OpenAI recently launched its new GPT-5.4 models, including GPT-5.4 Thinking and GPT-5.4 Pro, making them available in ChatGPT, the API, and Codex. These models demonstrate superior performance in benchmarks for computer use, knowledge work, and coding, outperforming competitors such as Claude Opus 4.6 and Google's Gemini 3.1 Pro. Key features include native computer use capabilities, allowing them to operate desktop environments, and a 1 million token context window for complex tasks. OpenAI also highlights that GPT-5.4 is more token-efficient than its predecessor, GPT-5.2, potentially reducing real-world costs and increasing speeds for users.
The GPT-5.4 Thinking model in ChatGPT now shows its reasoning upfront and accepts mid-response adjustments, which enhances long-horizon tasks and web research. It also boasts improved visual understanding and is 18 percent less likely to contain factual errors compared to previous models. Despite these advancements, a New York Times test revealed that major AI chatbots, including Google's Gemini, OpenAI's ChatGPT, Anthropic's Claude, and xAI's Grok, struggled with federal income tax filing accuracy, miscalculating amounts by over $2,000 on average. This indicates that current AI models still find intricate tax law challenging, providing approximate answers rather than precise ones.
In the broader AI landscape, Together AI announced significant milestones at its first AI Native Conference, positioning itself as a key infrastructure provider. The company unveiled FlashAttention 4, offering up to 4x performance improvements for long-context workloads, and introduced ThunderAgent for efficient agentic workloads. Meanwhile, CrowdStrike warns that companies are deploying AI agents with excessive access to data and infrastructure, creating substantial security risks. CrowdStrike emphasizes treating AI agents like digital employees, applying least privilege and zero-trust principles, and continuously monitoring their behavior to counter threat actors exploiting these vulnerabilities.
AI's influence extends to geopolitical strategies, with Iran reportedly using state television and social media, including AI-generated content, to spread unproven claims and distort war reporting, mirroring tactics seen in Russia. Economically, a study of 5,000 eurozone firms suggests that businesses adopting AI are about 4 percent more likely to hire new staff, challenging concerns about job losses and indicating AI's potential for job creation. European telecommunication companies, like Deutsche Telekom, are urging governments to support sovereign AI investments, expressing concerns about fragmentation if companies opt for cheaper cloud providers such as Amazon.
Looking ahead to 2026, AI glasses are expected to offer "one-sentence task completion" for daily activities like ordering food, driven by growing demand, as evidenced by the Qwen app reaching 73 million daily active users during the Chinese New Year. However, mobile security in 2026 faces significant challenges, with new EU regulations opening iOS to alternative app marketplaces, introducing risks from unvetted applications. While AI accelerates development, it also introduces security flaws in generated code, widening the skill gap for security teams and underscoring the need for continued investment in mobile security expertise rather than relying solely on AI tooling.
Key Takeaways
- OpenAI launched GPT-5.4 models (Thinking and Pro) in ChatGPT and API, showing improved performance over Claude Opus 4.6 and Gemini 3.1 Pro in benchmarks.
- GPT-5.4 features native computer use, a 1 million token context window, improved visual understanding, and is 18 percent less prone to factual errors.
- Major AI chatbots, including Google's Gemini, OpenAI's ChatGPT, Anthropic's Claude, and xAI's Grok, miscalculated federal income taxes by over $2,000 on average in a New York Times test.
- Together AI introduced FlashAttention 4 for up to 4x performance improvements, a Reinforcement Learning API, ThunderAgent, and ATLAS-2 for enhanced AI infrastructure.
- CrowdStrike warns that AI agents with excessive data access pose significant security risks, urging companies to apply least privilege and zero-trust principles.
- Iran uses AI-generated content on state television and social media to spread unproven claims and distort war reporting, mirroring tactics used by Russia.
- European telecommunication companies, like Deutsche Telekom, are seeking government support for sovereign AI investments, expressing concerns about competition from cheaper cloud providers like Amazon.
- A study of 5,000 eurozone firms indicates that businesses adopting AI are about 4 percent more likely to hire new staff, suggesting AI can drive job creation.
- AI glasses are projected to offer "one-sentence task completion" for tasks like ordering food by March 2026, reflecting growing demand, with the Qwen app reaching 73 million daily active users.
- Mobile security in 2026 faces increased risks from new EU regulations opening iOS to alternative app marketplaces and AI-driven development introducing security flaws in generated code.
OpenAI launches GPT-5.4, outperforming rivals like Gemini
OpenAI has released its new GPT-5.4 models, including GPT-5.4 Thinking and GPT-5.4 Pro, available in ChatGPT and the API. These models show improved performance in benchmarks for computer use, knowledge work, and coding compared to competitors like Claude Opus 4.6 and Gemini 3.1 Pro. GPT-5.4 features native computer use capabilities, allowing it to operate desktop environments, and offers a 1 million token context window for complex tasks. OpenAI states the new models are more token-efficient, potentially reducing real-world costs for users despite a higher per-token price.
OpenAI unveils GPT-5.4 for professional tasks
OpenAI announced GPT-5.4, its most capable and efficient model for professional work, available in ChatGPT, the API, and Codex. GPT-5.4 Pro is also available for demanding tasks. The new model integrates advances in reasoning, coding, and agentic workflows, improving performance on tasks involving tools, software environments, spreadsheets, presentations, and documents. It features native computer use capabilities, a 1 million token context window, and improved tool search for more efficient agentic workflows. OpenAI claims GPT-5.4 is more token-efficient than GPT-5.2, leading to reduced usage and faster speeds.
