On April 23, OpenAI released GPT-5.5, codenamed "Spud", the first fully retrained base model since GPT-4.5. The model excels in agent-style coding, knowledge work, and scientific research, achieving 82.7% on Terminal-Bench 2.0, 58.6% on SWE-Bench Pro, 84.9% on GDPval, and 78.7% on OSWorld-Verified. It matches GPT-5.4 in latency but consumes fewer tokens, priced at $5 per million input tokens and $30 for output—approximately double that of GPT-5.4. Safety risk is rated "high", and the API remains closed. ChatGPT and Codex with GPT-5.5 are now available to Plus, Pro, Business, and Enterprise users, alongside the release of GPT-5.5 Pro.

On April 23, Tencent released and open-sourced the HunYuan Hy3 preview language model, featuring a fast-slow thinking fused MoE architecture with 295B total parameters and 21B active parameters, supporting a 256K context window. It scored 74.4% on SWE-bench Verified (up from 53.0% in the previous version), increased BrowseComp from 28.7% to 67.1%, and averaged 88.4 points on Tsinghua's doctoral math qualifying exam. Training began in late January, and the model was completed in under three months. API pricing is set at $0.18 per million input tokens and $0.59 for output. It has been deployed across products including Yuanbao, QQ, and WeChat Reading, and listed on Tencent Cloud TokenHub. The rebuild was led by Tencent's Chief AI Scientist Yao Shunyu.

Anthropic's restricted cybersecurity AI model, Claude Mythos Preview, was accessed without authorization on its release day. A small group of users exploited predictable system naming patterns, prior Mercor data leak information, and third-party vendor accounts to infer access endpoints and gain inference privileges. Anthropic confirmed it is investigating and stated that model weights were not leaked. The breach did not rely on sophisticated techniques but rather foreseeable security oversights, discovered by journalists rather than the company—dubbed a "humiliation" by experts. Previously, Mythos had also been exposed due to unencrypted website data, compounding failures that severely damage Anthropic’s credibility as an AI safety leader.

SK Hynix reported a first-quarter 2026 operating profit of 37.61 trillion KRW (approximately $25.4 billion), a fivefold year-on-year increase and well above expectations, with sales nearing 52.58 trillion KRW. Growth was driven by surging demand for high-bandwidth memory (HBM) chips fueled by the AI boom and significant price increases in DRAM and NAND. The company plans to invest $8 billion in ASML's extreme ultraviolet lithography equipment to expand capacity. However, Samsung has taken the lead in next-generation HBM4 chips, planning investments exceeding 110 trillion KRW. SK Group Chairman warned of potential future losses due to technological shifts and highlighted energy supply as a key challenge for AI data centers.

Intel's Q1 2026 revenue reached $13.58 billion, up 7.2% year-on-year and surpassing the expected $12.42 billion; adjusted EPS came in at $0.29, far above the anticipated $0.01. Data center revenue rose 22% year-on-year to $5.1 billion, driven by rising demand for AI CPUs. Second-quarter revenue guidance of $13.8–14.8 billion also exceeded forecasts. Despite a widened net loss of $4.28 billion, market sentiment remains positive about Intel's turnaround. The company is advancing 18A process production and has partnered with several of Elon Musk’s companies to develop and manufacture high-performance chips at the Taichung Wafer Fab Park.

OpenAI has released Privacy Filter, an open-source bidirectional token classification model designed to detect and mask personally identifiable information (PII) in text. Built with a sparse MoE architecture, it has 1.5B total parameters but only 50M active parameters, supports a 128K context window, identifies eight types of PII, and replaces them with placeholders, achieving a 96% F1 score on the PII-Masking-300k benchmark. Released under Apache 2.0 license, it supports client-side data anonymization via WebGPU execution in browsers. This move opens up OpenAI’s internal training data sanitization pipeline, though experts note the tool is only auxiliary and cannot address deeper privacy risks like model re-identification.

Huaqin Technology completed its H-share listing on the Hong Kong Stock Exchange on April 23, establishing an "A+H" dual capital platform, with shares peaking 17% higher on the first day. It issued 58.5 million H-shares at HK$77.70 each, raising HK$4.55 billion. As one of the world’s largest electronics ODMs, it reported 2025 revenue of RMB 171.44 billion (up 56% year-on-year) and net profit of RMB 4.05 billion. Data center business revenue surpassed RMB 40 billion, with AI servers accounting for over 70%. Seventeen cornerstone investors, including Morgan Asset Management, UBS, and Taikang Life, subscribed approximately $290 million.

AI-powered business intelligence analytics platform Omni Analytics has raised $120 million in a Series C round, reaching a $1.5 billion valuation—more than doubling from last year. The round was led by ICONIQ, with participation from Alphabet’s GV and Redpoint Ventures. The platform enables users to automatically generate data visualizations through chat interfaces, uses semantic models to unify business terminology and resolve query ambiguity, and supports AI agents interacting via MCP servers. Revenue tripled over the past year, with clients including Perplexity AI and dbt Labs. Developers can embed custom chart components, and data definitions can be shared with external AI models.

Hugging Face has launched ML Intern, an open-source AI agent capable of autonomously executing full machine learning research workflows, including reading papers, extracting datasets, and running GPU training jobs. It outperformed Claude Code in scientific reasoning benchmarks, improving Qwen3-1.7B's score on GPQA from 10% to 32% after 12 rounds of supervised fine-tuning (Claude Code’s best: 22.99%). It also surpassed Codex by 60% on HealthBench. The agent supports intelligent loops of up to 300 iterations, with context management and sandboxed execution capabilities. Hugging Face offers early users a total of $1,000 in GPU and Anthropic credit allowances.

Surging demand for AI servers has led chipmakers to prioritize high-margin AI products, causing shortages in key components for general-purpose servers. Power management IC delivery cycles have extended to 35–40 weeks, and Baseboard Management Controllers to 21–26 weeks. Samsung’s plan to shut down an 8-inch wafer fab further tightens capacity. TrendForce has revised its 2026 server shipment growth forecast from 20% down to 13%, although AI server shipments are still projected to grow by about 28%. Cloud providers securing early orders are exacerbating the impact on enterprise users, while declining global 8-inch wafer capacity intensifies supply pressure on analog chips.