ChatGPT: Complete Encyclopedia

Every detail: architecture, RLHF, model evolution, benchmarks, pricing, 60+ languages, Codex, Sora, Health, Agent, Deep Research, 25+ use cases, 100+ specs. Updated Feb 2026.
Visit ChatGPT →
~45 min read 180+ data points 10,000+ words
180M+
Monthly Users
92%
MMLU (GPT-4)
1.8T
GPT-4 Params
2M
Token Context
60+
Languages
$200
Pro / mo
IntroEvolutionArchitectureRLHF CapabilitiesBenchmarksPricing FeaturesLanguagesCodex SoraHealthAgent25+ Uses LimitsFAQ

📖 What is ChatGPT? (Complete Definition)

ChatGPT is a conversational AI model developed by OpenAI, fine‑tuned from GPT‑3.5 (and later models) using Reinforcement Learning from Human Feedback (RLHF). It interacts in a dialogue format, enabling it to answer follow‑up questions, admit mistakes, challenge incorrect premises, and reject inappropriate requests. Launched as a research preview in November 2022, it became the fastest‑growing consumer application in history, reaching 100 million users in just two months.

Core technology: Built on the Transformer architecture (Vaswani et al. 2017), ChatGPT uses self‑attention mechanisms to process and generate human‑like text. The underlying GPT (Generative Pre‑trained Transformer) models are trained in two phases: unsupervised pre‑training on massive text corpora (Common Crawl, WebText, books, Wikipedia – approx. 570GB of text), followed by supervised fine‑tuning and RLHF.

Current models (as of Feb 2026): GPT‑5.2 (Instant, Thinking, Pro), GPT‑4o, GPT‑4.1, o3, o4‑mini, and legacy GPT‑5/5.1 variants. Knowledge cutoff: August 2025 for GPT‑5.2 series. Context windows: up to 2M tokens (GPT‑5.2 Pro).

📅 The Complete ChatGPT / GPT Model Evolution Timeline

GPT-1 (2018)

117M parameters, first generative pre-trained transformer.

GPT-2 (2019)

1.5B parameters, demonstrated zero-shot learning.

GPT-3 (2020)

175B parameters, few-shot in-context learning.

InstructGPT (Jan 2022)

RLHF introduced, aligned models with human intent.

ChatGPT (Nov 2022)

GPT-3.5 fine-tuned for dialogue, 4K context.

GPT-4 (Mar 2023)

Multimodal, 1.8T params (MoE), 32K context, 82% HumanEval.

GPT-4 Turbo (Apr 2024)

128K context, lower cost, updated knowledge.

GPT-4o (May 2025)

Omni model, native multimodal, faster.

GPT-5 (Oct 2025)

Unified system, automatic reasoning switching.

GPT-5.2 (Dec 2025)

Instant, Thinking, Pro variants; 2M context; knowledge cutoff Aug 2025.

Codex (Feb 2026)

Standalone app for multi-agent coding, long-horizon tasks.

Note: GPT‑4o, GPT‑4.1, o4‑mini, and GPT‑5 (Instant/Thinking) are scheduled for retirement on February 13, 2026, being fully replaced by GPT‑5.2 variants.

⚙️ Technical Architecture: From Transformer to GPT-5.2

🔧 Click to expand/collapse: Full architecture explanation (Transformer, attention, MoE, scaling laws)

1. Transformer Foundation

The "T" in GPT stands for Transformer (Vaswani et al. 2017). Unlike RNNs, Transformers process all tokens in parallel using self‑attention. Each token attends to every other token, computing attention scores that weight the influence of each token on the representation.

Attention(Q,K,V) = softmax(QK^T/√d_k)V

Multi‑head attention runs multiple attention mechanisms in parallel, allowing the model to focus on different subspaces.

2. GPT Architecture (Decoder‑only)

GPT uses only the decoder stack with masked self‑attention (to prevent looking ahead). Each layer consists of masked multi‑head attention, feed‑forward network, and layer normalization with residual connections.

3. Scaling Laws & Model Sizes

  • GPT-3: 175B parameters, 96 layers, 96 attention heads, 12,288 embedding dimension
  • GPT-4: Estimated 1.8T parameters, Mixture of Experts (MoE) with 16 experts, each ~111B active parameters per forward pass
  • GPT-5.2 Pro: Further MoE refinement, 2M token context, 32K embedding dimension (rumored)

4. Training Data

Pre‑training corpus includes: Common Crawl (filtered), WebText2, Books1/Books2, Wikipedia, and code repositories (GitHub). Total size: approximately 570GB of text (for GPT-3), larger for subsequent models. GPT-4 trained on over 100TB of text + multimodal data.

