GPT-4o is one of the most talked about AI models of recent years. OpenAI launched it in May 2024 as a fast model that could work with text, images and audio. Millions of people used it in ChatGPT, and developers built apps on top of it.
Much has changed since then. OpenAI removed GPT-4o from ChatGPT in February 2026, and its API status is now more complicated. This guide explains what the model is, how it works, what it costs, where you can still use it, and how people reacted when it left.
What is GPT-4o?
It is a large language model built by OpenAI. The letter o stands for omni, which means every. The name points to the main idea. One model handles text, images, audio and video input, and it can reply in text, audio or images.
Older systems used separate models for each job. GPT-4o replaced that chain with a single neural network trained on text, vision and audio together. This made replies faster and more natural, especially in voice chats.
- Maker: OpenAI
- Announced: May 13, 2024
- Type: multimodal large language model
- Input: text, images, audio and video
- Output: text, audio and images
- Languages: more than 50
- Status: retired from ChatGPT on February 13, 2026
GPT-4o by OpenAI: Release Date and Background
OpenAI announced GPT-4o on May 13, 2024, at its Spring Update event. Chief technology officer Mira Murati led a live demo. OpenAI said it offered GPT-4 level intelligence with faster speed and stronger voice, text and vision skills. Features rolled out in stages over the following weeks.
It followed GPT-4 from March 2023 and GPT-4 Turbo from November 2023. A smaller sibling, GPT-4o mini, arrived on July 18, 2024. In October 2024, OpenAI also published a system card with safety tests and capability details.
| Date | Event |
| May 13, 2024 | Model announced, text and image rollout begins |
| July 18, 2024 | Mini version released |
| August 2024 | Structured Outputs added to the API |
| September 2024 | Advanced Voice Mode reaches Plus and Team users |
| October 1, 2024 | Realtime API introduced |
| October 25, 2024 | System card published |
| November 20, 2024 | Update raises maximum output to 16,384 tokens |
| April 30, 2025 | GPT-4 retired from ChatGPT and replaced by this model |
| February 13, 2026 | Removed from ChatGPT |
What is GPT-4o mini?
The mini model is a smaller, cheaper version of the main model. OpenAI released it on July 18, 2024, and it replaced GPT-3.5 Turbo as the ChatGPT default small model. It also gave free users a GPT-4 class model for the first time. Reports say it came from distillation, where a small model learns from a large one.
It scored 82 percent on the MMLU benchmark, against 69.8 percent for GPT-3.5 Turbo. OpenAI said it costs about 60 percent less than that older model. Developers use it for high volume jobs such as support replies, receipt processing and email drafts.
| Feature | Main model | Mini model |
| Launch date | May 13, 2024 | July 18, 2024 |
| API input price now | $2.50 per million tokens | $0.15 per million tokens |
| API output price now | $10.00 per million tokens | $0.60 per million tokens |
| Context window | 128,000 tokens | 128,000 tokens |
| Maximum output | 16,384 tokens | 16,384 tokens |
| Knowledge cutoff | October 2023 | October 2023 |
| Best for | Complex and multimodal work | High volume, simple tasks |
How does GPT-4o work?
The model is a Generative Pre-trained Transformer. It was trained end to end on text, images and audio at the same time. Because of this, one network can read a photo, hear a voice and write a reply without passing work to other models.
Earlier voice mode used three steps in a chain. Tone, background noise and extra speakers were lost at the first step. The single model keeps that detail. A new tokenizer also helps, because it splits many languages into fewer tokens.
| Step | Old voice mode | Omni model |
| Hearing | Whisper turned speech into text | The same network hears the audio |
| Thinking | GPT-4 saw only the text | The model also sees tone, noise and speakers |
| Speaking | A text to speech model read the reply | The same network creates the audio |
| Reply time | 2.8 seconds with GPT-3.5, 5.4 seconds with GPT-4 | 0.32 seconds on average |
- Hindi, Marathi, Tamil, Telugu and Gujarati need 2.9 to 4.4 times fewer tokens.
- Arabic needs about 2 times fewer tokens.
- Chinese, Japanese, Korean and Vietnamese need 1.4 to 1.7 times fewer tokens.
GPT-4o Parameters: How Big Is the Model?
