GPT Models Cheat Sheet

Updated 2026-04-28

Next Topic: GRPO (Group Relative Policy Optimization) Cheat Sheet

🧠Study flashcards on this topic103 cards · spaced repetition→

GPT (Generative Pre-trained Transformer) models are large language models developed by OpenAI that use transformer architecture to generate text, analyze images, process audio, and perform complex reasoning. The lineup has expanded dramatically: the current flagship, GPT-5.5 (released April 2026), delivers 1M-token context and native computer-use capabilities, while the GPT-4.1 family offers a 1M-token non-reasoning alternative, and open-weight models (gpt-oss) are now available under Apache 2.0. OpenAI's new Responses API has replaced Chat Completions as the recommended primitive for agent-style workloads, bringing stateful context, built-in tools, and MCP server support into a single call.

What This Cheat Sheet Covers

This topic spans 14 focused tables and 140 indexed concepts, 103 flashcards. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.

Table 1: Core ArchitectureTable 2: Model Variants and SpecificationsTable 3: Tokenization and ContextTable 4: API Parameters and ConfigurationTable 5: API Endpoints and IntegrationTable 6: Agentic Tools and Built-in FeaturesTable 7: Multimodal CapabilitiesTable 8: Fine-tuning and CustomizationTable 9: Decoding and Generation StrategiesTable 10: Model Selection and Use CasesTable 11: Performance and OptimizationTable 12: Pricing and Cost ManagementTable 13: Open-Weight ModelsTable 14: Limitations and Considerations

Quick IndexSubscribe to unlock

A jump-to index of every table row in this cheat sheet.

Mind MapSubscribe to unlock

An interactive map of every table and concept in this topic.

Table 1: Core Architecture

Underneath every GPT model is the same transformer machinery, and these are its load-bearing pieces — self-attention that lets each token weigh every other, autoregressive decoding that writes one token at a time, BPE tokenization, and the training stages that turn raw web text into an aligned assistant. Understanding how they fit together explains both why these models are so capable and where their quirks come from.

Component	Example	Description
Transformer architecture	Multi-head self-attention + feed-forward layers	Foundation of all GPT models — processes token sequences in parallel, enabling efficient long-range dependency modeling.
Self-attention mechanism	Query-Key-Value matrices compute token relationships	• Each token attends to all others, assigning importance weights based on semantic relevance • enables context-aware representations.
Autoregressive generation	Predicts token $t_{n+1}$ given $t_1, t_2, ..., t_n$	Generates text one token at a time, conditioning each prediction on all previously generated tokens for coherent outputs.
Multi-head attention	8–96 attention heads process different aspects simultaneously	Parallel attention mechanisms allow the model to focus on multiple representation subspaces (syntax, semantics, entities) at once.
Pre-training + fine-tuning	Unsupervised learning on web text → RLHF alignment	Two-stage process: model learns language patterns from massive datasets, then human feedback refines outputs for helpfulness and safety.
Byte Pair Encoding (BPE)	"unhappiness" → ["un", "happiness"]	• Tokenization algorithm that iteratively merges the most frequent byte pairs • balances vocabulary size with coverage of rare words.
Layer normalization	Applied before/after each transformer block	Stabilizes training by normalizing activations across features, improving gradient flow in deep networks.
Positional encoding	Sinusoidal or rotary (RoPE) position embeddings	Since transformers lack inherent sequence order, positional encodings inject absolute or relative position information into token representations.
Mixture-of-Experts (MoE)	GPT-4 MoE: ~1.76T total params, ~220B active	Routes each token through a sparse subset of expert sub-networks, enabling massive parameter counts at lower inference cost.

Table 2: Model Variants and Specifications

The current OpenAI lineup at a glance — flagship GPT-5.5, the fast non-reasoning GPT-4.1 family, dedicated reasoning models like o3 and o4-mini, native audio, the open-weight gpt-oss pair, and the legacy entries kept around for compatibility. Each row pairs context size and per-token price with the kind of work it's actually built for, which is the fastest way to narrow down which model you should be calling.

Model	Example	Description
GPT-5.5	1M context, 128K output, $5/$30 per 1M tokens	• Current flagship — released April 23, 2026; first fully retrained base since GPT-4.5 • natively omnimodal (text, image, audio, video); reasoning levels: none/low/medium/high/xhigh • 82.7% Terminal-Bench 2.0, 85% ARC-AGI-2.