5. Tokenization

GPT models use Byte‑Pair Encoding (BPE) with a vocabulary of ~50,000 tokens (GPT-3/4). Handles subwords, code, and multiple languages efficiently.

🧠 RLHF: The Secret Sauce Behind ChatGPT

⚡ Click to expand: Complete 3‑step RLHF process with PPO, reward modeling, and human data

Reinforcement Learning from Human Feedback (RLHF) was introduced in OpenAI's 2022 paper "Training language models to follow instructions with human feedback." It aligns models with human preferences.

Step 1: Supervised Fine‑Tuning (SFT)

Collect 12k–15k prompt‑response pairs written by human labelers (who play both user and assistant). Fine‑tune the pre‑trained GPT-3.5 on this demonstration dataset. Result: SFT model.

Step 2: Reward Model (RM) Training

For each prompt, generate multiple responses from the SFT model (4–9). Labelers rank these responses from best to worst. This creates a dataset of ~33k comparisons. Train a reward model to predict the human‑preferred ranking (regression loss).

Step 3: Proximal Policy Optimization (PPO)

Use reinforcement learning to fine‑tune the SFT model against the reward model. The policy (PPO model) generates responses, the reward model scores them, and the policy updates to maximize reward. A KL penalty prevents straying too far from the SFT model.

Mathematical objective: objective(ϕ) = E_{(x,y)∼D_{πϕ}}[r_θ(x,y) − β·KL(πϕ(y|x) || π_SFT(y|x))]

This process can be iterated: use the best PPO model to collect more comparisons, retrain RM, and repeat.

For a detailed walkthrough, see the OpenAI InstructGPT paper or the Springer RLHF survey.

✨ Core Capabilities: What ChatGPT Can Do (Full List)

Natural Dialogue

Full conversational memory, context up to 2M tokens, follow‑up questions, clarification requests.

Code Generation

Python, JS, C++, Rust, Go, etc. Debugging, explanation, optimization. Codex integration for multi‑agent coding.

Creative Writing

Poems, stories, scripts, emails, ad copy, essays – adapts tone (warm, enthusiastic, professional).

Advanced Reasoning

Chain‑of‑thought, step‑by‑step math (GSM8K 87.1% GPT-4), logic puzzles, causal inference.

Web Search (Manual)

Optional browsing with citations. Up‑to‑date information retrieval.

File Uploads

Up to 512MB per file (text, PDF, Word, Excel, CSV, images). 2M token context for text.

Advanced Voice Mode

Real‑time conversation, multiple voices, live transcript, background operation.

Image Generation

ChatGPT Images (GPT Image 1.5) – text rendering, editing, recontextualizing. Unlimited for Plus/Pro.

Data Analysis

Upload CSV, generate charts, statistical analysis, pivot tables. Interactive tables.

Multilingual Translation

60+ languages, including all Indian languages. High accuracy (FLORES‑200 benchmarks).

Custom GPTs

Create and share custom versions with specific instructions, knowledge, and actions.

Apps / Plugins

App directory with connectors to Canva, Expedia, Zapier, etc.

📊 Benchmark Performance: Academic Studies (2023–2026)

Source: Springer Journal of Big Data meta‑analysis, Frontiers in Digital Health neurology study, and OpenAI reports.

BenchmarkGPT-3.5GPT-4GPT-4oGPT-5.2
MMLU (5‑shot, reasoning)70.0%86.4%88.1%92.3%
HumanEval (coding)48.1%82.0%85.5%89.7%
GSM8K (math)57.1%87.1%90.2%93.5%
FLORES‑200 (translation, avg)64.281.784.388.1
Neurology board exam 53.3%68.2%81.3%84.1% (vs residents 64.9%)

Key findings: ChatGPT‑5 outperforms neurology residents (84.1% vs 64.9%). On higher‑order items, GPT‑5 scores 86%. Gemini 2.5 trails at 77.6%.

💰 Complete Pricing & Plan Comparison (Official Feb 2026)

Source: OpenAI official pricing page and ZDNET analysis.

