# GPT-4o

The "omni" model that unified text, audio, and vision in a single natively multimodal architecture, with audio response speeds averaging 320ms, a 16x improvement over prior voice pipeline approaches.

- **Model ID:** `openai/gpt-4o`
- **Type:** chat
- **Providers:** azure, openai
- **Context window:** 128,000
- **Maximum output tokens:** 16,384
- **Pricing:** $2.5/1M input tokens, $10.0/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/gpt-4o

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'openai/gpt-4o',
  prompt: 'Why is the sky blue?'
})
```

## About

GPT-4o was announced on May 13, 2024 at OpenAI's Spring Updates event. The "o" stands for "omni," reflecting the model's foundational design: rather than connecting separate specialist models for different modalities, GPT-4o was trained end-to-end across text, audio, image, and video. This architectural choice enables sub-400-millisecond audio responses. Prior approaches chained a speech recognition model, a language model, and a text-to-speech model together, introducing latency at each boundary. GPT-4 Turbo-based voice averaged 5.4 seconds per turn.

GPT-4o matched GPT-4 Turbo on text and code in English while costing less in the API, with notable improvements on non-English text. This made it the default for developers who previously used GPT-4 Turbo: an upgrade in multimodal capability at a lower price.

The model accepts any combination of text, audio, image, and video as input and can generate text, audio, and image outputs. This flexibility spans real-time voice assistants, vision pipelines that analyze photographs or documents, and agents that process video frames alongside textual context.

## What to consider

For applications that mix modalities (for example, a voice interface that also accepts image uploads), a single model endpoint simplifies both architecture and cost accounting compared to separate specialist models.

## When to use

### Best For

- Real-time voice applications where the native audio processing eliminates pipeline latency between speech recognition, reasoning, and synthesis
- Mixed-modality workflows that process both images and text together, such as visual question answering or document analysis with figures
- Applications that need GPT-4-class text quality with lower API cost than GPT-4 Turbo
- Multilingual applications that benefit from GPT-4o's improved non-English text performance
- Products that may expand from text-only to multimodal inputs over time and want a single model to support that evolution

### Consider Alternatives When

- Maximum coding capability is required, and GPT-4.1 benchmarks are meaningfully better for your specific use case
- Cost is the primary driver and GPT-4.1 mini provides sufficient quality
- The task requires deep multi-step reasoning where the o-series reasoning models' chain-of-thought approach is superior

## Best for

- **Real-time voice applications:** Native audio processing eliminates pipeline latency between speech recognition, reasoning, and synthesis
- **Mixed-modality workflows:** Processing both images and text together for visual question answering or document analysis with figures
- **Cost-efficient GPT-4 quality:** Applications that need GPT-4-class text quality with lower API cost than GPT-4 Turbo
- **Multilingual applications:** Products that benefit from GPT-4o's improved non-English text performance
- **Evolving multimodal products:** Apps that may expand from text-only to multimodal inputs over time with a single model supporting that evolution

## Consider alternatives

- **Maximum coding capability:** GPT-4.1 benchmarks are meaningfully better for some coding use cases
- **Cost-driven workloads:** GPT-4.1 mini provides sufficient quality at lower cost
- **Deep multi-step reasoning:** The o-series reasoning models' chain-of-thought approach is superior for complex analytical problems

## Frequently asked questions

### What does "omni" mean in GPT-4o's name?

It reflects the model's end-to-end native training across text, audio, image, and video modalities, rather than being a combination of separate specialist models connected by a pipeline.

### How much faster is GPT-4o's audio response compared to earlier voice pipelines?

GPT-4o averages 320 milliseconds for audio responses; the prior GPT-4 Turbo-based voice approach averaged 5.4 seconds, making GPT-4o approximately 16x faster for voice.

### How does GPT-4o's API pricing compare to GPT-4 Turbo?

GPT-4o launched at lower API cost than GPT-4 Turbo while matching its performance on English text and code, and improving on non-English languages.

### What input and output modalities does GPT-4o support?

Inputs: text, audio, image, video. Outputs: text, audio, and image. This breadth makes it flexible for diverse multimodal application architectures.

### Is the "gpt-4o" model alias the same as a specific dated snapshot?

No. The alias `gpt-4o` points to the latest stable version, which may be updated over time. Dated snapshots like gpt-4o-2024-05-13 or gpt-4o-2024-11-20 pin to specific releases.

### Does routing GPT-4o through AI Gateway add latency?

AI Gateway is designed as a lightweight routing layer. For most applications, the observability, caching, and authentication benefits outweigh any marginal overhead.

### What are typical latency characteristics?

This page shows live throughput and time-to-first-token metrics measured across real AI Gateway traffic.

## Links

- [Model page](https://vercel.com/ai-gateway/models/gpt-4o)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://learn.microsoft.com/en-us/azure/ai-foundry/foundry-models/concepts/models-sold-directly-by-azure?pivots=azure-openai&tabs=global-standard-aoai%2Cstandard-chat-completions%2Cglobal-standard#gpt-5.1)
- [Provider pricing](https://platform.openai.com/docs/pricing)