Technology Overview
The AI technology stack for commerce — foundational models, agentic AI, LLMs, privacy & security, investment costs, and what comes next — based on Book Part 5 of the AI Best Practices for Commerce reference.
Comparing Public LLMs
Which model, for which task, at what cost
OpenAI: The Reasoning Specialists
OpenAI entered 2025 with three distinct model families: the GPT-4 series for general-purpose intelligence, the o1 series for complex reasoning, and the GPT-4o family optimized for speed and multimodality.
The GPT-4 models, released in March 2023 and iteratively improved through 2024, became workhorses of retail AI. GPT-4 Turbo’s 128,000-token context window enabled consideration of complete product catalogs in single inferences. The model excelled at understanding complex specifications and providing nuanced recommendations, parsing technical details to help customers compare products across multiple dimensions.
Synthetic data generation emerged as GPT-4’s unexpected retail superpower. When retailers needed training data for specialized models but lacked sufficient labeled examples, GPT-4 could generate thousands of synthetic product descriptions with variations in materials, construction techniques, and attributes. This pattern of using frontier models to create training data for specialized models became standard practice by 2024.
Code generation proved valuable for retailers building custom integrations, data pipelines, and automation scripts. GPT-4 understood not just programming languages but also common retail data formats, API patterns, and ecommerce platforms like Shopify and BigCommerce, enabling faster development of business-critical systems.
The o1 reasoning models, released in September 2024, took a different approach. Rather than immediately generating responses, these models spent additional compute time on internal “chain of thought” reasoning. For complex retail problems requiring multi-step logic, this deeper reasoning could improve outcomes. The tradeoff: response times of 10-30 seconds and costs roughly 3-4x higher than GPT-4 made them unsuitable for customer-facing applications but valuable for strategic analysis and complex problem-solving.
Pricing for OpenAI models evolved significantly. GPT-4 originally launched at $30 per million input tokens and $60 per million output tokens. By late 2024, GPT-4o had dropped to $3 per million input and $10 per million output tokens, an 83% reduction in output costs and 90% reduction in input costs over 16 months. GPT-4o mini, released in July 2024, offered even more aggressive pricing at $0.15 input and $0.60 output per million tokens, competing directly with smaller models from other providers.
The weakness at scale remained cost. Running GPT-4 across millions of daily interactions could easily generate $100,000+ monthly bills for mid-size retailers, driving many toward hybrid architectures using GPT-4 for complex queries and cheaper models for routine tasks.
| Provider | Key Strengths | Best Use Cases | Notable Limitation |
|---|---|---|---|
| OpenAI GPT-4/o1 | Reasoning depth, code gen, synthetic data; o1 for complex multi-step logic | Marketplace automation, engineering acceleration, complex analysis | Cost: $100K+/month for mid-size retailers at scale |
| Anthropic Claude | Long context (200K tokens), tool use reliability, safety/compliance | Customer service, search relevance, business process automation | Vision less capable than Gemini; no native web access |
| Google Gemini | Best-in-class vision, native web search, 1M token context window | Visual product discovery, fashion analytics, Workspace integration | Reasoning depth lags o1; tool use less consistent than Claude |
| Meta Llama | Open-weights: self-host, customize freely, no per-token API cost | High-volume classification, privacy-sensitive data, edge deployment | Infrastructure expertise required; no support SLA |
| Mistral | MoE efficiency, exceptional cost/performance, European data sovereignty | Embedded devices, budget-conscious deployments, EU retailers | Reasoning depth and context window lag larger models |
Last updated: March 12, 2026