OpenAI GPT-5.4: 1 млн токенов и computer-use
Main chat
A chat for vibe coders: news, guides, live cases, marketplace, and finding executors.
On March 5, 2026, OpenAI officially unveiled the GPT-5.4, the most powerful and efficient frontier model to date. It is immediately available in ChatGPT (as GPT-5.4 Thinking and Pro), API and Codex. The main emphasis is on real professional use: complex code, agent workflow, working with documents, tables and full-fledged computer management.
In short: this is not just a “big model”, but a ready-made tool for developers that really speeds up work and reduces the number of errors.
What's new about GPT-5? 4
- Context of up to 1 million tokens is ideal for long-term projects, large codebase analysis and long-term agent tasks.
- Native computer-use – the model can manage the browser itself, click on coordinates, fill out forms, send letters and perform complex workflow (the demo with Playwright Interactive is impressive).
- Upfront planning + steerability - now ChatGPT builds a response plan first and can adjust it along the way. Plus, preambles and mid-response instructions work much better.
- ** Improved multimodality** – better recognition of documents, tables, presentations and visuals. ChatGPT for Excel add-in.
- Tool search and token efficiency – the model chooses the right tools and spends fewer tokens (faster and cheaper to use).
There are two versions:
- *GPT-5.4 Thinking - for normal tasks (replaces GPT-5.2 Thinking).
- GPT-5.4 Pro - maximum performance for the most complex cases (Pro/Enterprise only).
Benchmark: a direct comparison with previous models
OpenAI has published an official table comparing GPT-5.4 with GPT-5.3-Codex and GPT-5.2. The numbers speak for themselves – especially in agent tasks, computer-use, and professional work.
| Бенчмарк | GPT-5.4 | GPT-5.3-Codex | GPT-5.2 |
|---|---|---|---|
| GDPval (wins or ties) | 83.0% | 70.9% | 70.9% |
| SWE-Bench Pro (Public) | 57.7% | 56.8% | 55.6% |
| OSWorld-Verified | 75.0% | 74.0%* | 47.3% |
| Toolathlon | 54.6% | 51.9% | 46.3% |
| BrowseComp | 82.7% | 77.3% | 65.8% |
*Note: A new image resolution setting was used for GPT-5.3-Codex.
Additional GPT-5.4 record results:
- Spreadsheet modeling: 87.3% (was 68.4%)
- Presentations (as measured by people): 68.0% preference vs. GPT-5. 2
- WebArena-Verified: **67.3% **
- MMMU-Pro: **81.2% **
- ARC-AGI-1: 93.7% (was 86.2%)
Especially strongly increased performance in working with tables, presentations and agent tasks. For developers, that means fewer bugs in the code, faster prototyping, and more reliable autonomous agents.
Availability and prices
Now (gradual rollout from March 5):
- ChatGPT Plus / Team / Pro – GPT-5.4 Thinking (GPT-5.2 Thinking will remain for another 3 months).
- Pro and Enterprise - GPT-5.4 Pro.
- API:
gpt-5.4andgpt-5.4-pro(Codex supports 1M context experimentally).
Prices rose (as expected):
| Модель | Input (за 1M токенов) | Output (за 1M токенов) |
|---|---|---|
| gpt-5.2 | $1.75 | $14 |
| gpt-5.4 | $2.50 | $15 |
| gpt-5.2-pro | $21 | $168 |
| gpt-5.4-pro | $30 | $180 |
Batch/Flex is twice as cheap and Priority is twice as expensive.
What this means for Vibcoders and Developers
GPT-5.4 is no longer just a chat. This is a full-fledged colleague who:
- it launches the browser itself and does routine activities,
- it analyzes huge repositories,
- builds complex models in Excel and presentations,
- create reliable agents without constant supervision.
For those building products on the OpenAI API and Codex, this is a direct performance upgrade and a new level of automation.