OpenAI ChatGPT 5.5 Release Analysis: What’s New

Lisa Ernst · 07.04.2026 · Artificial Intelligence · 10 min

The landscape of artificial intelligence evolves at a dizzying pace, and keeping up with the latest advancements can feel like a full-time job. When OpenAI announced its newest flagship large language model, GPT-5.4, on March 5, 2026, it marked another significant leap forward. This model, along with its smaller counterparts, promises to redefine how individuals and businesses interact with AI, particularly in professional and coding environments.

Source: A dynamically rendered graphic illustrates the comprehensive nature of the new OpenAI GPT-5.4, symbolizing its integration and advanced capabilities across AI applications.

Quick Summary: GPT-5.4 Highlights

• Release Date: March 5, 2026 (GPT-5.4), March 17, 2026 (GPT-5.4 mini and nano).
• Unified Model: Combines GPT and Codex strengths, offering native computer use.
• Expanded Context Window: Up to 1 million tokens experimentally (from 400,000 in GPT-5.2).
• Performance Boost: Significant improvements in reasoning, coding, and agentic workflows.
• Reduced Errors: 33% less likely to make incorrect singular statements than GPT-5.2.
• GPT-5.4 Thinking: New ChatGPT feature for outlining complex query approaches.
• Enhanced Safety: Classified with "High cyber capability" under OpenAI’s Preparedness Framework.
• New Variants: GPT-5.4 mini (faster, efficient) and GPT-5.4 nano (smallest, cost-effective).
• Availability: Rolling out to ChatGPT, Codex, and API users; GPT-5.2 Thinking retires June 5, 2026.

GPT-5.4: A Unified Frontier in AI

OpenAI officially launched GPT-5.4 on March 5, 2026. This iteration represents OpenAI’s most powerful and efficient frontier model for professional applications, merging advancements in reasoning, coding, and agentic workflows, as detailed in the official introduction to GPT-5.4. Notably, GPT-5.4 combines the strengths of the GPT and Codex lines into a single, unified system, also highlighted in the release announcement.

A key innovation in GPT-5.4 is its native computer use capabilities, making it the first general-purpose model with this feature, according to OpenAI’s communication. Agents can now operate computers and execute complex, cross-application workflows, as explained in the model’s overview. The model also features an experimental context window of up to 1 million tokens in Codex and the API, significantly expanding its capacity for processing information, which is a major point in the introduction. This represents a substantial increase from GPT-5.2’s 400,000 tokens, as contrasted in the performance comparisons.

GPT-5.4 improves tool usage efficiency through an enhanced tool search function, helping agents find and utilize the correct tools more effectively, as stated in the technical details. This reduces token consumption in tool-heavy workflows by 47% while maintaining accuracy, a significant improvement highlighted in the efficiency metrics.

Performance Benchmarks and Key Improvements

GPT-5.4 demonstrates notable performance gains across various benchmarks compared to its predecessor, GPT-5.2. On the GDPval benchmark, GPT-5.4 matches or surpasses industry experts in 83.0% of comparisons, a significant increase from GPT-5.2’s 70.9%, as detailed in the official benchmarks. An internal benchmark for spreadsheet modeling tasks in investment banking also showcases GPT-5.4’s superior analytical abilities, achieving an average of 87.3% compared to GPT-5.2’s 68.4%, further elaborated in the performance report. Human evaluators preferred GPT-5.4’s presentations in 68.0% of cases over GPT-5.2, according to the user feedback results.

The model significantly reduces factual errors; GPT-5.4 is 33% less likely to make incorrect singular statements and 18% less likely to contain errors in full responses than GPT-5.2, as indicated by the error rate analysis. For agentic environments, GPT-5.4 achieves a 75.0% success rate on OSWorld-Verified, surpassing GPT-5.2’s 47.3% and even human performance at 72.4%, a remarkable achievement presented in the agentic performance metrics. On WebArena-Verified, it reaches a 67.3% success rate with DOM and screenshot-guided interaction, an increase from GPT-5.2’s 65.4%, as shown in the WebArena results. Online Mind2Web shows an impressive 92.8% success rate with screenshot-based observations for GPT-5.4, according to the Mind2Web benchmarks.

Visual perception capabilities also see advancements, with GPT-5.4 achieving an 81.2% success rate on MMMU-Pro without tool usage, up from GPT-5.2’s 79.5%, as noted in the multimodal performance data. Document understanding also improves, demonstrated by an average error rate of 0.109 on OmniDocBench for GPT-5.4, down from GPT-5.2’s 0.140, as shown in the document processing benchmarks. For image detail, GPT-5.4 supports an "Original" setting for full fidelity up to 10.24 million pixels or a maximum dimension of 6000 pixels, information available in the image handling specifications. The "High" detail setting supports up to 2.56 million pixels or a maximum dimension of 2048 pixels, also detailed in the documentation.

In coding, GPT-5.4 integrates the programming strengths of GPT-5.3-Codex with knowledge work and computer usage capabilities, as explained in the coding section of the announcement. It matches or exceeds GPT-5.3-Codex on SWE-Bench Pro with lower latency, a key metric highlighted in the coding benchmarks. A /fast mode in Codex, powered by GPT-5.4, delivers up to 1.5 times faster token speed, information found in the Codex updates. The model excels at complex frontend tasks, producing more aesthetic and functional results, as described in the developer testimonials. An experimental Codex feature, "Playwright (Interactive)," facilitates visual debugging of web and Electron applications, as mentioned in the experimental features section.