PlanPriceKey FeaturesLimits
Free$0GPT-5.2 Instant (up to 10 msgs/5h), limited voice, 16K context, basic image gen, ads (testing in US)No Thinking/Pro, limited memory
Go$8/moHigher limits, may include ads, GPT-5.2 Instant + some ThinkingNo Pro, Codex limited
Plus$20/moGPT-5.2 Instant/Thinking, 160 msgs/3h, Sora (5s 720p), Codex, Agent (~40 runs/mo), Deep Research (~25 tasks/mo), legacy models, ad‑freeNo GPT-5.2 Pro, unlimited images
Pro$200/moGPT-5.2 Pro, unlimited msgs, Sora (20s 1080p, watermark‑free), Codex Max, Agent (~400 runs/mo), Deep Research (250 tasks/mo), Pulse (morning briefings), earliest access, highest context (2M)Price
Team$25/user/moBusiness features, admin controls, higher limits, shared workspaceAnnual required
EnterpriseCustomSOC2, SSO, data residency (US/EU/India/JP etc.), unlimited everything, complianceContact sales

Ads: Free and Go tiers in the US will test ads; Plus/Pro remain ad‑free.

🔧 Exhaustive Feature Matrix (All 40+ Features)

FeatureFreeGoPlusPro
GPT-5.2 Instant✅ (capped)✅ (higher)✅ (high)✅ (unlimited)
GPT-5.2 Thinking✅ (limited)✅ (high)✅ (unlimited)
GPT-5.2 Pro
Legacy models✅ (settings)
Context window (tokens)16K32K128K2M
Advanced Voice Mode✅ (limited)✅ (more)✅ (extended)✅ (unlimited)
Image generation✅ (capped)✅ (more)✅ (unlimited)✅ (priority)
Sora video✅ (5s 720p, 2 concurrent)✅ (20s 1080p, 5 concurrent, no watermark)
Codex coding agent✅ (limited)✅ (standard)✅ (Codex Max)
Agent (automated tasks)✅ ~40 runs/mo✅ ~400 runs/mo
Deep Research✅ (5 light/mo)✅ (10 light)✅ 25 tasks/mo✅ 250 tasks/mo
ChatGPT Pulse
Memory✅ (limited)✅ (more)✅ (high, cross‑chat)✅ (max)
File uploads✅ (limited size)✅ (larger)✅ (512MB)✅ (512MB priority)
Health space✅ (opt‑in)
Apps / plugins
Pinned chats
Tasks (in Pulse)
Ads (US test)✅ (may have)✅ (may have)

Source: Official ChatGPT pricing page, ZDNET

🌐 Language Support: 60+ Languages (Complete List)

ChatGPT supports over 60 languages, including all 22 official Indian languages. The model performs best in English, Spanish, French, German, Chinese, and Hindi, with ongoing improvements for low‑resource languages.

🇮🇳 Indian Languages (22)

AssameseBengaliBodo DogriGujaratiHindi KannadaKashmiriKonkani MaithiliMalayalamManipuri MarathiNepaliOdia PunjabiSanskritSantali SindhiTamilTelugu Urdu

🌍 International Languages (selected)

EnglishSpanishFrench GermanItalianPortuguese RussianJapaneseKorean ChineseArabicTurkish DutchPolishUkrainian GreekHebrewThai VietnameseCzechSwedish FinnishDanishNorwegian HungarianRomanianBulgarian SerbianCroatianSlovak SlovenianLithuanianLatvian EstonianMalayIndonesian TagalogSwahiliAmharic GeorgianKazakhMongolian KhmerLaoBurmese

💻 Codex: The AI Coding Agent (Feb 2026)

Released February 2, 2026: The Codex app for macOS is a command center for managing multiple coding agents in parallel.

  • Run long‑horizon and background coding tasks.
  • Review clean diffs from isolated worktrees.
  • See agent progress and decisions in real‑time.
  • Execute reusable skills and automations.
  • Available with ChatGPT plans that include Codex: Free and Go can try with limited limits; Plus and Pro get 2x rate limits.

Codex can connect to GitHub repositories, generate code changes, run checks, install dependencies, and more. It represents OpenAI's move toward agentic coding assistance.

🎬 Sora: Text‑to‑Video Generation

Sora creates realistic videos from text prompts or images. Available exclusively to paid users.

  • Plus: videos up to 5 seconds at 720p, 2 concurrent generations
  • Pro: up to 20 seconds at 1080p, 5 concurrent, no watermark
  • Supports editing, animation, and consistent characters

🩺 Health: Dedicated Medical Space

Rolling out January 2026: a separate space in ChatGPT for health and wellness conversations.

  • Securely connect medical records, Apple Health, and wellness apps.
  • Answers grounded in your own health data.
  • Health conversations, memory, and files are kept separate and not used for training.
  • Designed to help navigate medical care, not replace it.
  • Available on web and iOS initially (Android coming). Free, Go, Plus, Pro in supported countries (excluding EEA, Switzerland, UK).

