GPT Model Family Comparison
The GPT-6 tiers, what reasoning effort replaced, when to use the Responses API over Chat Completions, how the bill is calculated, and what is different on Azure.
GPT-4o ("omni", 2024) made text, image, and audio input standard in one model. Two years on it is a legacy model, and the current OpenAI lineup is the GPT-6 family (September 2026): GPT-6 Astra ($10/$50 per million tokens) for the hardest work, GPT-6.1 Sol ($2/$10) as the default, and GPT-6 Luna ($0.10/$0.50) for high-volume tasks. All three share a 1.05M-token context and 128K output. The separate o1/o3 reasoning models are gone: reasoning is built in and controlled with a reasoning-effort setting.
The pipeline
Each stage in order — click any step to read what it does.
Interactive Pipeline Flow
Architecture diagrams
OpenAI model tiers and use cases.
Key design decisions
The trade-offs worth knowing before you build this.
Instead of prompting 'respond in JSON format' (unreliable), use response_format with json_schema. The model is constrained to valid JSON matching your Zod/Pydantic schema — eliminating parse errors in production.
GPT-6 Luna handles summarisation, classification, extraction, and light reasoning for 1/20th the price of GPT-6.1 Sol. A/B test your prompts on both models and upgrade only when quality gap affects user outcomes.
GPT-6 Astra costs 5x what GPT-6.1 Sol does, and high reasoning effort adds latency and output tokens on any tier. They pay off on hard math, science, and complex coding. For most product features (chat, search, summarisation), Sol or Luna at low effort is faster and cheaper with no visible quality loss.
Continue reading
All LLM Model Landscape articles →Claude: Anthropic's Model Family in Practice
The four Claude tiers, adaptive thinking and effort, the tool-use rules that differ from other providers, prompt caching, and where you can run it.
Llama: Open-Weight Models for Production
The Llama lineup and licence, serving with Ollama and vLLM, GPU memory sizing, and an honest look at when a hosted API is still the better choice.
Evals for Choosing a Model
Why public benchmarks do not transfer to your workload, how to build a model-selection eval, and why cost per successful task beats price per token.
Discussion
Sign in to share your feedback and join the discussion.