Module 2 — Model Comparison Analysis

1. Models Evaluated

Model	Provider	Inference Profile	Characteristics
Claude Sonnet 4.6	Anthropic	jp.anthropic.claude-sonnet-4-6	High capability, extended thinking
Nova Pro	Amazon	apac.amazon.nova-pro-v1:0	Balanced performance/cost
GPT-OSS 20B	OpenAI	openai.gpt-oss-20b-1:0	Cost-optimized reasoning

All models accessed via Amazon Bedrock Converse API with cross-region inference profiles.

2. Benchmark Results

Three prompt categories tested: short factual Q&A, analytical comparison, and structured summarization.

2.1 Latency (ms)

Model	Short Factual	Analytical	Structured	Average
Claude Sonnet 4.6	3,189	28,519	5,178	12,295
Nova Pro	659	2,300	1,140	1,366
GPT-OSS 20B	558	6,942	1,593	3,031

2.2 Output Tokens

Model	Short Factual	Analytical	Structured
Claude Sonnet 4.6	87	1,713	202
Nova Pro	66	541	210
GPT-OSS 20B	93	2,048*	469

* GPT-OSS 20B hit the default max_tokens limit (2048) on the analytical prompt.

2.3 Cost per Request (USD)

Model	Short Factual	Analytical	Structured	Total (3 calls)
Claude Sonnet 4.6	$0.00137	$0.02576	$0.00310	$0.03023
Nova Pro	$0.00022	$0.00174	$0.00068	$0.00264
GPT-OSS 20B	$0.00005	$0.00083	$0.00020	$0.00107

3. Analysis & Trade-offs

Claude Sonnet 4.6

Highest quality output with nuanced reasoning and well-structured responses
Extended thinking significantly increases latency on complex prompts (28s for analytical)
Cost is 11x Nova Pro and 28x GPT-OSS — premium pricing for premium output
Best for: customer-facing content generation, complex analysis, tasks where quality justifies cost

Nova Pro

Consistently low latency across all prompt types (avg 1.4s)
Concise but accurate responses — fewer tokens but relevant content
Mid-range cost with predictable performance envelope
Best for: real-time chat applications, high-throughput workloads, latency-sensitive use cases

GPT-OSS 20B

Lowest cost per token — ideal for budget-constrained scenarios
Reasoning model: returns reasoningContent blocks requiring special response parsing
Good latency on short tasks but variable on complex prompts
Best for: batch processing, internal tooling, cost-sensitive applications with moderate quality needs

4. Recommendations

For startup chat applications: Use Nova Pro as the default model (low latency, reasonable cost), offer Claude Sonnet 4.6 as a premium option for complex tasks, and use GPT-OSS 20B for background/batch operations where cost dominates.

Decision Matrix

Priority	Recommended Model	Rationale
Quality first	Claude Sonnet 4.6	Superior reasoning, best for customer-facing
Latency first	Nova Pro	Sub-second for short queries, consistent
Cost first	GPT-OSS 20B	28x cheaper than Claude, acceptable quality
Balanced	Nova Pro	Best latency/cost ratio for interactive use

GenAI Foundations — Model Comparison Analysis