What is GPT-5.4 and why does it matter?
GPT-5.4 is OpenAI’s latest flagship artificial intelligence model, released in March 2026. This neural network combines the best achievements in reasoning, programming, and agentic workflows. If you follow the development of AI, GPT-5.4 is the model that is redefining the standards of professional work with artificial intelligence.
In this article, we will break down the key capabilities of GPT-5.4, compare it with previous versions, and explain how to use it today.
- OpenAI’s best model for professional work in 2026
- 83% wins compared with professionals on the GDPval benchmark
- 75% successful tasks on the desktop (OSWorld-Verified) — above the human level (72.4%)
- Support for a context window of up to 1 million tokens
- 33% reduction in hallucinations compared with GPT-5.2
Key capabilities of GPT-5.4
Professional work at a new level
OpenAI developed GPT-5.4 with a focus on real professional tasks. On the benchmark GDPval, which tests agents across 44 professions in 9 key sectors of the US economy, GPT-5.4 outperformed or matched professionals in 83.0% of cases — versus 70.9% for GPT-5.2.
Particular progress was achieved in office document work:
- Spreadsheets: on an internal benchmark used by an investment bank analyst, GPT-5.4 scores 87.3% versus 68.4% for GPT-5.2
- Presentations: in 68% of cases, people preferred slides from GPT-5.4 — thanks to better aesthetics and visual variety
- Data accuracy: false claims appeared 33% less often, and full answers containing errors were 18% less often
Native computer control
GPT-5.4 is the first general-purpose OpenAI model with built-in computer use capabilities. This means the agent can independently control a computer: press buttons, enter text, navigate websites and applications.
The results speak for themselves:
- OSWorld-Verified (desktop navigation): 75.0% — above the human level (72.4%)
- WebArena-Verified (browser-based work): 67.3% successful tasks
- Online-Mind2Web: 92.8% successful tasks based on screenshots alone
For developers, this opens new horizons: creating agents capable of performing complex multi-step tasks in real software environments without constant human intervention.
Programming and Codex
GPT-5.4 inherited the strongest coding capabilities from GPT-5.3-Codex and added knowledge and skills in professional work to them. On the benchmark SWE-Bench Pro it scores 57.7% — surpassing GPT-5.3-Codex (56.8%) and GPT-5.2 (55.6%).
Codex now has a /fast mode, which delivers up to a 1.5× increase in token generation speed without loss of quality. An experimental skill has also been released: Playwright (Interactive), allowing Codex to visually debug web applications directly during development.
Improved tool use
One of GPT-5.4’s key innovations is tool search: instead of loading all tool definitions into context at once, the model requests them as needed. This reduces token consumption by 47% while maintaining the same accuracy.
On the benchmark BrowseComp, which measures the ability of AI agents to find hard-to-reach information on the internet, GPT-5.4 outperformed GPT-5.2 by 17 percentage points: 82.7% versus 65.8%. The GPT-5.4 Pro version sets a new record — 89.3%.
Controlled thinking in ChatGPT
In ChatGPT, GPT-5.4 Thinking now displays a preliminary reasoning plan before starting complex tasks. The user can adjust the direction in the middle of the process without starting over. This is especially useful for long, multi-step requests.
GPT-5.4 vs GPT-5.2: comparison across key benchmarks
| Benchmark | GPT-5.4 | GPT-5.2 | Gain |
|---|---|---|---|
| GDPval (professional work) | 83.0% | 70.9% | +12.1% |
| OSWorld-Verified (PC control) | 75.0% | 47.3% | +27.7% |
| BrowseComp (web search) | 82.7% | 65.8% | +16.9% |
| Toolathlon (tool use) | 54.6% | 45.7% | +8.9% |
| ARC-AGI-2 (abstract reasoning) | 73.3% | 52.9% | +20.4% |
Availability and pricing of GPT-5.4
GPT-5.4 is available today:
- ChatGPT: GPT-5.4 Thinking is available to Plus, Team, and Pro subscribers (replaces GPT-5.2 Thinking). GPT-5.2 Thinking will remain available for another three months in the Legacy Models section, after which it will be disabled on June 5, 2026.
- API: available as
gpt-5.4andgpt-5.4-pro - Codex: includes experimental support for a 1M-token context window
API pricing:
- gpt-5.4: $2.50 / M input tokens, $15 / M output tokens
- gpt-5.4-pro: $30 / M input, $180 / M output
- Batch and Flex — 2x cheaper than the standard price; Priority processing — 2x more expensive
Security and control
OpenAI classifies GPT-5.4 as a model with high cyber capabilities according to the Preparedness Framework. The company has implemented an enhanced cybersecurity stack: monitoring systems, access control, and asynchronous blocking of high-risk requests for clients with Zero Data Retention.
A major milestone was the study CoT controllability: it was established that GPT-5.4 Thinking is not capable of intentionally hiding its chain of reasoning, which makes chain-of-thought monitoring a reliable safety tool.
Frequently asked questions about GPT-5.4
How is GPT-5.4 different from GPT-5.2?
GPT-5.4 significantly outperforms GPT-5.2 across all key areas: professional work (+12%), computer control (+28%), web search (+17%). In addition, GPT-5.4 is OpenAI's first model with native computer use capabilities and support for a context window of up to 1 million tokens.
Can GPT-5.4 be used for free?
There is no free access to GPT-5.4. The model is available in ChatGPT for Plus ($20/month), Team, and Pro subscribers. In the API, pricing starts at $2.50 per million input tokens.
What is GPT-5.4 Pro and how is it different from regular GPT-5.4?
GPT-5.4 Pro is a version designed for maximum performance on the most difficult tasks. It is available in ChatGPT Pro and Enterprise plans, as well as via API. On a number of benchmarks, the Pro version shows better results: for example, BrowseComp 89.3% versus 82.7% for the standard version.
Does GPT-5.4 support Russian?
Yes, GPT-5.4 supports Russian. The model was trained on multilingual data and is capable of handling complex tasks in Russian, including document analysis, writing, and answering questions.
When will GPT-5.4 be available in Russia?
OpenAI is available in Russia through VPN or intermediary services that provide access to ChatGPT and the API. Official access from Russia remains restricted by the company's policy, but developers can use the API through third-party platforms.
Conclusion
GPT-5.4 is not just another version of a language model. It is a systemic leap in AI capabilities: from working with documents and code to computer control and web navigation. OpenAI has managed to bring together the best of recent developments in one model, making it more efficient, more accurate, and faster.
If you are a developer or a professional working with AI tools, GPT-5.4 is worth your attention right now. Start with the API (gpt-5.4) or try GPT-5.4 Thinking in ChatGPT — and see for yourself how far artificial intelligence has advanced.
Sources: OpenAI official announcement of GPT-5.4 · GDPval benchmark · GPT-5.4 System Card