.svg)
GPT-4 and GPT-5 are two generations of OpenAI models released more than two years apart. This guide sticks to what OpenAI has officially published: what each model is, how they differ, what they cost in the API and where they are available today.
.jpeg)
What is GPT-4?
OpenAI announced GPT-4 on March 14, 2023. It is a large multimodal model: it accepts image and text inputs and produces text outputs (OpenAI, GPT-4). Image input started as a research preview; it is covered in our guide to GPT-4 Vision.
Key facts from the launch page:
- On a simulated bar exam, GPT-4 scored around the top 10% of test takers, while GPT-3.5 scored around the bottom 10%.
- It scored 40% higher than GPT-3.5 on OpenAI's internal adversarial factuality evaluations and was 82% less likely to respond to requests for disallowed content.
- It was fine-tuned with reinforcement learning from human feedback (RLHF), and its pre-training data mostly stopped in September 2021.
- The API offered an 8,192-token context window, plus a 32,768-token version (gpt-4-32k) with limited access.
OpenAI did not disclose GPT-4's size. The technical report states that it "contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar" (GPT-4 Technical Report). Any parameter count you see for GPT-4 is an estimate, not an official figure. OpenAI later released faster variants, covered in our article on GPT-4 Turbo.
What is GPT-5?
OpenAI introduced GPT-5 on August 7, 2025 (OpenAI, Introducing GPT-5). Instead of asking users to pick a model, ChatGPT used a unified system with:
- a smart, efficient model that answers most questions;
- a deeper reasoning model, GPT-5 thinking, for harder problems;
- a real-time router that chooses between them based on the conversation, the complexity of the request and explicit intent (for example asking it to "think hard").
When usage limits were reached, a mini version took over, and GPT-5 pro offered extended reasoning for the most demanding tasks. GPT-5 became the default model in ChatGPT, replacing GPT-4o, o3, o4-mini, GPT-4.1 and GPT-4.5, and was also available to Free users, who moved to GPT-5 mini after their limits.
For developers, OpenAI released three sizes, gpt-5, gpt-5-mini and gpt-5-nano, with new controls such as a verbosity parameter, a "minimal" reasoning effort and custom tools (OpenAI, GPT-5 for developers).
Comparison of GPT-4 vs GPT-5
1. Architecture
GPT-4 is a single model. GPT-5 in ChatGPT is a system of models with a router, which is why the same chat can give a quick answer or take time to reason. Neither model's parameter count has been published by OpenAI.
2. Context window
GPT-4 launched with 8,192 tokens (32,768 for gpt-4-32k). In the API, GPT-5 accepts up to 272,000 input tokens and can return up to 128,000 output tokens, for a total context of 400,000 tokens.
3. Benchmarks
OpenAI reported these results for GPT-5 at launch:
- AIME 2025 (math, without tools): 94.6%
- SWE-bench Verified (real-world coding): 74.9%
- Aider Polyglot (coding): 88%
- MMMU (multimodal understanding): 84.2%
- HealthBench Hard: 46.2%
- GPQA (PhD-level science questions), GPT-5 pro: 88.4%
GPT-4 was evaluated on different tests in 2023 (bar exam, MMLU and other academic exams), so these numbers should not be read as a direct head-to-head. OpenAI's own comparisons for GPT-5 are against GPT-4o and o3, not the original GPT-4.
4. Reliability
According to OpenAI, with web search enabled GPT-5's responses were about 45% less likely to contain a factual error than GPT-4o, and about 80% less likely than o3 when GPT-5 was thinking. OpenAI also reported cutting sycophantic replies from 14.5% to less than 6% in its evaluations. For GPT-4, the reference point was GPT-3.5: 40% higher on internal factuality evaluations.
5. API pricing
At launch, GPT-4 cost $0.03 per 1,000 prompt tokens and $0.06 per 1,000 completion tokens ($0.06 and $0.12 for the 32K version). GPT-5 launched at $1.25 per million input tokens and $10 per million output tokens; gpt-5-mini at $0.25 and $2; gpt-5-nano at $0.05 and $0.40. Check OpenAI's pricing page before budgeting, as prices change.
6. Personalization
GPT-5 brought four preset personalities in ChatGPT (Cynic, Robot, Listener and Nerd). GPT-4 had no equivalent presets; customization went through custom instructions and, at the time, ChatGPT plugins.
Which model can you use today?
- GPT-4 was retired from ChatGPT on April 30, 2025 and replaced by GPT-4o. It remained available in the API (ChatGPT release notes).
- GPT-5 (Instant and Thinking) was retired from ChatGPT on February 13, 2026, along with GPT-4o, and remains available in the API (OpenAI Help Center).
- ChatGPT now runs newer models; OpenAI's release notes list the current lineup, including GPT-6, launched in ChatGPT on October 7, 2026.
In practice, the GPT-4 vs GPT-5 choice now mostly concerns developers maintaining API integrations: GPT-5 is far cheaper per token and offers a much larger context, while GPT-4 mainly matters for legacy applications.
Conclusion
GPT-5 changed how OpenAI's models are delivered: a routed system with a reasoning mode, much larger context and much lower API prices than GPT-4 at launch. Both models have since been replaced in ChatGPT, so rely on OpenAI's release notes for the current model and treat unofficial figures, such as parameter counts, with caution.
FAQ
When were GPT-4 and GPT-5 released?
GPT-4 was announced on March 14, 2023 and GPT-5 on August 7, 2025.
How many parameters do GPT-4 and GPT-5 have?
OpenAI has not published the parameter count of either model. Figures circulating online are estimates.
Is GPT-5 cheaper than GPT-4 in the API?
Yes. At launch, GPT-5 cost $1.25 / $10 per million input / output tokens, against $30 / $60 per million for GPT-4.
Can I still use GPT-4 or GPT-5 in ChatGPT?
No. GPT-4 left ChatGPT on April 30, 2025 and GPT-5 on February 13, 2026. Both remain available in the API.