OpenAI has not shared the parameter count for GPT-4o. It also keeps the size of its other recent models private. Any number you see online is an outside guess. It is not a confirmed figure.
Judge GPT-4o by results instead. The published specs are more useful, and the table below lists them. The bullets show how it performed on speed, price and quality.
| Spec | Value |
| Context window | 128,000 tokens |
| Maximum output | 16,384 tokens (4,096 at launch) |
| Knowledge cutoff | October 2023 |
| Parameter count | Not disclosed |
- Speed: about 110 tokens per second in one independent test
- Price: $2.50 input and $10 output per million tokens
- Quality: first place on four of six launch benchmarks
- Audio reply time: 320 milliseconds on average
What can GPT-4o do?
| Type | Input | Output | Note |
| Text | Yes | Yes | Chat, summaries, writing and code |
| Images | Yes | Yes in ChatGPT | The API model returns text only |
| Audio | Yes | Yes | Powered voice chats in ChatGPT |
| Video | Yes, as a camera or screen feed | No | OpenAI’s Sora makes video, not this model |
Voice and Audio
The model can hold spoken talks with an average reply time of 320 milliseconds. Its fastest replies take 232 milliseconds. That is close to human pace in conversation.
- Real time translation across more than 50 languages
- Speech in different emotive styles, including singing
- Reading mood from the tone of your voice
- Interrupting mid reply, which was promised at launch and arrived later
Vision, Text and Code
It can describe photos, read handwriting and answer questions about a screenshot. It also writes, summarizes, answers questions and helps with code. Sharing your screen while you talk can be simpler than uploading files.
- Menu and sign translation from a photo
- Chart reading and data analysis
- Code writing, explaining and debugging
- File uploads for summaries and Q&A
Learn more…….What Is Financial Management? Scope, Importance
GPT-4o vs GPT-4 and GPT-4 Turbo
GPT-4 launched in March 2023 and handled text well, but it relied on other tools for audio and images. GPT-4 Turbo added a 128,000 token window and better image reading. The newest of the three merged text, vision and audio into one model.
In practice, text and code quality is close between GPT-4 Turbo and the newer model. The big gains are speed, cost, vision, audio and non English text. OpenAI retired GPT-4 from ChatGPT on April 30, 2025, and said the newer model replaced it.
| Feature | GPT-4 | GPT-4 Turbo | GPT-4o |
| Release date | March 14, 2023 | November 2023 | May 13, 2024 |
| Context window | 8,192 tokens | 128,000 tokens | 128,000 tokens |
| Knowledge cutoff | September 2021 | December 2023 | October 2023 |
| Inputs | Text, limited images | Text, images | Text, images, audio |
| Voice design | Separate models | Separate models | One model |
| Non English token use | Standard | Standard | 1.4 to 4.4 times fewer |
| Launch API price per million tokens | $30 input, $60 output | $10 input, $30 output | $5 input, $15 output |
GPT-4o Performance and Benchmarks
OpenAI tested GPT-4o against GPT-4 Turbo, Claude 3 Opus, Gemini 1.5 Pro and others on six tests. It ranked first on four. These results date from 2024, and newer models have passed them.
Gains over GPT-4 Turbo on text were only a few percentage points. The bigger jump was speed. One independent test measured about 110 tokens per second, roughly three times faster than GPT-4 Turbo. Benchmarks also miss how well a model handles voice and images.
GPT-4o Use Cases
The model suits jobs where speed and mixed media matter. Its low delay made live conversation practical in ChatGPT for the first time. OpenAI’s demos showed math tutoring, code help and spoken translation between English and Spanish.
| Use case | Example task |
| Customer support | A chatbot that answers in real time |
| Translation | Live talk between an English and a Spanish speaker |
| Practice | Spoken rehearsal for interviews and sales calls |
| Coding | Explaining a plot or a script while you share your screen |
| Data work | Reading a chart and suggesting next steps |
| Accessibility | Describing a scene aloud for people with low vision |
Teams also tested it in law, health and education for summaries, record analysis and interactive lessons. Human review is still needed in high stakes areas.
How Good Is GPT-4o?
Reviewers praised the speed and the natural voice. They also found real gaps. Zapier’s writer saw wrong Sudoku grids and misread handwriting. DataCamp’s writer got made up football stats and a wrong plant species.