The model also demonstrates superior agentic web searching, improving by 17% over GPT-5.2 on BrowseComp, with GPT-5.4 Pro reaching 89.3%, as shown in the web browsing performance report. The ARC-AGI-2 benchmark shows a significant jump from 52.9% for GPT-5.2 to an impressive 73.3% for GPT-5.4, detailed in the ARC-AGI-2 results.

Source: m.36kr.com

The ARC-AGI-2 benchmark results highlight GPT-5.4’s significant improvement, rising from 52.9% (GPT-5.2) to an impressive 73.3%.

GPT-5.4 Thinking and Enhanced Safety

GPT-5.4 introduces "Thinking" as a feature within ChatGPT, which drafts its approach to complex queries with an initial outline, as explained in the Thinking feature description. Users can adjust instructions or change direction mid-response, according to the user interaction guidelines. This allows the model to process difficult tasks more thoroughly and maintain context better during extended reasoning, a benefit highlighted in the reasoning capabilities section. GPT-5.4 Thinking enhances deep web research, particularly for highly specific inquiries, as demonstrated in case studies.

OpenAI classifies GPT-5.4 under its Preparedness Framework as having "High cyber capability," as outlined in the safety and security report. The model incorporates a robust cybersecurity stack, monitoring systems, and trusted access controls, also detailed in the technical specifications. An open-source evaluation indicates that GPT-5.4 Thinking has a low ability to deliberately conceal its chain of thought (CoT), a positive sign for safety, as reported in the safety research findings.

Availability and Pricing

GPT-5.4 is rolling out incrementally across ChatGPT and Codex, as announced in the availability update. In the API, GPT-5.4 is available as gpt-5.4, and GPT-5.4 Pro as gpt-5.4-pro, details found in the API documentation. ChatGPT Plus, Team, and Pro users have access to GPT-5.4 Thinking, which replaces GPT-5.2 Thinking, as mentioned in the ChatGPT-4o update. GPT-5.2 Thinking will remain available as a legacy model until June 5, 2026, as specified in the transition plan. Enterprise and educational plans can enable early access through admin settings, found on the GPT-5.4 introduction page, the ChatGPT Enterprise page, and the ChatGPT Education page. GPT-5.4 Pro is available for Pro and Enterprise plans, as highlighted in the Pro and Enterprise offerings and the Enterprise solutions.

Pricing for GPT-5.4 Thinking in ChatGPT remains consistent with GPT-5.2 Thinking, detailed on the ChatGPT pricing page. For the API, GPT-5.4 is more expensive per token than GPT-5.2, though its increased token efficiency can lead to lower overall costs for complex tasks, as explained in the API pricing breakdown. Batch and Flex pricing are half the standard API rate, with priority processing doubling the rate, further elaborated in the API pricing details.

API Pricing Overview

API prices for GPT-5.4 are $2.50 per 1 million input tokens and $15 per 1 million output tokens, as listed on the developer pricing page. GPT-5.4 Pro API pricing stands at $30 per 1 million input tokens and $180 per 1 million output tokens, also specified in the Pro API pricing.

GPT-5.4 Mini and Nano

OpenAI released GPT-5.4 mini and nano on March 17, 2026, as announced in the official introduction. GPT-5.4 mini significantly improves upon GPT-5 mini in coding, reasoning, multimodal understanding, and tool usage, according to the mini model’s capabilities. It operates over twice as fast as GPT-5 mini, a remarkable speed improvement noted in the performance comparisons. GPT-5.4 mini achieves performance comparable to GPT-5.4 on SWE-Bench Pro and OSWorld-Verified, demonstrating its efficiency, as detailed in the benchmarks for mini.

Source: appuals.com

This comparison chart illustrates how GPT-5.4 mini achieves efficiency comparable to the full GPT-5.4 model for tasks like SWE-Bench Pro.

This smaller model is available in the API, Codex, and ChatGPT, as described on the introduction page for GPT-5.4. It supports text and image inputs, tool use, function calls, web search, file search, and computer usage, showcasing its versatility, detailed in the features of GPT-5.4 mini. GPT-5.4 mini features a context window of 400,000 tokens, as mentioned in the technical specifications. API pricing for GPT-5.4 mini is $0.75 per 1 million input tokens and $4.50 per 1 million output tokens, available on the API pricing page. Within Codex, GPT-5.4 mini utilizes 30% of the GPT-5.4 quota, as specified in the Codex integration details. Free and Go ChatGPT users can access GPT-5.4 mini through the "Thinking" feature, as noted in the ChatGPT features and the ChatGPT overview.

Source: vecteezy.com

The iconic OpenAI ChatGPT logo—a teal circular badge with a white interwoven design—represents the platform where users can access the new GPT-5.4 mini.

GPT-5.4 nano is the smallest and most cost-effective version of GPT-5.4, exclusively available through the API, discussed in the nano model’s introduction. It is recommended for classification, data extraction, ranking, and coding sub-agents, highlighting its specialized uses, as described in the use cases for nano. API prices for GPT-5.4 nano are $0.20 per 1 million input tokens and $1.25 per 1 million output tokens, detailed on the API pricing for nano.

Conclusion

The release of GPT-5.4 and its variants, mini and nano, marks a substantial advancement in AI capabilities, particularly for complex professional tasks and coding. Its enhanced performance across benchmarks, improved safety features, and expanded context window signify a leap towards more robust and versatile AI agents. As these models become more accessible, their impact on software development, data analysis, and general workplace efficiency will likely continue to grow. OpenAI’s continued developments underscore a commitment to pushing the boundaries of what AI can achieve, ensuring constant evolution in the field.

Sources