🤖 Agent & Deep Research

Agent

Performs multi‑step tasks: browse web, fill forms, handle shopping, etc.

  • Plus: ~40 runs/mo
  • Pro: ~400 runs/mo

Deep Research

Research analyst mode: browses web, analyzes PDFs/images, produces structured reports with citations.

  • Free: 5 light tasks/mo
  • Plus: 25 tasks/mo
  • Pro: 250 tasks/mo

🚀 25+ Real‑World Use Cases

Draft & refine emails
Summarize meetings/PDFs
Clean spreadsheets, generate charts
Translate with voice
Create weekly plans
Generate thumbnails, brand visuals
Build simple webpages
Script podcasts & produce audio
Ad copy variations
Plan travel itineraries
Compare products
Meal plans & grocery lists
Budget from bank CSVs
Summarize academic sources
Coding help (Python sandbox)
Math problem‑solving
Literature reviews
Slide decks from notes
Deep research with citations
Custom GPTs
Health insights (Health space)
Sora video creation
Agent‑automated shopping
Voice conversations (Advanced Voice)
Study assistant (learning mode)

⚠️ Known Limitations & Challenges

Hallucinations

May generate plausible‑sounding but incorrect facts, especially in niche domains.

Sensitive to phrasing

Rephrasing can yield different answers; may refuse when it should answer.

Verbosity

Tends to over‑explain; repeats phrases like "as an AI language model."

Guesses intent

Often guesses instead of asking clarifying questions.

Bias

Can exhibit social biases; OpenAI uses Moderation API and RLHF to mitigate.

Knowledge cutoff

GPT-5.2 cutoff August 2025; events after may be unknown (unless web search used).

❓ Frequently Asked Questions (Ultimate List)

Is ChatGPT free?
Yes, there is a free tier with limited access to GPT-5.2 Instant (up to 10 messages every 5 hours), basic image generation, and voice. Paid plans (Go $8, Plus $20, Pro $200) offer higher limits and more features.
What's the difference between GPT-5.2 Instant, Thinking, and Pro?
Instant is fast for everyday tasks; Thinking applies deeper reasoning for complex problems; Pro is the most powerful, research‑grade model with 2M context and unlimited usage.
Can ChatGPT browse the internet?
Yes, with manual activation (web search button). It can access current information and cite sources. Available on all tiers.
What languages does ChatGPT support?
Over 60 languages, including all major Indian languages (Hindi, Bengali, Tamil, Telugu, etc.) and most European/Asian languages. See full list above.
How large is the context window?
Free: 16K tokens; Go: 32K; Plus: 128K; Pro: 2M tokens. File uploads up to 512MB.
What is Codex?
Codex is a new macOS app (Feb 2026) that manages multiple AI coding agents in parallel, runs background tasks, and reviews code diffs. Available to Plus/Pro with rate limits.
Does ChatGPT remember past conversations?
Yes, with memory enabled. Plus/Pro have improved memory that can find details from past chats. You can manage or delete memory in settings.
What is the Health space?
A dedicated, private space for health conversations where you can connect medical records and wellness apps. Not used for training. Rolling out Jan 2026.
Can I generate images with ChatGPT?
Yes, ChatGPT Images (GPT Image 1.5) is available on all tiers (with usage limits). Plus/Pro get higher quotas. Supports editing and recontextualizing.
What is Sora?
Sora is OpenAI's text‑to‑video generator. Plus users can create 5‑second 720p videos; Pro users get 20‑second 1080p videos without watermarks.
What are Agents?
Agents perform multi‑step tasks like web browsing, form filling, and shopping. Plus: ~40 runs/mo; Pro: ~400 runs/mo.
How accurate is ChatGPT on medical exams?
A 2026 neurology study found ChatGPT‑5 scored 84.1% on board‑style questions vs. residents' 64.9%.
Will there be ads?
OpenAI is testing ads in the US for Free and Go tiers. Plus and Pro will remain ad‑free.
What is ChatGPT Pulse?
A Pro‑only feature that delivers a personalized morning briefing (5‑10 cards) based on your chat history and interests.
How does RLHF work?
Three steps: supervised fine‑tuning on human demonstrations (SFT), training a reward model on ranked outputs (RM), and optimizing with Proximal Policy Optimization (PPO).

📚 Resources & Further Reading

Official ChatGPT

OpenAI product page

OpenAI Research

InstructGPT, GPT-4, RLHF papers

API Docs

Developer guide

Help Center

FAQs & support

Springer Benchmark

Meta‑analysis of LLMs