ChatGPT Versions: All 37 GPT Models at a Glance

Do you remember when ChatGPT took the world by storm in November 2022? OpenAI's GPT models have completely changed how we work and use computers.

OpenAI, the makers of ChatGPT, started as a small non-profit organization and became the pioneers of a digital revolution.

But how do the different ChatGPT versions or GPT models actually differ? What milestones have they set? And what can you use them for?

In this article, I walk you through every GPT model to date. From the fresh GPT-5.6 preview (Sol, Terra, Luna, available via API and Codex only since late June 2026), through GPT-5.5 as the generally available flagship and GPT-5.5 Instant as the new ChatGPT default, all the way back to GPT-1.

What Are GPT Models?

GPT models are advanced language models based on neural networks, trained to generate human-like text. The "GPT" in their name stands for "Generative Pre-trained Transformer," which already describes the basic principles of these models:

• Generative: They can create new, original content.
• Pre-trained: They were pre-trained with enormous amounts of text.
• Transformer: They are based on the Transformer architecture, a groundbreaking AI technology for language processing.

The models are trained with huge amounts of text from the internet, e.g., Wikipedia and Common Crawl, and learn to recognize patterns and connections in language.

This way, they can not only understand text but also generate text themselves, which is often barely distinguishable from human-written text.

Before we look at the individual models, here are the key milestones of GPT development from 2018 to today.

GPT-5.6 (Limited Preview)

Released: June 26, 2026 (preview)

On June 26, 2026, OpenAI announced the GPT-5.6 family (Sol, Terra, Luna) as a limited preview. The models are currently rolled out via API and Codex only, to around 20 selected partner organizations. There is no ChatGPT rollout yet, and GPT-5.5 remains the generally available standard.

GPT-5.6 Sol: The New Preview Flagship

Sol is the flagship of the GPT-5.6 family and introduces two new mechanisms: a new "Max reasoning effort" setting and an "Ultra" mode with sub-agents that breaks particularly complex tasks into smaller sub-tasks. On Terminal-Bench 2.1, Sol reaches 88.8%, and 91.9% in Ultra mode. On ExploitBench, Sol is competitive with Claude Mythos 5 while using roughly one-third of the output tokens.

Key features:

• Terminal-Bench 2.1: 88.8% (Ultra mode 91.9%)
• Max reasoning effort: New, higher reasoning tier for particularly hard tasks
• Ultra mode: Sub-agent architecture for complex multi-step workflows
• 1 million token context
• API pricing: $5/million input, $30/million output, $0.50 cache read
• Availability: Preview, API and Codex only for selected partners

GPT-5.6 Terra: Balanced Tier

According to OpenAI, Terra matches GPT-5.5 quality at roughly half the price. That makes Terra the most interesting model for teams already using GPT-5.5 in production and looking to cut their inference cost in half without giving up noticeable quality. Terra scores 82.5% on Terminal-Bench 2.1.

• Terminal-Bench 2.1: 82.5%
• API pricing: $2.50/million input, $15/million output, $0.25 cache read
• Positioning: GPT-5.5 quality at roughly half the price
• Availability: Preview, API and Codex only

GPT-5.6 Luna: Fast and Cheap

Luna is the smallest, fastest, and cheapest variant in the GPT-5.6 family. Luna scores 84.3% on Terminal-Bench 2.1. A Cerebras deployment with up to 750 tokens per second is planned for July 2026 according to OpenAI.

• Terminal-Bench 2.1: 84.3%
• API pricing: $1/million input, $6/million output, $0.10 cache read
• Cerebras deployment: Up to 750 tokens/sec planned for July 2026
• Availability: Preview, API and Codex only

One important note: GPT-5.6 runs under the US government framework (executive order from June 2, 2026), which is why it is limited to a small number of partner organizations. If you are shipping today, GPT-5.5 is still the right choice. GPT-5.6 sits "on top" of the existing lineup rather than replacing it.

GPT-5.5

Released: April 23, 2026

GPT-5.5 is currently the most powerful generally available model from OpenAI and the biggest jump on agentic tasks since GPT-5. On Terminal-Bench 2.0 it scores 82.7% versus 75.1% for GPT-5.4, a gap of over 7 percentage points. Since June 26, 2026, GPT-5.6 has been running as a preview on top (API and Codex only). For ChatGPT users, GPT-5.5 stays the current top variant.

