

3️⃣ Cost Optimization Tips

• Token management → main driver of cost savings
  • Keep prompts concise
  • Limit output length to what’s necessary
• Use Batch Mode → up to 50% cheaper
• Choose smaller models → generally cheaper
• Adjust hyperparameters (Temperature, Top-K, Top-P)
  • Affects model behavior but not pricing

---

📝 Final Summary (Exam/Practical Points)

• On-Demand = Flexibility / Batch = Bulk & Discounts / Provisioned = Guaranteed Performance
• Cost order: Prompt Engineering < RAG < Instruction Fine-tuning < Domain Adaptation
• Cost-saving keys: Token management + Batch Mode