The rest of this section is part of the full cheat sheet

Table 3: Tokenization and Context

GPT models don't see words — they see tokens, and almost every cost, limit, and quirk traces back to how text gets chopped up. The entries here cover the tokenizer libraries and encodings, what a context window actually holds, and prompt caching, giving you the mental model you need to estimate spend and avoid blowing past the window.

Concept	Example	Description
tiktoken library	`import tiktoken` `enc = tiktoken.get_encoding("cl100k_base")` `enc.encode("hello world")`	• OpenAI's official fast BPE tokenizer • provides exact token counts for GPT models • essential for cost estimation and prompt engineering.

The rest of this section is part of the full cheat sheet

Table 4: API Parameters and Configuration

The knobs you set on each API request to shape its behavior — temperature and top_p for randomness, reasoning_effort for thinking depth, the penalties for curbing repetition, and stateful options like store and previous_response_id that carry context between calls. Tuning these well is often the difference between a model that feels sharp and one that wanders.

Parameter	Example	Description
temperature	`temperature=0.7`	• Controls randomness of outputs • 0 = deterministic, 2 = maximum creativity • adjusts sampling probability distribution.

The rest of this section is part of the full cheat sheet

Table 1: Core Architecture

Component	Example	Description
Transformer architecture	Multi-head self-attention + feed-forward layers	Foundation of all GPT models — processes token sequences in parallel, enabling efficient long-range dependency modeling.
Self-attention mechanism	Query-Key-Value matrices compute token relationships	• Each token attends to all others, assigning importance weights based on semantic relevance • enables context-aware representations.
Autoregressive generation	Predicts token $t_{n+1}$ given $t_1, t_2, ..., t_n$	Generates text one token at a time, conditioning each prediction on all previously generated tokens for coherent outputs.
Multi-head attention	8–96 attention heads process different aspects simultaneously	Parallel attention mechanisms allow the model to focus on multiple representation subspaces (syntax, semantics, entities) at once.
Pre-training + fine-tuning	Unsupervised learning on web text → RLHF alignment	Two-stage process: model learns language patterns from massive datasets, then human feedback refines outputs for helpfulness and safety.
Byte Pair Encoding (BPE)	"unhappiness" → ["un", "happiness"]	• Tokenization algorithm that iteratively merges the most frequent byte pairs • balances vocabulary size with coverage of rare words.
Layer normalization	Applied before/after each transformer block	Stabilizes training by normalizing activations across features, improving gradient flow in deep networks.
Positional encoding	Sinusoidal or rotary (RoPE) position embeddings	Since transformers lack inherent sequence order, positional encodings inject absolute or relative position information into token representations.
Mixture-of-Experts (MoE)	GPT-4 MoE: ~1.76T total params, ~220B active	Routes each token through a sparse subset of expert sub-networks, enabling massive parameter counts at lower inference cost.

Table 2: Model Variants and Specifications

Model	Example	Description
GPT-5.5	1M context, 128K output, $5/$30 per 1M tokens	• Current flagship — released April 23, 2026; first fully retrained base since GPT-4.5 • natively omnimodal (text, image, audio, video); reasoning levels: none/low/medium/high/xhigh • 82.7% Terminal-Bench 2.0, 85% ARC-AGI-2.

Table 3: Tokenization and Context

Concept	Example	Description
tiktoken library	`import tiktoken` `enc = tiktoken.get_encoding("cl100k_base")` `enc.encode("hello world")`	• OpenAI's official fast BPE tokenizer • provides exact token counts for GPT models • essential for cost estimation and prompt engineering.

Table 4: API Parameters and Configuration

Parameter	Example	Description
temperature	`temperature=0.7`	• Controls randomness of outputs • 0 = deterministic, 2 = maximum creativity • adjusts sampling probability distribution.

GPT Models (GPT-4, GPT-5) Cheat Sheet

Table 1: Core Architecture

Table 2: Model Variants and Specifications

Table 3: Tokenization and Context

Table 4: API Parameters and Configuration

GPT Models (GPT-4, GPT-5) Cheat Sheet

Table 1: Core Architecture

Table 2: Model Variants and Specifications

Table 3: Tokenization and Context

Table 4: API Parameters and Configuration