GPT-5.5: Agentic Workhorse

GPT-5.5 is specifically designed for multi-tool workflows. It can autonomously write code, run web research, and analyze data across multiple tools without every step needing explicit instruction.

Key features:

• 1 million token context: Via API and in ChatGPT
• Terminal-Bench 2.0: 82.7% (vs. 75.1% for GPT-5.4)
• GDPval: 84.9% on real-world knowledge work tasks
• API pricing: $5/million input tokens, $30/million output tokens
• Availability: ChatGPT Plus, Pro, Business, Enterprise, and Codex

GPT-5.5 Pro: The Research Partner

GPT-5.5 Pro is the top variant for the most demanding tasks, especially in science, finance, and law. It delivers 39.6% on FrontierMath Tier 4 (GPT-5.5 base: 35.4%) and 90.1% on BrowseComp (vs. 84.4% for the base variant).

• FrontierMath Tier 4: 39.6% on the hardest math benchmark
• BrowseComp: 90.1% on complex web research
• API pricing: $30/million input tokens, $180/million output tokens
• Availability: ChatGPT Pro, Business, and Enterprise

GPT-5.5 Instant: The New Default Model

On May 5, 2026, OpenAI rolled out GPT-5.5 Instant. It replaces GPT-5.3 Instant as the default model in ChatGPT and, according to internal evaluations, produces 52.5% fewer hallucinated claims on high-stakes prompts in medicine, law, and finance.

• Hallucinations: 52.5% fewer than GPT-5.3 Instant on high-stakes prompts
• AIME 2025: 81.2% (vs. 65.4% for GPT-5.3 Instant)
• Personalization: Can refer back to past conversations, files, and Gmail
• More concise answers: 30.2% fewer words, 29.2% fewer lines
• Availability: Default for all ChatGPT plans, exposed via API as "chat-latest"

GPT-5.4

Released: March 2026

GPT-5.4 was the most powerful model from OpenAI at release and brought two significant innovations: native computer-use and a massively expanded context window of up to 1 million tokens via the API. It has since been succeeded by GPT-5.5 but remains active as a cheaper workhorse model.

GPT-5.4 Thinking: Reasoning with Computer-Use

GPT-5.4 Thinking combines the reasoning capabilities of the GPT-5 series with native computer-use functionality. The model can directly operate the user's computer, navigate websites, fill out forms, and autonomously execute complex multi-step tasks.

Key features:

• Computer-Use: Native understanding of screen content and mouse control
• 1M Token Context: Processing extremely long documents and codebases via the API
• Improved Benchmarks: 92.8% on GPQA Diamond, 75.0% on OSWorld-Verified
• Availability: For ChatGPT Plus, Team, Pro, and Enterprise users

GPT-5.4 Pro: Maximum Performance

GPT-5.4 Pro is the most powerful variant, exclusive to Pro and Enterprise users. It uses extended reasoning time and delivers the best results for particularly demanding tasks.

• BrowseComp: 89.3% (web research benchmark)
• SWE-Bench Pro: 57.7% on professional software engineering tasks
• Priority Access: No wait times, even during high demand

GPT-5.3

Released: February 2026

GPT-5.3 focused on speed and efficiency and was the default model for ChatGPT users until early May 2026, when GPT-5.5 Instant took over that role.

GPT-5.3 Instant: Predecessor Model

GPT-5.3 Instant initially replaced GPT-5.1 Instant as the default model in ChatGPT. Since May 5, 2026 it has itself been superseded by GPT-5.5 Instant. OpenAI is keeping it available as an option for paid users for roughly three months before phasing it out.

• 400K Token Context: Same context as GPT-5.1, but faster processing
• Knowledge Cutoff: August 2025 (with web search for current information)
• Availability: All plans (Free with limit, Go unlimited, Plus/Pro/Business/Enterprise)

GPT-5.3 Codex: Agentic Coding

GPT-5.3 Codex is specifically optimized for the Codex platform and API usage. It powers OpenAI's Codex Agent, which can autonomously write, test, and deploy code.

• Agentic Workflows: Can independently execute tasks in repositories
• 400K Token Context: Processing large codebases
• Availability: Via the API and Codex platform

GPT-5.2

Released: January 2026

GPT-5.2 was an incremental update to GPT-5.1 with improvements in reasoning speed and multimodality. It has already been superseded by GPT-5.3 as the default model and has legacy status (available until June 2026).

