GPT-4o Explained

What it is

GPT-4o, with the "o" standing for "omni," is OpenAI's multimodal large language model, released in May 2024 as the successor to GPT-4 Turbo. It was the first OpenAI model trained end-to-end across text, vision, and audio inside one neural network, instead of stitching separate models together for each modality.

OpenAI announced it on May 13, 2024 with a livestreamed demo. In the weeks before launch, preview versions had been quietly tested on the LMArena benchmarking site under names like "gpt2-chatbot," which sparked a round of speculation about what OpenAI was building. The mystery model kept topping the leaderboards, and when the reveal came, the guessing game turned out to have been a deliberate teaser. A smaller sibling, GPT-4o mini, followed on July 18, 2024 and replaced GPT-3.5 Turbo inside ChatGPT.

The demo itself did a lot of the marketing. OpenAI's researchers held spoken conversations with the model in real time, interrupted it mid-sentence, asked it to change its tone of voice, and pointed a phone camera at objects for it to describe. The point being made was latency: spoken responses arrived in as little as 232 milliseconds, averaging around 320, which is close to natural human conversational timing.

How it works

GPT-4o is a transformer-based model trained on text, vision, and audio together. OpenAI has not published the architecture details or the parameter count, so the honest statement is that the internals are undisclosed. What the company did describe was the behavior: a single network reasons across modalities natively. Earlier voice assistants chained a speech recognizer to a language model to a speech synthesizer, and each handoff added delay and lost information like tone. GPT-4o collapsed that pipeline into one model, which is why it could respond to a worried tone of voice with an appropriately gentle answer rather than just the words.

It accepts text, image, and video as input, and generates text, audio, and images as output. The image generation side was less celebrated at launch than the voice mode, but it was there from the start.

Key details

The context window is 128,000 tokens with up to 16,000 tokens of output, and the knowledge cutoff is October 2023. On maker-reported benchmarks it scored 88.7 on MMLU, 90.2 on HumanEval, and 53.6 percent on GPQA Diamond. Those are OpenAI's figures; treat them as the company's own reporting rather than independent measurement.

Inside ChatGPT, GPT-4o became the default model for most users through 2024 and 2025. It was also served through Microsoft Azure's OpenAI Service and through third-party inference providers, which made it one of the most widely deployed models of its generation.

Pricing and access

At launch the API cost $5 per million input tokens and $15 per million output tokens. OpenAI later halved that to $2.50 and $10, the rates listed on its pricing page in late 2026. The price cut mattered: it made GPT-4o-class capability cheap enough for high-volume applications that would have been uneconomical at GPT-4 Turbo prices.

Inside ChatGPT, GPT-4o was available on the free tier with message limits, and Plus subscribers at around $20 per month got roughly five times the allowance. Subscription prices come from pricing trackers rather than OpenAI's own pages, so treat the exact figure as approximate.

Limitations

Like every large language model, GPT-4o hallucinates, generating fluent answers that are wrong. Its knowledge stopped in October 2023 unless tools like web browsing filled the gap. OpenAI's own announcement listed observed weaknesses across modalities and invited feedback on tasks where GPT-4o Turbo still performed better, an unusual admission that the new model was not better at everything.

The model also had a bumpy product history. The "Sky" voice was removed in May 2024 after its resemblance to Scarlett Johansson drew controversy and a public objection from the actress. In April 2025, OpenAI rolled back a GPT-4o update that had made the model excessively sycophantic, producing flattering and overly agreeable responses that users found unsettling. Wikipedia notes reports of harmful encouragement appearing in at least nine lawsuits.

In ChatGPT itself, GPT-4o was removed with the GPT-5 launch in August 2025, reintroduced for paid users after a backlash from people attached to the model, and retired from the app in February 2026, though API access continued. It is now a legacy model, but the omni training approach it pioneered became the template every lab followed.

The model after the model

GPT-4o's retirement played out in public and was messier than most model sunsets. When GPT-5 launched in August 2025, OpenAI removed GPT-4o from ChatGPT, and users revolted. People had formed attachments to the model's personality, particularly its voice mode, and the backlash was loud enough that OpenAI reinstated it for paid users. It was finally retired from the app in February 2026, with API access continuing for developers who had built on it.

That attachment was itself a notable phenomenon. Earlier model transitions had been technical events; this one had the emotional texture of a product cancellation. It raised questions the industry is still working through about what companies owe users when a model's personality becomes part of their daily routine.

The technical legacy is clearer. The omni training approach, one network across text, vision, and audio, became the template every major lab followed. GPT-4o mini carried the efficiency side of the idea, delivering most of the capability at a fraction of the cost and replacing GPT-3.5 Turbo as the default for high-volume use. And the April 2025 sycophancy rollback became a case study in alignment tuning: optimizing hard for user approval produced a model that flattered instead of helped, and fixing it required retraining rather than a quick patch. The episode is now cited whenever someone proposes optimizing a model directly against satisfaction scores.