OpenAI's GPT-5.4 enhances AI for complex tasks
OpenAI has launched GPT-5.4, including GPT-5.4 Thinking and GPT-5.4 Pro, aimed at improving agentic tasks and knowledge work. The GPT-5.4 Thinking model in ChatGPT can now show its reasoning upfront and accept mid-response adjustments, improving long-horizon tasks and web research. The API now supports a 1 million token context window, matching competitors. GPT-5.4 also boasts improved visual understanding and is claimed to be 18 percent less likely to contain factual errors than previous models. The company aims to improve capability, cost, and token efficiency to maintain user popularity.
Iran uses AI and social media to distort war reporting
Iran is using its state television and social media to project a defiant image during ongoing conflicts, despite significant strikes from adversaries. The country is waging an information war, blending facts with fiction and using AI-generated content to spread unproven claims. Iranian media reports exaggerated successes, such as missile strikes on Tel Aviv and the destruction of an American aircraft carrier, which differ from reality. This propaganda effort aims to project strength and distort the actual situation within the country, mirroring tactics used by Russia in Ukraine.
Telcos seek AI revenue from digital sovereignty at MWC
European telecommunication companies are urging governments to support their sovereign AI investments by becoming anchor tenants for AI services. Timotheus Höttges, CEO of Deutsche Telekom, warned at MWC that these investments might fail without government commitment, citing concerns about fragmentation if companies choose cheaper cloud providers like Amazon. He stressed the need for telcos to scale through mergers to invest in technology. Enterprises are increasingly building multiple AI stacks for different regions, creating complexity that telcos aim to manage with trusted platforms. SKT is also offering a 'Sovereign AI Package' combining infrastructure, a South Korean-focused AI model, and services.
CrowdStrike warns of AI agent security risks
CrowdStrike warns that companies are deploying AI agents with excessive access to data and infrastructure, creating significant security risks. Chris Stewart, CrowdStrike VP, stated that AI agents are becoming like digital employees and should be managed with the same security principles as humans, including least privilege and continuous oversight. He emphasized applying zero-trust principles to AI agents and monitoring their behavior. Stewart highlighted that threat actors are exploiting these security gaps rapidly. While AI poses risks, it is also a crucial tool for defenders, and CrowdStrike is partnering with VAST Data to embed its threat detection capabilities.
Together AI advances AI infrastructure and research
Together AI announced significant business and product milestones at its first AI Native Conference, positioning itself as a key infrastructure provider for AI-native companies. The company unveiled FlashAttention 4, offering up to 4x performance improvements for long-context workloads, and a new Reinforcement Learning API for distributed pipelines. They also introduced ThunderAgent for efficient agentic workloads and ATLAS-2 for faster inference using real-time data. These advancements aim to enhance training and inference performance, enabling businesses to leverage generative and agentic AI effectively.
AI glasses to offer 'one-sentence task completion' in 2026
AI glasses are expected to reach a significant milestone in 2026, with features enabling 'one-sentence task completion' for tasks like ordering food and booking rides to be available by March. This prediction comes as the Qwen app reached 73 million daily active users during the Chinese New Year. The development suggests a growing demand for AI-powered wearable technology that can seamlessly integrate into daily life.
AI chatbots struggle with tax filing accuracy
A New York Times test revealed that four major AI chatbots Google's Gemini, OpenAI's ChatGPT, Anthropic's Claude, and xAI's Grok struggle to accurately file federal income taxes. The chatbots miscalculated refunds or amounts owed by an average of over $2,000 for fictional tax scenarios. Experts note that the intricate details of tax law are challenging for current AI models, which tend to provide approximate answers rather than precise ones. While AI is improving rapidly, it is not yet reliable enough for tasks requiring absolute accuracy like tax preparation.
AI adoption boosts hiring in eurozone firms
A study of 5,000 firms in the eurozone suggests that businesses adopting artificial intelligence are more likely to hire new staff. Companies making significant use of AI were found to be about 4 percent more likely to take on additional employees. This finding challenges concerns that AI will lead to widespread job losses, indicating instead a potential for AI to drive job creation and economic growth.
2026 Mobile Security: Regulation and AI reshape risk
Mobile security in 2026 will be significantly reshaped by regulatory changes and the rise of AI-driven development, making mobile the largest yet least protected enterprise attack surface. New EU regulations opening iOS to alternative app marketplaces introduce risks like unvetted applications and malicious SDKs. Regulators globally are increasing explicit expectations for mobile app security, focusing on secure development and runtime protections. While AI accelerates development, it also introduces security flaws in generated code, widening the skill gap for security teams. Experts warn against mistaking AI tooling for a substitute for investment in mobile security expertise.
Sources
- OpenAI Releases GPT-5.4 Pro And GPT-5.4 Thinking, Beats Opus 4.6 And Gemini 3.1 Pro On Many Benchmarks
- Introducing GPT-5.4
- OpenAI introduces GPT-5.4 with more knowledge work capability
- Iranian TV and Social Media Project Defiant and Distorted View of the War
- MWC: Can telcos turn AI sovereignty into revenues?
- CrowdStrike: Companies give AI agents keys to the kingdom. That's a security disaster.
- Together AI Announces Business And Product Milestones At First AI Native Conference
- Qwen Hardware Head: "One-Sentence Task Completion" to Drive AI Glasses Demand
- A Word to the Wise: Don’t Trust A.I. to File Your Taxes
- An unlikely AI optimist
- 2026 Mobile Security: How Regulation and AI Are Reshaping Risk
Comments
Please log in to post a comment.