AI News

2026-05-29 · FridayPublished daily at 08:00

AI Daily

Issue 38292026-05Friday
30major stories21sources12firsthand announcements4new modelsAbout 12 minutes to read
Today's Highlights
  1. 1
  2. 2
  3. 3
In This Issue
  1. 01Model Releases/Updates4 items
  2. 02Product Releases/Updates8 items
  3. 03Industry News8 items
  4. 04Paper research2 items
  5. 05Tips and perspectives8 items

01Model Releases/Updates

Anthropic:NewsroomFirsthand

Claude Opus 4.8 released: Comprehensive upgrades in coding, agent skills, and reasoning.

Anthropic has released Claude Opus 4.8, a next-generation model and upgrade to Opus 4.7. It has improved across benchmarks for coding, agent skills, reasoning, and practical knowledge work. Claude Opus 4.8 is available now at the same price as its predecessor. New features include user control over task effort, a “dynamic workflows” feature in Claude Code, and a 2.5x speed mode for Opus 4.8 priced at one-third of the previous rate. Early testers report more reliable judgment on agent tasks and more efficient tool use. The model scored 84% on Online-Mind2Web, outperforming Opus 4.7 and GPT-5.5. Its honesty and alignment have also improved, and its rate of missed coding errors has fallen by about 75%.

Original text
xAI:NewsFirsthand

Grok Build 0.1 on API

xAI's latest coding model, Grok Build 0.1, has entered public testing through the xAI API. Trained specifically for agentic coding tasks, it supports web development, debugging, and MCP, and is also the model powering Grok Build CLI. Its inference speed exceeds 100 tokens/second, with pricing at $1/m tokens for input and $2/m tokens for output. Beyond coding, it is also suited to general-purpose agent and tool-calling scenarios, and is available through OpenRouter and Vercel AI Gateway.

Original text
Google AI for Developers

Nano Banana Pro and Nano Banana 2 officially launch

🍌 Nano Banana Pro [gemini-3-pro-image] and Nano Banana 2 [gemini-3.1-flash-image] are now officially released and available for production use through the Gemini API. Check out these excellent community examples to see what the two models can actually do 🧵↓

Original text
SenseTime

SenseTime Upgrades Its Infographic Generation Model, Enhancing Several Core Capabilities

SenseTime introduced an upgraded version of its infographic generation model, SenseNova-U1-8B-MoT-Infographic. The 8B-parameter model has been optimized across four key areas: improved text accuracy and readability, with less repetition and inappropriate scaling; more consistent and sensible layouts, with more stable backgrounds; higher-quality charts and diagrams; and new support for rendering academic content. The post includes links to the model page on Hugging Face and a page demonstrating its capabilities.

Original text

02Product Releases/Updates

Claude:BlogFirsthand

Introducing dynamic workflows in Claude Code.

Claude Code introduces “Dynamic Workflows,” enabling Claude to handle complex tasks end to end. By dynamically writing scripts, it runs dozens to hundreds of subagents in parallel within a single session, then verifies the results before presenting them. It is suited to tasks requiring analysis from multiple angles, such as finding bugs across codebases and large-scale migrations (for example, porting Bun from Zig to Rust). The feature is now available in research preview for Claude Code CLI, desktop, the VS Code extension, and platforms including the API, Amazon Bedrock, and Vertex AI. It is available to Max and Team users, as well as users on enabled Enterprise plans.

Original text
Perplexity

Perplexity Computer Now Integrates with Microsoft Office

Perplexity Computer is now available in Microsoft Excel, Word, PowerPoint, and Outlook. Use Computer directly in the app sidebar to coordinate work, draft documents, build models, create presentations, and handle email. Available now: https://www.perplexity.ai/hub/products/integrations/microsoft

Original text
Mistral AI:NewsFirsthand

Search Toolkit released.