GPT-5.2 primarily served as a transition model and brought under-the-hood improvements in inference efficiency that paved the way for GPT-5.3 and GPT-5.4.

GPT-5.1

Released: November 12, 2025

After mixed user reviews of GPT-5, OpenAI made a significant overhaul of the ChatGPT experience with GPT-5.1. The version released on November 12, 2025 reorganizes the GPT-5 family around two main variants (GPT-5.1 Instant and GPT-5.1 Thinking) and introduces groundbreaking personalization features.

GPT-5.1 Instant: Smarter and Warmer

GPT-5.1 Instant was the primary model for everyday conversations and replaced the previous standard GPT-5 model. It has since been superseded by GPT-5.3 Instant.

Key Features:

• Adaptive Reasoning: For the first time, GPT-5.1 Instant can decide for itself when it needs to "think" before answering. For simple questions, it responds immediately; for more complex tasks, it automatically activates reasoning capabilities for more thorough answers
• Warmer Default Tone: The model has been trained to be more conversational and approachable while remaining helpful and precise
• Improved Instruction Following: Significantly better ability to follow instructions correctly and literally
• Same Technical Specifications: 400,000 input tokens, 128,000 output tokens like GPT-5
• API Availability: Available via gpt-5.1-chat-latest in the API

GPT-5.1 Thinking: Intelligent Thinking Time Adjustment

GPT-5.1 Thinking is the specialized variant for complex problem-solving and adjusts its reasoning time more intelligently than its predecessor.

Key Features:

• Adaptive Thinking Time: Precisely adjusts internal thinking time to the complexity of the task. Faster for simple questions, more thorough for difficult problems
• Improved Persistence: Better ability to stick with complex, multi-step problems
• Easier to Understand: Reasoning process is more transparent and traceable than with GPT-5
• API Name: Available as GPT-5.1 in the API

Personalization: Customizable AI Personality

With GPT-5.1, OpenAI introduces comprehensive personalization features that allow users to customize how ChatGPT responds.

Customizable Properties:

• Tone Settings: Choose between predefined styles like "Professional," "Candid," or "Quirky"
• Humor Level: Control how much humor ChatGPT brings to responses
• Formality Level: Adjustable from very formal to very informal
• Emoji Usage: Control whether and how often emojis are used
• Response Length: Preference for concise or detailed responses
• Warmth: How empathetic and friendly the responses should sound

These personalization options are saved at the account level and affect all future conversations with ChatGPT.

API Improvements for Developers

GPT-5.1 brings significant improvements for developers and agentic workflows:

No Reasoning Mode:

• New parameter reasoning_effort='none' forces the model to never use reasoning tokens
• Makes GPT-5.1 usable similarly to GPT-4.1, GPT-4o, and other non-reasoning models
• Better performance on parallel tool calls, coding tasks, instruction following, and web search tools compared to GPT-5 with 'minimal' reasoning

Extended Prompt Caching:

• Up to 24 hours cache retention (previously much shorter)
• Faster responses on follow-up questions at lower cost
• Activation via parameter: prompt_cache_retention='24h'

New Developer Tools:

• apply_patch Tool: Creates, updates, and deletes files in your codebase using structured diffs. Enables iterative, multi-step code editing workflows without custom tool descriptions
• shell Tool: Allows the model to suggest shell commands that a host integration can execute and return outputs for further reasoning. Useful for debugging, environment inspection, and automated scripts (requires strict execution policies and sandboxing)

Performance and Benchmarks

OpenAI claims "significant improvements" on AIME 2025 and Codeforces but doesn't provide specific numbers. Independent benchmarks show a mixed picture:

• AIME 2025: GPT-5.1 Thinking achieves approx. 94.0% (practically unchanged from GPT-5's 94.6%)
• Independent Tests: Benchmarks from Vals show no significant differences between GPT-5 and GPT-5.1 on AIME, SWE-bench, and Terminal Bench
• LiveCodeBench: GPT-5.1 performs slightly better than GPT-5
• LiveCodeBench Pro: GPT-5 is slightly ahead of GPT-5.1 (olympiad-level problems)

The main advantage of GPT-5.1 lies less in raw benchmark scores and more in the improved user experience through adaptive reasoning, warmer tone, and personalization options.

