DeepSeek’s V4‑Pro Discount Becomes Permanent; Anthropic Launches Claude Opus 4.8

The AI landscape this week is characterized by rapid technological iteration, a surge of capital, and growing ethical reflection.

DeepSeek (May 22) : Announces that the 75% temporary discount on its V4‑Pro model is now permanent. Input token price drops to $0.435 per million, output token price to $0.87 per million, and cache‑hit price to $0.003625 per million, making it one of the most price‑competitive frontier models. Source: DeepSeek API documentation .

Microsoft (May 22) : Releases the Fara 1.5 browser‑computer agent family (4B, 9B, 27B parameters). On the Online‑Mind2Web benchmark the model scores 72%, outperforming OpenAI’s Operator and Google’s Gemini 2.5 in web navigation and task automation. Source: MarkTechPost / AI Agents Directory .

OpenAI (May 22) : Files a confidential IPO registration statement (S‑1 draft) with the SEC. Goldman Sachs and Morgan Stanley act as lead underwriters, JPMorgan Chase also participates. Target listing window is September‑November 2026, with a potential valuation of up to $1 trillion. Source: CryptoBriefing .

Apple (May 23) : Registers the sub‑domain genai.apple.com ahead of WWDC 2026, two weeks away. Reports suggest WWDC will showcase a new Siri architecture built on Google Gemini and allow users to choose default providers such as ChatGPT, Gemini, and Claude. Source: AI Weekly (citing 9to5Mac) .

Pope Francis (May 25) : Issues the first papal encyclical on AI, titled Magnifica Humanitas (42,300 words), focusing on protecting human dignity in the AI era. The document discusses AI’s impact on human relationships, employment, power concentration, and autonomous weapons; Anthropic co‑founder Chris Olah speaks at the release. Source: Vatican Newsroom .

Andrej Karpathy (May 25) : Joins Anthropic’s pre‑training team. Former OpenAI founding member and ex‑Tesla AI senior director, Karpathy will lead a new team to accelerate pre‑training research using Claude itself, marking a major talent win for Anthropic. Source: TechCrunch .

xAI (May 25) : Launches an early‑access version of Grok Build, a terminal‑native AI coding‑agent tool. Features include an interactive terminal UI, Plan Mode for plan‑then‑execute workflows, parallel sub‑agent collaboration, and headless automation. Currently limited to SuperGrok and X Premium Plus subscribers. Source: TECHi .

Microsoft (May 26) : Releases the MAI‑Image‑2.5 image‑generation model. On the third‑party Arena benchmark it ranks third for text‑to‑image, showing significant improvements in text rendering reliability, commercial visual quality, and scene‑structure understanding over MAI‑Image‑2.0. Source: Microsoft AI official blog .

Microsoft (May 27) : Publishes the open‑source Agent Governance Toolkit, providing policy enforcement, zero‑trust identity management, and sandbox isolation for AI agents. The toolkit covers the OWASP Agentic Top 10 security standards and supports multiple programming languages. Source: AI Heartland / AI Agents Directory .

YouTube (May 27) : Announces a major update to AI video labeling. Starting May, YouTube will automatically detect and label videos with substantial realistic AI‑generated or modified content, moving away from reliance on creator disclosure. AI labels will be more prominent, with long‑form video tags displayed below the player and short‑form tags overlaid on the video. Source: YouTube official blog .

Anthropic (May 28) : Completes a Series H financing round of $65 billion, bringing post‑money valuation to $965 billion—surpassing OpenAI for the first time. Lead investors include Altimeter and Sequoia, with Amazon contributing $5 billion. The company expects to achieve profitability this quarter. Source: Wall Street Journal (via 华尔街见闻) .

SpaceX (May 28) : Clarifies that its Colossus AI data center is rented to Anthropic for only six months, denying a long‑term lease. Anthropic pays $1.25 billion per month for the compute cluster, originally scheduled through 2029. Source: Sina Finance / Global Market Report .

Anthropic (May 29) : Officially releases the flagship model Claude Opus 4.8, just 41 days after Opus 4.7. The new model shows significant gains in coding, reasoning, and agent tasks, reducing code‑defect omission rate to one‑quarter of the previous generation. It also introduces a 2.5× faster “Fast” mode that cuts cost by two‑thirds while keeping standard pricing unchanged. Source: 36 Kr .

DeepMind CEO Demis Hassabis (May 29) : In a Stanford talk, claims AI is entering a “species‑level transition” with a development speed roughly ten times that of the Industrial Revolution, warning that the next decade will have minimal fault tolerance and urging stronger international AI regulatory coordination. Source: Stanford Daily .

The collection of announcements underscores how pricing strategies, model releases, corporate financing, and ethical discourse are jointly reshaping the competitive dynamics of the generative AI market.