OpenAI's omni-era workhorse — text and image in, text out, with a cached-input rate that halves the cost of repeated context.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
The name promises more than the API delivers, and that is the first thing worth clearing up.
GPT-4o was introduced as OpenAI's omni-era model, and the demos showed voice and vision working together in real time. Through the API you get text and image input, and text output. No generated speech, no generated images. If you budgeted for a single model to handle a voice product end to end, budget again.
What you do get is a dependable general-purpose model with one genuinely clever pricing feature.
GPT-4o runs about $2.50 per million input tokens and $10.00 per million output tokens. Cached input drops to roughly $1.25 per million — half price.
That discount is not a rounding detail; it changes which architectures make sense. Any design that resends a stable prefix on every call — a long system prompt, a product catalogue, a policy document, a fixed few-shot block — pays half for that portion once it is cached. Structure prompts so the stable material sits at the front and the variable part at the end, and the saving arrives without changing a line of application logic.
Get the order backwards and you pay full price for the same tokens. It is the cheapest optimisation available here and the one most often missed.
The knowledge cutoff is October 2023.
Measured against today that is a long silence. Anything released, renamed, repriced or deprecated since then is invisible to the model unless you put it in the prompt yourself. For a support assistant answering questions about your own product that hardly matters, because your documentation is in context anyway. For anything that reasons about the current state of the world, it matters a great deal, and retrieval stops being optional.
Context is 128,000 tokens. Respectable when it launched, ordinary now that million-token windows are widely available.
Output ceiling is 16,384 tokens — double what several contemporaries allowed, and enough for a substantial file or a long structured response in one pass.
It calls tools reliably, it handles images without a separate vision pipeline, and it is probably the most thoroughly documented model in production anywhere. When something behaves oddly at two in the morning, someone has already written up the fix. That is worth more than a benchmark point.
| Field | Value |
|---|---|
| Provider | OpenAI |
| Context window | 128,000 tokens |
| Maximum output | 16,384 tokens |
| Knowledge cutoff | October 2023 |
| Input | text, image |
| Output | text |
| Tool calling | Supported |
| Reasoning mode | Not supported |
| Indicative pricing | $2.50 in / $1.25 cached in / $10.00 out per million |
| Released | 13 May 2024 |
Pricing and availability change; confirm against your own account before you plan a budget around them.
See how it stands against newer options in the AI model directory, or read Gemini 2.5 Flash if the 128K window is what you are trying to escape.
Use the recorded facts as a shortlist, then validate the model against representative inputs and production constraints.
GPT-4o is categorized for Audio-text-to-text, Automatic speech recognition, Image-text-to-text, Image-to-text, Text generation, Text-to-text AI models. Its current record accepts text, image, audio and produces text. Confirm file formats, preprocessing, and provider-specific request schemas before implementation.
The directory records 128K context and 16.4K maximum output. Listed token prices are $2.50 input and $10.00 output per 1M tokens. Treat missing values as unknown and recheck current commercial terms.
Recorded output speed is 110 t/s and time to first token is 0.42s. Hosting route, region, prompt length, concurrency, and provider load can materially change both measurements.
This page separates sourced model facts from incomplete fields. Benchmark evidence is displayed only when its source is linked and verified. Before choosing GPT-4o, test task quality, tool reliability, safety behavior, data controls, rate limits, and total cost on the exact route you intend to use.
Each score below comes from a published result we could trace back to its source. Read them as evidence about the specific task a benchmark measures — a model that leads on one can easily trail on another.
Related approved models in Text-to-text AI models. Compare specifications and verify fit for your workload.
Anthropic
The 2024 Sonnet that set the standard for practical coding work — now well behind the current Claude line at identical pricing.
Google's long-context multimodal model with up to 2M token windows.
Meta
Meta's largest open-weights model, competitive with frontier closed models.
Xai via Requesty
A frontier multimodal model optimized specifically for high-performance agentic tool calling.
DeepSeek
Official DeepSeek V4 Flash release with enhanced agentic capabilities and integrated DSpark speculative decoding