Availability and Rollout

GPT-5.1 was gradually rolled out starting November 12, 2025:

• Phase 1: Paid subscribers (Pro, Plus, Go, Business)
• Phase 2: Free users and non-logged-in users
• Phase 3: Enterprise and Edu users
• Legacy Access: The previous GPT-5 models remain available in a legacy dropdown for three months

GPT-5

Released: August 7, 2025

With GPT-5, OpenAI achieved a significant milestone: unifying reasoning capabilities and standard language model functionality under a single interface. The model series was released on August 7, 2025, and marks the first time that free users also get access to a reasoning model.

The GPT-5 family consists of three main variants, each optimized for different use cases and enabling flexible trade-offs between performance, cost, and latency.

GPT-5 (Main Model)

GPT-5 sets new standards in several important areas and achieves top results across various benchmarks.

Technical Specifications:

• Input limit: 400,000 tokens
• Output limit: 128,000 tokens (including invisible reasoning tokens)
• Knowledge cutoff: September 30, 2024
• Multimodal input: Text and images
• Output: Text only
• API prices: $1.25 per million input tokens, $10 per million output tokens

Benchmark Results:

• Math: 94.6% on AIME 2025 (without tools)
• Programming: 74.9% on SWE-bench Verified, 88% on Aider Polyglot
• Multimodal Understanding: 84.2% on MMMU
• Healthcare: 46.2% on HealthBench Hard
• Science: 82.1% on GPQA Diamond (PhD level)

Reasoning Levels:

GPT-5 offers four configurable reasoning levels controlled via the API parameter reasoning_effort:

• Minimal: New level for fastest responses with basic reasoning
• Low: Fast responses with moderate reasoning
• Medium: Balanced speed and reasoning depth
• High: Maximum reasoning capabilities for complex tasks

Hallucination Reduction:

GPT-5 shows significant improvements in factual accuracy: With web search enabled, responses are 45% less error-prone than GPT-4o and even 80% less error-prone than OpenAI's o3 model.

GPT-5 mini

GPT-5 mini is the efficient variant of the GPT-5 family, offering strong performance at significantly reduced costs.

Key Features:

• API prices: $0.25 per million input tokens, $2 per million output tokens
• Knowledge cutoff: May 30, 2024
• Same token limits as GPT-5 (400k input, 128k output)
• Supports all four reasoning levels
• Beats Gemini 2.5 Flash in price-performance ratio
• Ideal for high-volume applications with moderate requirements

Performance Characteristics:

• About 70-80% of GPT-5 performance on complex tasks
• Higher hallucination rate than GPT-5, but still better than previous generations
• Excellent for structured data processing and standard programming tasks
• Supports parallel tool calls and structured outputs

GPT-5 nano

GPT-5 nano is the smallest and most cost-effective model in the GPT-5 family, optimized for high-volume applications with simpler tasks.

Specifications:

• API prices: $0.05 per million input tokens, $0.40 per million output tokens
• Knowledge cutoff: May 30, 2024
• Extremely cost-effective (25x cheaper than GPT-5 for input)
• Same token limits and reasoning level support
• Hallucination rate: 7.3% on FActScore (trade-off between size and reliability)

Use Cases:

• Simple text generation and completion
• Basic classification and categorization
• Rapid prototyping
• Mass processing of simple requests
• Chatbot applications with basic requirements

GPT-5 pro

GPT-5 pro is the premium variant with extended reasoning, specifically developed for the most demanding tasks.

Outstanding Performance:

• Perfect Math: 100% on AIME 2025 with tools, 94.6% without
• PhD-Level Science: 88.4% on GPQA Diamond without tools
• Unlimited access for ChatGPT Pro users ($200/month)
• Limited access for Plus users with higher usage limits
• Extended reasoning time for the most complex problems

Special Capabilities:

• Multi-step mathematical proofs
• Complex scientific analyses
• Advanced codebase refactoring
• Detailed technical documentation
• Research-level problem solving

Shared Features of the GPT-5 Family

All GPT-5 models share important core features:

• Tool Integration: Support for web search, file search, image generation, and more
• Structured Outputs: Native support for JSON and other structured formats
• Streaming: Real-time output for better user experience
• Prompt Caching: Cost savings on repeated requests
• Batch API: Efficient mass processing
• Verbosity Parameter: Control of response verbosity (low, medium, high)

o3 and o4

Released: January 31, 2025 (o3-mini), April 16, 2025 (o3/o4-mini), June 10, 2025 (o3-pro)

With the o3 series, OpenAI introduced its latest generation of reasoning models specifically developed for complex problem-solving and logic tasks.

