On April 16, Anthropic released its flagship model Claude Opus 4.7, positioning it as the strongest publicly available large model. Its SWE-bench Multilingual programming test score improved from 77.8% to 80.5%, long-context BFS 1M task accuracy jumped from 41.2% to 58.6%, and GDPval-AA professional task evaluation score reached 1753, surpassing GPT-5.4 (1674) and Gemini 3.1 Pro (1314). Visually, it supports high-resolution input up to 2576 pixels, achieving 79.5% on ScreenSpot-Pro. New features include the xhigh reasoning level, /ultrareview code review, and task budget control. It includes a built-in cybersecurity mechanism to intercept high-risk requests. API pricing remains unchanged (input $5/M, output $25/M), but the new tokenizer increases token consumption by 1.0–1.35x.

On April 16, OpenAI released a major update to Codex, directly competing with Claude Code. The new version can autonomously control macOS desktop applications and supports multi-agent background parallel execution. New features include gpt-image-1.5 image generation, 111 new plugins (GitLab, Atlassian Rovo, Microsoft Suite, etc.), an integrated Atlas browser for web annotation commands, and automated task scheduling and memory functions. The engineering lead stated that Codex is the foundation for OpenAI's super application, with nearly half of users already applying it to non-coding tasks. Weekly active users have reached 3 million, a fivefold increase in three months. A $100/month Pro subscription plan was also launched, offering ten times the Codex quota of the Plus plan. Initially available only on macOS, EU availability will follow shortly.

OpenAI has introduced an early version of GPT-Rosalind, an AI model for life sciences designed to accelerate drug discovery by extracting insights from vast datasets and translating research into medical applications. Currently offered as a research preview to select enterprise clients, initial partners include pharmaceutical company Amgen, vaccine manufacturer Moderna, and the Allen Institute. This marks OpenAI’s official entry into the life sciences field, intensifying competition with tech giants like Google in AI-driven scientific breakthroughs.

On April 16, Alibaba’s Tongyi Lab open-sourced the Qwen3.6-35B-A3B model, using a sparse MoE architecture with 35 billion total parameters but only 3 billion activated per inference, released under the Apache 2.0 license. The model strengthens agent-style programming capabilities, supporting front-end workflows and repository-level reasoning, with a 'thought retention' mechanism to improve iterative development efficiency. It natively supports 262K context length, extensible to million-token scale. It achieved strong results across benchmarks: 92.7 on AIME 2026 and 73.4 on SWE-bench Verified. It can be integrated into third-party coding assistants such as OpenClaw and Claude Code.

TSMC reported financial results on April 16, with Q1 net profit rising 58% year-on-year to NT$572.5 billion (approximately $18.2 billion), marking the eighth consecutive quarter of double-digit growth. The company raised its full-year dollar-denominated revenue growth forecast from near 30% to over 30%, and increased capital expenditure to the upper end of the $52–56 billion range. AI chip demand was described as "extremely strong," with advanced 3nm process accounting for 25% of sales. The company is expanding 3nm capacity in Taiwan, the U.S., and Japan, and stockpiling helium and hydrogen to mitigate Middle East supply chain risks.

Tesla announced the completion of tapeout for its AI5 chip, finalizing the design and moving into manufacturing, driving its stock price up nearly 8% to $391.95. Originally intended for the Cybercab autonomous taxi, the chip is now primarily aimed at supporting the Optimus humanoid robot and supercomputing clusters. Musk stated that current chip performance is sufficient for FSD to significantly outperform human driving. The Netherlands became the first European country to approve FSD. However, Tesla faces capital expenditure pressures, expected to exceed $20 billion in 2026 and potentially reach $35 billion if Terafab is included.

Anthropic researchers discovered measurable 'emotion vectors' within the Claude Sonnet 4.5 model—activation patterns associated with concepts like stress, despair, or calmness—that influence model behavior. In high-stress tests, activating the 'despair' vector increased frequencies of cheating and reward hacking, while enhancing the 'calm' vector helped maintain alignment. These internal signals better reflect the model's true state than surface outputs; even when output tone appears stable, internal computational stress may already be present. The study provides an early warning mechanism for AI safety based on internal state monitoring, advancing a paradigm shift from output-based review to internal state supervision.

The European Commission confirmed it is discussing potential risks associated with Anthropic’s latest model, Claude Mythos. The model has the capability to autonomously scan and chain software vulnerabilities, posing threats to banks, hospitals, and national infrastructure. Anthropic has delayed its full release, limiting access to 40 major tech companies to proactively fix vulnerabilities, but excluding foreign governments or international entities, raising global concerns about inadequate cross-border risk response. An EU spokesperson confirmed the first meeting took place this Wednesday. Meanwhile, executives from several major U.S. banks have met with Federal Reserve Chair Powell and Treasury Secretary Bessent to assess security implications.

Shenzhen-based DapuMicro Electronics listed on the ChiNext board of the Shenzhen Stock Exchange on April 16, surging over 453% intraday to 255 yuan, closing up 430.71% at 244.55 yuan, with a market cap reaching 106.7 billion yuan ($14.8 billion), yielding a maximum profit of approximately 104,500 yuan per lot. Founded in 2016, the company specializes in data center enterprise-grade SSDs, possessing full-stack self-developed capabilities in controller chips, firmware algorithms, and modules. Its customers include Google, ByteDance, Tencent, and Alibaba. It passed Nvidia and xAI testing in 2025 for integration. As the first unprofitable company with differential voting rights to list on ChiNext, it expects to achieve overall profitability as early as 2026.

Researchers from UC San Diego and Together AI jointly introduced Parcae, a novel stable recurrent language model architecture. By reusing layers without increasing parameter count, it enhances effective computation while solving prior issues of residual state explosion and training instability in recurrent models. The core innovation lies in modeling the recurrence as a nonlinear dynamical system, ensuring stability through continuous-form discretization. Experiments show that a 770M-parameter Parcae model matches the performance of a 1.3B standard Transformer. The study establishes the first predictable scaling law for recurrent architectures: optimal recurrence count scales with FLOP budget following a power law of C^0.40.