Mistral AI releases the public preview of Search Toolkit, a composable framework for building production-grade search pipelines for AI applications. It aims to address the excessive engineering time teams spend assembling search infrastructure from scattered data ingestion, retrieval, and evaluation tools. Search Toolkit brings these together in a single framework with shared interfaces, allowing teams to focus more on improving search quality. The open-source tool can be deployed in the cloud, on-premises, or at the edge, and supports retrieval scenarios such as enterprise search and RAG.

Original text
MiniMax

MiniMax M2.7 Free Agentic Coding Available for a Limited Time

Free agentic coding with MiniMax M2.7 on @OpenHandsDev? Yes, please! Available for a limited time ⚡👀

Original text

Replit Canvas: Agent design tool released.

Original text
Google Developers BlogFirsthand

Accelerate your integration workflow with the Google Pay & Wallet Developer MCP server

Google introduced the Google Pay & Wallet Developer MCP server, an open-standard tool designed to securely connect AI development assistants and IDEs to live APIs and account context. Developers can search official documentation, validate Wallet pass definitions, check integration status, and manage merchant accounts without leaving their development environment. The integration aims to reduce development friction by minimizing context switching and providing real-time, reliable AI support, accelerating development workflows.

Original text
TechCrunch:AI

Sesame, a conversational AI startup founded by Oculus's founders, releases its iOS app

Sesame, an AI startup founded by the founder of Oculus, released its iOS app, bringing conversational AI agents to the public. The app offers a more natural back-and-forth interaction experience and is designed differently from traditional chatbots, aiming to make users feel more like they are talking with a real person.

Original text

MuleRun launches on Alibaba Cloud Marketplace, offering round-the-clock AI labor

Original text

03Industry News

Anthropic:NewsroomFirsthand

Anthropic Raises $65 Billion in Series H Funding at a $965 Billion Valuation

Anthropic announces a $65 billion Series H round led by Altimeter Capital and others, valuing the company at $965 billion post-money. The company says enterprise deployment of its flagship Claude models continues to grow, with annualized revenue exceeding $47 billion. The funding will support AI safety and interpretability research, expand computing capacity to meet demand for Claude, and scale products and partnerships. Anthropic has recently expanded its computing capacity significantly and announced that Claude is available on all three major cloud platforms: AWS, Google Cloud, and Microsoft Azure.

Original text
X.PIN

DeepSeek plans to immediately apply for a STAR Market IPO after completing its financing.

Exclusive: DeepSeek plans to apply for an A-share IPO on the STAR Market immediately after completing its current financing round of about $50 billion (350 billion yuan). Source: A large fund manager participating in the round.

Original text

Sam Altman and Dario Amodei Have Both Walked Back Predictions That AI Would Trigger an Employment Crisis

Original text
Mistral AI:NewsFirsthand

AI Now Summit 2026

Mistral AI announced several developments at AI Now Summit 2026. Its industrial AI solution, “Mistral for Industrial Engineering,” combines physical models with engineering knowledge and is working with Airbus, BMW Group, and ASML to accelerate engineering design and optimization. Its AI agent product Vibe has been upgraded with enhanced reasoning and agent-task capabilities. The company also plans to bring a new 10 MW data center in Les Ulis online in the third quarter of 2026, focused on inference operations to strengthen computing capacity security.

Original text

Qwen3.7-Max tops OpenRouter's popular large-model rankings.

Original text
OpenRouter:AnnouncementsFirsthand

OpenRouter raises $113 million in Series B funding

OpenRouter, an AI model aggregation platform, announced that it has raised $113 million in a Series B round. The round was led by CapitalG, with participation from several firms including NVentures and ServiceNow Ventures. Existing investors Andreessen Horowitz and Menlo Ventures also took part.

Original text
Ars Technica:AI

Apple is working to fit Google's massive Gemini model into the iPhone to power the new Siri

Apple is trying to integrate a large Gemini model into the iPhone to support new Siri capabilities. Because of the model's large size, fully local processing may not be feasible, making a cloud component highly likely.