The common problem was confidence. The model often gave a fluent answer when it should have said it was unsure. Always check facts, numbers and any text it reads from images or audio.
| Worked well | Fell short |
| Fast, natural voice replies | Misread a Sudoku grid and added random numbers |
| Reading messy handwriting | Swapped one outline point for a similar phrase |
| Explaining code and plots | Reported football stats wrongly |
| Finding sources on the web | Named the wrong plant species |
| Restyling a photo as a painting | Kept little resemblance to the original |
GPT-4o Limitations and Risks
GPT-4o can hallucinate, which means it states false details in a confident tone. Its knowledge stops at October 2023. Its 128,000 token window is smaller than Gemini 1.5 Pro, which claims 2 million. Its reasoning is weaker than OpenAI’s o series models.
Audio adds new risks. Cloned voices could fuel scam calls, so OpenAI limited output to four preset voices at launch. On its Preparedness Framework, OpenAI rated the model Medium risk because of persuasion. It scored Low in cybersecurity, biological threats and model autonomy.
- Some answers can show bias.
- Tricky adversarial prompts can fool it.
- Strong accents and technical words cause transcription errors.
- OpenAI’s own outtakes showed failed translation between two non English languages.
| Preparedness area | Rating |
| Cybersecurity | Low |
| Biological, chemical, radiological and nuclear threats | Low |
| Persuasion | Medium |
| Model autonomy | Low |
How much does GPT-4o cost?
In ChatGPT, GPT-4o was free with limits. Plus cost $20 a month and gave about five times more messages. API pricing is charged per million tokens, and it has dropped since launch.
| Model | Input | Output |
| Main model at launch | $5.00 | $15.00 |
| Main model now | $2.50 | $10.00 |
| Mini model | $0.15 | $0.60 |
Cached input costs $1.25 per million tokens. The new tokenizer also lowers bills in many languages because fewer tokens are needed. Prices can change, so check OpenAI’s pricing page before you budget.
- Example job: 1 million input tokens and 200,000 output tokens
- Main model at launch prices: $8.00
- Main model now: $4.50
- Mini model: $0.27
| ChatGPT plan in December 2024 | Price | Access |
| Free | $0 | Limited messages, some features locked |
| Plus | $20 a month | Full access |
| Pro | $200 a month | Full access plus o1 models |
| Team | $25 per user a month | Group access |
GPT-4o API
Developers reached GPT-4o through Chat Completions, Responses, Batch and other endpoints. It accepts text and image input and returns text, including Structured Outputs. Streaming, function calling, fine tuning and predicted outputs are supported. Audio and video are not supported on the main model.
Audio versions were separate models, such as the realtime and audio variants. OpenAI has scheduled those for shutdown on January 20, 2027. The November 2024 and August 2024 snapshots remain listed. The May 2024 snapshot is deprecated.
| Usage tier | Requests per minute | Tokens per minute |
| Tier 1 | 500 | 30,000 |
| Tier 2 | 5,000 | 450,000 |
| Tier 3 | 5,000 | 800,000 |
| Tier 4 | 10,000 | 2,000,000 |
| Tier 5 | 10,000 | 30,000,000 |
How to access GPT-4o
You can no longer pick it in ChatGPT. The main route today is the OpenAI API. Some cloud platforms have offered it too, but they follow their own retirement schedules. If you only want a chatbot, use the newer models in ChatGPT.
- Create an OpenAI API account.
- Add billing and create an API key.
- Send a request to the Chat Completions or Responses endpoint with the GPT-4o model name.
- Watch OpenAI’s deprecations page for shutdown dates.
python
from openai import OpenAI
client = OpenAI()
reply = client.chat.completions.create(
model=”gpt-4o”,
messages=[{“role”: “user”, “content”: “Hello”}],
)
print(reply.choices[0].message.content)
Store your key in an environment variable, not in your code. Test on a small budget first. Rate limits depend on your usage tier.
Chat GPT-4o Free
At launch on May 13, 2024, free ChatGPT users got GPT-4o with tighter message limits. Plus users got five times more messages. After free users hit the cap, ChatGPT switched them to a smaller model.
