Everything about GPT-5.4: pricing, context window, features (tool_search, Computer Use, Compaction), and the expensive pro variant for hard tasks.
Full content
## GPT-5.4 Overview
Released March 5, 2026. OpenAI's most advanced model for professional work.
### GPT-5.4 (Flagship)
- **Context window**: ~1.05M tokens
- **Max output**: 128K tokens
- **Pricing**: $2.50/M input, $15/M output
- **Supports**: structured outputs, distillation, code interpreter, hosted shell, skills
- **Input**: text + images
- Most factual and efficient model — uses fewer tokens, faster responses
### GPT-5.4-pro (Hard Tasks)
- **Reasoning levels**: medium / high / xhigh
- **Pricing**: $30/M input, $180/M output
- Same context window (1.05M) and 128K output
- **Does NOT support**: structured outputs, distillation, code interpreter, hosted shell
- Use for tasks requiring maximum accuracy
### Key Innovations
#### tool_search
New capability that efficiently searches across many tools and loads only relevant ones. Saves tokens and improves tool selection.
#### Computer Use (Built-in)
The model can work with software in a build → run → verify → fix loop.
#### Compaction (Built-in)
Preserves important context across long agent action sequences.
#### Mid-response Interruption
Users can stop the model mid-response and add instructions or change direction.
### Benchmarks
- SWE-Bench Pro: 57.7%
- OSWorld-Verified: 75.0%
- BrowseComp: 82.7%
- ARC-AGI-2: 74.0% (Pro: 83.3%)
### Pricing Notes
- Price multiplier for prompts over ~272K tokens
- ~10% surcharge for regional endpoints (data residency)
### Codex Tip: 1M Context
```
model = "gpt-5.4"
model_context_window = 1000000
model_auto_compact_token_limit = 900000
```
### Snapshots
- gpt-5.4-2026-03-05
- gpt-5.4-pro-2026-03-05