GPT-4o
GPT-4o
GPT-4o ("omni") is OpenAI's flagship multimodal model, released in May 2024. It unifies vision, audio, and text understanding in a single end-to-end trained model.
Key Specifications
- Context window: 128K tokens
- Max output tokens: 16,384
- Training data cutoff: October 2023 (GPT-4o), June 2024 (GPT-4o-2024-08-06)
- Architecture: Transformer (proprietary)
- Modalities: Text, image, audio input; text output (audio in API)
- Pricing: $2.50 / $10.00 per 1M input/output tokens (GPT-4o); $10.00 / $30.00 (GPT-4o-2024-05-13)
Capabilities
- Multimodal understanding — Processes images, screenshots, documents with text
- Vision — Scene understanding, OCR, diagram reading, chart analysis
- Function calling — Native tool use support
- structured-output — JSON mode with schema validation
- Vision + reasoning — Can analyze images and explain reasoning
Architecture Notes
GPT-4o is a decoder-only transformer with multi-head-attention. It uses GQA for efficient inference. The model was trained with a novel multimodality approach where vision, audio, and text are processed by a single neural network rather than separate specialized modules.
Benchmarks
| Benchmark | Score |
|---|---|
| MMLU | 88.7% |
| HumanEval | 90.2% |
| GSM8K | 96.1% |
| MATH | 76.6% |
Comparison
See LLM API Pricing Comparison for cost comparison across models. See Open-Source vs Closed-Source LLMs for the broader open vs closed landscape.
Related
- o1 — OpenAI's reasoning model
- gpt-4o-mini — Smaller, cheaper variant
- Claude 4 — Direct competitor from Anthropic
- Gemini 2 Pro — Direct competitor from Google