That is no longer possible, because the model left ChatGPT on February 13, 2026. The free plan now runs a newer model. Treat any site that promises free access with care.
| Topic | May 2024 | Now |
| Model on the free plan | Main model with a cap | A newer model |
| After the cap | Switch to a smaller model | Not applicable |
| Text chats | Limited | Generous, with guardrails |
| Files, images and voice | Limited, some locked | Capped |
| Price of the free plan | $0 | $0 |
Chat GPT-4o Download
You cannot download the model itself. It is a closed model that runs on OpenAI’s servers. What you can download is the ChatGPT app for iPhone, Android and desktop.
The macOS app launched with the model on May 13, 2024. Today those apps run newer models. Any file that claims to be a GPT-4o download is a red flag. It could be malware or a scam.
- Safe: the official ChatGPT app from the App Store or Google Play
- Safe: the desktop app from OpenAI’s own site
- Risky: unknown APK files
- Risky: unofficial installers that use OpenAI’s name
GPT-4o on Hugging Face
OpenAI never released the weights, so the model is not on Hugging Face as an open download. Pages that use its name are usually demos, datasets or apps that call OpenAI’s API.
If you want an open model from OpenAI, look at its gpt-oss family, released with open weights in 2025. Other open families on the platform include Llama, Mistral and Qwen.
- gpt-oss from OpenAI
- Llama from Meta
- Mistral models
- Qwen models
What People Say About GPT-4o on Reddit
Reddit became a main home for fans of the model. Users praised its warm, casual tone and its creative writing. Many called newer models colder. In August 2025, similar backlash pushed OpenAI to bring it back for paid users.
The #Keep4o campaign grew on Reddit and X before the 2026 shutdown. A Change.org petition passed 20,000 signatures. Some users cancelled plans or tried rival chatbots. OpenAI said only 0.1 percent of daily users still chose the model.
- Praise: warm tone, creative help and a steady style
- Complaint: only about two weeks of notice
- Concern: some people leaned on it for emotional support
- Other side: many users moved on without trouble
GPT-4o Retirement Date
GPT-4o Retirement Date is the day OpenAI stopped offering the model in a product. GPT-4o left ChatGPT on February 13, 2026, while the API version has separate shutdown dates.
In ChatGPT
OpenAI announced the change in late January 2026. The model left ChatGPT on February 13, 2026, along with GPT-4.1 and other older models. Business, Enterprise and Edu plans kept it in custom GPTs until April 3, 2026. After that, it was fully retired across ChatGPT.
In the API
The API works differently. OpenAI removed the chat latest snapshot on February 17, 2026. The dated May 2024 snapshot shuts down on October 23, 2026. OpenAI’s help page says the main model stays in the API, with no shutdown date listed yet. OpenAI promises at least six months of notice for generally available models.
| Date | What happens |
| February 13, 2026 | Removed from ChatGPT |
| February 17, 2026 | Chat latest API snapshot removed |
| April 3, 2026 | Ends in custom GPTs for Business, Enterprise and Edu |
| October 23, 2026 | May 2024 API snapshot shuts down |
| January 20, 2027 | Realtime and audio versions shut down |
| February 26, 2027 | Transcription versions shut down |
- Move any app on the May 2024 snapshot before October 23, 2026.
- Test your prompts on a newer model.
- Check OpenAI’s deprecations page every month.
- In ChatGPT, use the newer GPT-5 series models that replaced it.
Frequently Asked Questions
Is ChatGPT 4o available for free?
It was at launch, with message limits. It is no longer in ChatGPT, so free users now get a newer model.
What’s the difference between GPT-4 and GPT-4o?
GPT-4 mainly handles text and images through separate tools. GPT-4o is one model for text, images and audio, and it is faster and cheaper.
Is ChatGPT 4o gone?
Yes, in ChatGPT. It was retired on February 13, 2026. Parts of it remain in the API for now.
How to use GPT-4o?
Use the OpenAI API and call the model by name. It can no longer be selected in ChatGPT.
Is there a 100% free ChatGPT?
Yes. ChatGPT has a free plan that costs nothing. It has limits on files, images and voice.
Conclusion
GPT-4o showed what one model for voice, vision and text could do. It cut delay to human speed, lowered costs and improved support for many languages. Those ideas now shape most leading AI systems.
Today the model is a piece of AI history and a smaller API option. Most people should use the newer models in ChatGPT. Developers should check OpenAI’s deprecations page and plan a move before the October 23, 2026 shutdown of the May 2024 snapshot.
