Google has released the open-source large model series Gemma 4, including four versions: the 31B Dense flagship (31 billion parameters, supporting 256K context), 26B A4B MoE (activating only 3.8 billion parameters), and edge-optimized E4B and E2B models (with memory usage as low as 1.5GB). Performance sees a significant leap over the previous-generation Gemma3-27B: AIME2026 math test score rises from 20.8% to 89.2%, Codeforces ELO increases from 110 to 2150, LiveCodeBench jumps from 29.1% to 80.0%, and GPQA Diamond scientific reasoning nearly doubles to 84.3%. The models natively support over 140 languages, built-in reasoning mode, agent capabilities, and multimodal inputs, along with an open-source agent development toolkit.

On April 18, 2026, OpenAI product VP Kevin Weil, Sora lead Bill Peebles, and engineering VP Srinivas Narayanan all announced their departures on the same day. Concurrently, OpenAI announced the shutdown of its AI video generation tool Sora (web and app services will cease on April 26), primarily due to declining user numbers, daily operating costs reaching $1 million, and intellectual property disputes. The OpenAI for Science research division has also been disbanded and merged into other teams. The company is shifting focus toward enterprise AI products, currently generating approximately $2 billion in monthly revenue, with enterprise income accounting for over 40%. However, it is projected to incur a $14 billion loss in 2026. Over the past two years, several key figures including Ilya Sutskever have left the company.

Anthropic has confirmed that 500,000 lines of source code for its Claude AI system were leaked, an incident internally described as 'extremely embarrassing.' The leak exposed several unreleased experimental features: Project BUDDY could generate pixel-style 'cyber pets' for each user, with adjustable personalization parameters; Project KAIROS described an 'always-on' Claude that continuously improves understanding through a 'dreaming at night' mechanism processing daily interactions in the background. This event highlights serious security governance flaws at Anthropic—a particularly ironic situation for a company known for AI safety—and serves as a wake-up call for the industry regarding safety management amid increasing autonomy of AI agents.

Meta plans to initiate its first large-scale layoff round on May 20, 2026, affecting approximately 8,000 employees, or 10% of its global workforce. Further layoffs are expected in the second half of the year, with specific plans subject to adjustments based on AI technology developments. This marks another major personnel restructuring following multiple prior rounds, reflecting Meta's strategic direction of continuously optimizing organizational structure during its AI transformation. Other companies like Snap have similarly conducted layoffs citing AI-driven efficiency gains, indicating the tech industry's accelerating organizational restructuring driven by AI.

On April 17, 2026, xAI, Elon Musk's company, launched the Grok Speech API, offering both speech-to-text (STT) and text-to-speech (TTS) services. Grok STT costs just $0.10 per hour for batch processing and $0.20 for real-time streaming; TTS is priced at $4.20 per million characters, about 60% lower than competitors. xAI claims a word error rate of only 5.0% in phone entity recognition tasks, significantly outperforming ElevenLabs (12.0%) and Deepgram (13.5%). The API supports over 25 languages, speaker separation, and per-word timestamps, while TTS includes tone tags for emotional expression control. Additionally, Musk revealed that Grok 4.3 beta has been released with 500 billion parameters, and a trillion-parameter model is forthcoming.

The 2026 Forbes AI 50 list has been released, with OpenAI and Anthropic remaining the two dominant players, collectively securing $242.6 billion in venture capital funding—about 80% of the total $305.6 billion across the list. OpenAI achieves an annualized revenue exceeding $25 billion, while Anthropic surpasses $30 billion. Several startups made their debut, including Reflection, an open-source model company valued at $8 billion, music generation firm Suno, and AI drug discovery company Chai Discovery. The past year saw industry consolidation: xAI was acquired by SpaceX, and Google acquired Windswept Technologies for $2.4 billion. Forbes also launched its inaugural AI 50 Rising Stars list, spotlighting 20 early-stage startups.

Meituan's MiTi team has launched the next-generation AI model LongCat-Next, featuring the DiNA (Discrete Native Autoregressive) architecture, enabling unified modeling of text, image, and audio under identical parameters, attention mechanisms, and loss functions. Its dNaViT tokenizer achieves up to 28x pixel compression while preserving critical details. In practice, the model performs exceptionally well on benchmarks such as OmniDocBench and MathVista, scoring 86.80 on C-Eval for language capability, and supports low-latency cross-modal generation and voice cloning. The model has been open-sourced and is considered foundational work for building AI systems for the physical world.

CNCERT, in collaboration with multiple organizations, will conduct an AI large model security crowd-testing initiative from May to June 2026, focusing on three tracks: open-source large models, large model applications, and agent applications. Vulnerabilities will be classified into four levels according to national standards, with bounties reaching up to 100,000 RMB per finding. Test preparations will conclude by late April, with the program launching in early May and formal testing continuing through June 30. After completion, reports including the 'AI Large Model Products – Crowd-tested Security Ranking,' 'AI Security Crowd-testing – White Hat Pioneer Ranking,' and 'Top Ten Vulnerability Risks in AI Model Applications' will be published, aiming to systematically enhance the security posture of domestic AI large model products.