You might currently be experiencing a déjà vu, or perhaps you have already found yourself lost in the vastness of OpenAI’s model catalog; in any case, we completely understand. On March 3, just three days ago, Sam Altman’s company announced its new model, ChatGPT 5.3 Instant, and now it has surprised us yet again with the introduction of ChatGPT 5.4, a cutting-edge AI designed for professional use that incorporates the coding capabilities of GPT 5.3‑codex.
GPT 5.4 has already begun a phased rollout in its Thinking and Pro versions, both in ChatGPT as well as in the API and in Codex. GPT‑5.4 Thinking is now available to Plus, Team, and Pro plans, replacing GPT‑5.2 Thinking (which shall remain available until June 5, 2026, via the model selector). Meanwhile, GPT‑5.4 Pro is already accessible for the Pro and Enterprise plans. Users who are currently on the Free plan shall not yet be able to access the potential of this model.
While ChatGPT 5.3 Instant was developed as a solution to the excess “cringe” of its predecessor, ChatGPT 5.2 Instant—which could sometimes display an awkward tone, whether for being overly authoritative or for making unwarranted assumptions about the user’s intentions or emotions—ChatGPT 5.4 is being introduced as a model that incorporates the best of OpenAI’s latest advancements in reasoning, coding, and agent workflow orchestration.
Furthermore, the company has structured this release along two distinct lines to meet different needs. On one hand, GPT-5.4 Thinking stands out for its enhanced deep reasoning, web research, and context maintenance capabilities. One of its major innovations is the ability to now display a preliminary thought plan, allowing the user to “adjust the direction of the response mid-way as it is being generated and obtain a final output more closely tailored to their needs, without requiring additional turns,” according to what OpenAI explains.
On the other hand, for those who require maximum power for mission-critical tasks, GPT-5.4 Pro is now available. This “big brother” has been optimized for complex decision-making and large-scale data analysis, delivering superior performance in corporate environments where margin for error must be minimal.
GPT-5.4 incorporates the advanced coding capabilities of GPT-5.3-Codex. OpenAI has succeeded in merging the best of its code-specialized models with much smoother general reasoning. “The result is a model that handles real, complex tasks with precision, effectiveness, and efficiency, delivering exactly what you requested with less back-and-forth,” the company affirms.
Theme park simulation game developed with GPT 5.4 starting from a single instruction, utilizing Playwright Interactive for in-browser gameplay testing and image generation for the isometric asset set. Source: OpenAI
This model has also been deployed in Codex and the API, being “the first general-purpose model we have released with next-generation native computer usage capabilities, enabling agents to operate computers and carry out complex workflows across all applications. It supports up to 1 million context tokens, allowing agents to plan, execute, and verify long-term tasks.”
For those seeking speed, a /fast mode has been introduced in Codex, which enables token generation rates up to 1.5 times faster without sacrificing any degree of intelligence.
The GPT-5.4 model introduces improvements in interaction with external tools, enabling agents to operate in more complex ecosystems with greater reliability. Through agent tool invocation, the model has increased its accuracy in deciding when and how to execute actions during the reasoning process.
For developers using the API, the most significant new feature is “Tool Search”. Previously, it was necessary to include all tool definitions in the initial prompt, consuming thousands of tokens and driving up both costs and latency. Now, “with Tool Search, GPT‑5.4 receives a simplified list of available tools along with the Tool Search function. When the model needs to use a tool, it can consult its definition and add it to the conversation at that moment.”
One of the major challenges of generative AI is “hallucinations.” OpenAI has made this aspect a priority and has ensured that GPT 5.4 is its most factual model to date. According to the company, individual claims have a 33% lower probability of being false compared to GPT 5.2. More broadly, its responses show an 18% reduction in the likelihood of committing errors.
With regard to its model’s security, OpenAI has reinforced GPT-5.4 with safeguards such as an expanded cybersecurity stack, which “includes monitoring systems, reliable access controls, and asynchronous blocking for higher-risk requests for clients in zero data retention (ZDR) environments, along with continued investment in the broader security ecosystem.”
In addition, the company has achieved significant advances in Chain of Thought (CoT) monitoring. OpenAI has developed a new open-source metric to determine whether the model can “trick” us by concealing its reasoning. The results indicate that GPT-5.4 Thinking exhibits a low ability to hide its reasoning. This ensures greater transparency and allows security tools to audit how the model arrives at its conclusions.
Photo: OpenAI
Your email address will not be published. Required fields are marked *
Δ