Skip to main content
OpenAI’s API provides a rich set of model families and variants, each optimized for specific use cases—from code completion to conversational AI. Choosing the right model can help you balance performance, cost, and accuracy.

​
Model Families and Variants

A model family groups related variants by their primary function. Each variant carries a version identifier reflecting its training improvements and feature set. Below is a high-level overview:
Model Family (Use Case)VariantOptimized For
Davinci (Natural Language)text-davinci-002Instruction following
text-davinci-003Advanced NLU and generation
Codex (Code Completion)code-davinci-002Code autocomplete and synthesis
GPT-3.5 (Chat & Completion)gpt-3.5-turboLow-latency chat and multi-turn support

​
GPT-3.5 Variants Overview

The GPT-3.5 series balances capability and cost with varying token limits and data cutoffs:
VariantMax TokensTraining Data ThroughPrimary Use Case
text-davinci-0034,096June 2021Natural language tasks
gpt-3.5-turbo4,096September 2021Chat-style interfaces
The image is a table listing GPT-3.5 model variants, including their descriptions, maximum tokens, and training data. It categorizes models into "Chat" and "Word/Code Completion" types.
gpt-3.5-turbo is optimized for multi-turn dialogues and often delivers lower latency and cost for chat applications.
OpenAI continually refines these variants, improving response quality, speed, and cost-effectiveness.

​
Fine-Tuning Models

Fine-tuning allows you to customize a base model on your domain-specific dataset. Benefits include:
  • Domain-tailored responses with higher relevance
  • Shorter prompts, reducing token usage and cost
  • Faster inference, since context is baked into the model
The image illustrates the process of model fine-tuning, showing a flow from custom data to a fine-tuned model that generates responses. It highlights benefits like higher quality results, token savings, and lower latency.

​
Fine-Tuning Workflow

  1. Format your dataset in JSONL with prompt–completion pairs.
  2. Upload data and create a fine-tuning job via the CLI or API.
  3. Monitor training, evaluate performance, and deploy your custom endpoint.
Ensure your training data is clean, balanced, and representative. Low-quality data can degrade performance.
Embeddings convert text into vector representations for advanced applications:
Use CaseDescription
Semantic SearchRetrieve documents based on contextual meaning
RecommendationSuggest similar content or products
Entity ClassificationCategorize text into predefined classes
OpenAI’s text-embedding-ada-002 model generates high-quality embeddings at scale.
The image illustrates an embedding model process where the text "anatine amigos" is converted into a vector using the "text-embedding-ada-002" model.

​
Other Model Families

OpenAI’s ecosystem extends beyond GPT and Codex:

Watch Video