This series is particularly optimized for STEM applications (science, technology, engineering, and mathematics).

o3-mini (January 31, 2025):

• The latest available reasoning model
• Optimized exclusively for text processing (unlike o1, which can also process images)
• Specialized in improved reasoning capabilities
• Supports structured outputs and function calls/tools
• Enormous context window of 200,000 tokens with 100,000 token output capacity
• Trained with data up to October 2023
• Supports web search for access to current information

o3-mini-high (February 2025):

• Extended version of o3-mini with higher reasoning intensity
• Three levels of reasoning intensity selectable (low, medium, high)
• At medium reasoning level, performance equals that of the o1 model on demanding reasoning evaluations like AIME and GPQA
• 24% faster response time compared to o1-mini (average 7.7 seconds vs. 10.16 seconds)
• In A/B tests, experts preferred the responses 56% of the time over o1-mini
• 39% reduction in serious errors on complex real-world questions

The o3 series represents a significant advance in the development of AI models that can understand complex relationships and draw logical conclusions. While o1 is optimized for multimodal applications, the o3 series focuses on pure text processing with superior reasoning capabilities, especially in STEM areas.

o3-pro: The Smartest Reasoning Variant

Released: June 10, 2025

o3-pro is the premium version of the o3 model, optimized for particularly demanding tasks.

Key Features:

• Like o1-pro, o3-pro is a version of OpenAI's smartest model, designed for longer thinking and most reliable answers
• Users particularly prefer this model for areas like mathematics, science, and programming
• 87% cheaper than o1-pro with comparable or better performance
• API prices: $20 per 1 million input tokens, $80 per 1 million output tokens
• OpenAI recommends using background mode for long-running tasks to avoid timeouts
• Available for Pro and Team users, replaces o1-pro in the model picker
• Enterprise and Edu users received access one week later

o3-pro represents the pinnacle of the o3 series and offers the best reasoning performance for users willing to pay for maximum intelligence and reliability.

o4 Series: Fast and Efficient Reasoning

The o4 series was introduced in April 2025 and represents OpenAI's latest generation of small but powerful reasoning models.

o4-mini (April 16, 2025):

• Smaller model optimized for fast, cost-effective reasoning
• Achieves remarkable performance for its size and cost, especially in math, programming, and visual tasks
• Best model on the AIME 2024 and 2025 benchmarks
• For the first time, reasoning models can use and combine all tools in ChatGPT agentically
• Includes web search, Python analysis, visual input processing, and image generation
• Supports significantly higher usage limits than o3 thanks to its efficiency
• 300 messages per day for regular users
• Strong option for high volume and high throughput on questions that benefit from reasoning

o4-mini-high:

• Extended variant of o4-mini with higher reasoning intensity
• Limited to 100 messages per day due to more intensive processing
• Offers better performance on more complex tasks than o4-mini
• Same tool integration as o4-mini
• Ideal for users seeking a balance between speed and reasoning depth

The o4 series makes advanced reasoning accessible to more users and offers an efficient alternative to the larger o3 models for many use cases.

GPT-4.1

Released: April 14, 2025

The GPT-4.1 family is a new model family with comprehensive improvements, including major advances in programming, instruction following, and long context understanding.

The family includes three variants:

GPT-4.1 (Main Model):

• Processes up to 1 million tokens of context (an increase from 128,000 with GPT-4o)
• Outperforms GPT-4o at context lengths up to 128K tokens and maintains strong performance up to 1 million tokens
• Trained to follow instructions more accurately and literally than predecessors
• Significantly more reliable at recognizing relevant text and ignoring distractions
• Considerably improved diff capabilities compared to earlier GPT models
• Thomson Reuters achieved 17% better accuracy on multi-document reviews with CoCounsel

GPT-4.1 mini:

• Smaller, more efficient variant with the same context improvements
• Available in the OpenAI API and as a fallback model for free users
• Available in the model picker under "more models" for paying users
• Offers a balanced ratio between performance and cost

GPT-4.1 nano:

• The smallest and fastest model in the family
• Achieves 80.1% on MMLU, 50.3% on GPQA, and 9.8% on Aider polyglot coding
• Even outperforms GPT-4o mini on important benchmarks
• Ideal for applications that prioritize speed and efficiency

Key Differences from GPT-4o:

• Context Window: GPT-4.1 processes up to 1 million tokens vs. 128,000 with GPT-4o
• Interpretation: GPT-4.1 is more literal in interpretation and requires more specific instructions
• Reliability: Better performance at recognizing relevant information in long documents
• Programming: Superior diff capabilities for code changes
• Prompt Migration: Existing GPT-4o prompts may need adjustment as GPT-4.1 infers less user intent

The GPT-4.1 family represents a significant advance over GPT-4o, especially in processing long contexts, precise instruction following, and complex programming tasks.

GPT-4.5

Released: February 27, 2025

GPT-4.5 is OpenAI's general Large Language Model, released on February 27, 2025 initially as a "Research Preview" for developers and ChatGPT Pro subscribers.

Key Features:

• Significantly reduced error rate when outputting facts compared to GPT-4o and o1
• More natural interaction feel through improved ability to understand user intent
• Higher "emotional intelligence" (EQ) in interaction
• Better suited for tasks like writing, programming, and practical problem-solving
• Staged introduction: first for Pro subscribers ($200/month), then for Plus and Team subscribers, and finally for Edu and Enterprise users

According to OpenAI CEO Sam Altman, GPT-4.5 offers a special quality in interaction despite non-record-breaking benchmark results, which he described on Twitter as "a kind of magic I haven't felt before." This statement suggests advances in the perceived naturalness of AI interaction that go beyond purely quantitative metrics.

o1

Released: August to December 2024

With the o1 models, OpenAI introduced a new generation of language models particularly focused on "reasoning" (logical thinking and problem-solving). The latest developments in this series have further improved capabilities.

o1 (2024-12-17):

• The most powerful model in the o1 series with outstanding reasoning capabilities
• Supports both text and image processing
• Function calls and structured output possible
• Context window of 200,000 tokens with 100,000 token output capacity
• Ideal for complex problem-solving tasks and advanced programming

o1-mini (2024-09-12):

• Smaller, more efficient version of o1
• Optimized for speed and cost efficiency
• Context window of 128,000 tokens with 65,536 token output capacity
• Particularly effective for programming tasks that require speed and lower resource consumption
• Despite smaller size, significantly more capable than comparable predecessor models

The o1 models represent OpenAI's focus on deeper understanding and more complex problem-solving and pave the way for AI applications in even more demanding areas.

GPT-4o Audio

Released: October to December 2024

As an extension of the GPT-4o family, OpenAI developed specialized audio models optimized for speech processing and audio interactions.

GPT-4o-audio-preview (2024-12-17):

• Specialized in audio and text generation
• Supports conversion of text to speech and audio
• Context window of 128,000 tokens
• Ideal for applications that need natural-sounding speech output

GPT-4o-realtime-preview (2024-12-17 and 2024-10-01):

• Optimized for real-time audio processing with low latency
• Perfect for "speech in, speech out" conversational interactions
• Ideal for support agents, assistants, translators, and other applications that require fast response time
• Supports natural real-time audio conversations

These audio models significantly expand the use cases of GPT models and enable more natural, speech-based interactions with AI systems.

GPT-4o

Released: May 2024 to November 2024 (various versions)

GPT-4o ("o" stands for "omni") brought further improvements, particularly in speed and multimodal processing. It has been updated multiple times since its introduction, with the latest versions from November 2024.

General Key Features:

• Near real-time responsive conversations
• Improved image and audio processing
• Cost-effective usage (lower costs compared to GPT-4)
• More efficient processing with consistent quality
• Simultaneous understanding of text, image, and audio in a conversation

Various Versions:

• GPT-4o (2024-11-20): The latest version with improved creative writing capability, structured output, and support for parallel function calling
• GPT-4o (2024-08-06): An intermediate version with improved multimodality features
• GPT-4o (2024-05-13): The original version that already showed significant improvements over GPT-4 Turbo

GPT-4o-mini (2024-07-18):

• Faster, more affordable model as an alternative to GPT-3.5 Turbo
• Supports text and image processing
• Offers JSON mode and parallel function calls
• Ideal for applications that need a good balance between performance and cost

GPT-4o represents an important step toward natural and intuitive human-machine interaction and makes the use of AI even more accessible and intuitive.

GPT-4 Turbo

Released: November 2023

GPT-4 Turbo was introduced as a significant update to GPT-4 and brought substantial improvements in areas such as efficiency, knowledge currency, and multimodal capabilities.

Key Features:

• Significantly more efficient processing with consistent or improved quality
• More current knowledge (training data up to April 2023, later updated)
• Improved context window of 128,000 tokens
• Enhanced vision capabilities with better image analysis
• Introduction of JSON mode for structured outputs
• Improved function calling capabilities
• Optimized for developer workflows and API applications

GPT-4 Turbo stood out particularly for its better price-performance ratio and made advanced AI capabilities accessible for a broader range of applications.

It served as an important bridge between the original GPT-4 and the later introduced GPT-4o family and supported more complex multimodal applications while reducing costs.

GPT-4

Released: March 2023

GPT-4 represented another major leap forward. Although OpenAI didn't provide exact information about the parameter count, the performance was significantly increased.

Key Features:

• Multimodal capabilities (can understand and interpret images)
• Significantly improved reasoning and problem-solving ability
• Higher accuracy on complex tasks
• Longer context (can process up to 32,000 tokens, later extended to 128,000)
• Better capabilities in mathematical calculations and programming

GPT-4 outperformed human experts on various standardized tests and showed impressive capabilities in areas such as bar exams, medical knowledge tests, and programming tasks.

GPT-3.5

Released: November 2022 (as ChatGPT)

Strictly speaking, GPT-3.5 is not a completely new model but a further development of GPT-3. However, it became the foundation for ChatGPT, which fundamentally changed how the general public perceives AI.

Key Features:

• Optimized for dialogues and conversations
• Trained through Reinforcement Learning from Human Feedback (RLHF)
• Better handling of harmful, unethical, or misleading requests
• Ability to recognize and admit mistakes
• Memory within a conversation

ChatGPT was the first AI application to reach one million users within just 5 days and became synonymous with generative AI. With this model, the true AI boom began for the general public.

GPT-3

Released: June 2020

With GPT-3, the big breakthrough came. The model was so much more powerful than its predecessors that it attracted attention beyond the AI community. GPT-3 was the first model that truly showed that AI might be capable of writing human-like text.

Key Features:

• 175 billion parameters (100 times more than GPT-2)
• Training on even more data, including books, Wikipedia, and various websites
• Ability to generate code
• Completing various tasks without specific training (few-shot learning)
• Significantly improved understanding of context and nuances

GPT-3 was offered commercially via an API and formed the basis for numerous applications, from text creation to programming. It made clear that we had arrived at a turning point in AI development.

GPT-2

Released: February 2019

GPT-2 caused a stir when OpenAI initially decided not to release the full model immediately, out of concern about possible misuse, such as creating fake news. This decision sparked a debate about ethical issues in AI research.

Key Features:

• 1.5 billion parameters (more than 10 times as many as GPT-1)
• Training on 8 million websites
• Significantly improved text generation with better coherence
• Could generate plausible continuations to given text beginnings

GPT-2 already showed impressive capabilities and made clear that the chosen path (simply more parameters and more training data) led to increasingly better results.

GPT-1

Released: June 2018

GPT-1 was OpenAI's first attempt to develop a large language model. With 117 million parameters (the "tuning knobs" of the model), it was impressive for its time but seems almost tiny compared to today's models.

Key Features:

• 117 million parameters
• Training on a dataset of books
• Could generate simple texts, but with clear weaknesses in coherence and logic

GPT-1 served primarily as a proof of concept and showed the potential of the Transformer architecture for text generation. However, the practical applications were still severely limited.

Comparison of All GPT Models

To better understand the development and differences between the models, here's an updated tabular comparison:

It gets even more revealing when you put coding performance in relation to API price and compare it with Claude and Gemini.

GPT-5.3 Codex is the best price-performance model for code

Ideal: strong + cheap

OpenAI

Anthropic

Google

DeepSeek

Moonshot AI

Efficiency frontier (best price-performance)

Sources: OpenAI, Anthropic, Google DeepMind

PNG|SVG

CC BY 4.0

gradually.ai