Original text
OpenAI: Website updatesFirsthand

OpenAI's Frontier Governance Framework

OpenAI released its “Frontier Governance Framework,” outlining how its AI safety, security, and risk management practices align with new regulatory requirements in the EU and California. The framework is intended to govern the development and deployment of its frontier models.

Original text

04Paper research

Rohan Paul

hexoai open-sources the SIA framework: AI agents achieve recursive self-improvement

hexoai has open-sourced the SIA (Self-Improving AI) framework. The framework demonstrates that AI agents can not only optimize their external workflows (harnesses), but also directly update their own model weights through task feedback, enabling autonomous improvements in domain knowledge and capabilities rather than relying solely on human-provided prompts or tool improvements. The paper reports that SIA improves performance by 56.6% on the LawBench benchmark, reduces runtime on GPU kernels by 91.9%, and improves performance by 502% over the baseline on single-cell RNA denoising tasks.

Original text
LMSYS BlogFirsthand

The SGLang team partners with AMD to make large-scale disaggregated inference for DeepSeek-R1 on AMD Instinct™ MI355X GPUs competitive in total cost of ownership

SGLang and AMD collaborated on a series of full-stack optimizations that deliver a highly competitive total cost of ownership on AMD Instinct™ MI355X GPUs running inference for the DeepSeek-R1 large model. At an interactive latency of 129 tok/s/user, the cost is $0.169 per million tokens, 5% lower than the NVIDIA B200 (Dynamo TRT-LLM) solution and 40% lower than the B200 (SGLang) solution. For throughput, 24 AMD GPUs achieved 2,436 tok/s/GPU, 1.25 times the per-GPU throughput of the B200 SGLang setup using 48 GPUs. Key optimizations include MoRI hybrid FP4/FP8 quantization all-to-all communication, MoRI-IO K…

Original text

05Tips and perspectives

IT Home · Artificial Intelligence

People's Daily interviews He Tingbo of Huawei: The new Kirin phone chip arriving this autumn will deliver a “leap” in performance over last year's model

He Tingbo of Huawei proposed a new semiconductor evolution path called the “Tao (τ) Law,” replacing “geometric scaling” with “time miniaturization” such as logic folding as the new guiding principle. She said Huawei has independently developed 381 chips based on this approach over the past 6 years. New Kirin smartphone chips will be released this fall; they are the first complete “Tao chips,” with “leap” improvements in performance and integration compared with last year.

Original text
IT Home · Artificial Intelligence

NVIDIA launches the AI framework Polar, boosting Codex benchmark scores by 594.74%

NVIDIA’s research team open-sourced Polar, an agent reinforcement learning framework. It connects to GRPO training by placing an agent at the model API boundary, without requiring rewrites of existing agent execution frameworks such as Codex CLI, Claude Code, Qwen Code, and Pi. Experiments show that with the Qwen3.5-4B model, Polar raised Codex’s pass@1 score on SWE-Bench Verified from 3.8% to 26.4% (a 594.74% increase). Its prefix_merging technique reduced training steps from 1185 to 218, making training about 5.39 times faster and raising average GPU utilization from 20.4% to 87.7%.

Original text
MarkTechPost

A coding guide to building semantic, hybrid, sparse, and quantized vector search systems powered by pgvector

This tutorial builds a complete pgvector experimental environment in Google Colab, demonstrating how PostgreSQL can serve modern AI applications as a vector database. It covers installing PostgreSQL, compiling the pgvector extension, establishing a connection through Psycopg, and registering the vector type for smooth integration with Python. Finally, SentenceTransformers is used to create and store embedding vectors.

Original text
OpenRouter

OpenRouter now offers Flex and Priority service tiers for supported models

Tip: You can use the Flex and Priority tiers for supported models (OpenAI, Google Vertex, etc.). See each model page for pricing information. Documentation: https://openrouter.ai/docs/guides/features/service-tiers

Original text

Four steps to secure AI-generated applications

Original text
Google Developers BlogFirsthand

How the community uses Tunix and TPUs to train Gemma to “think”.

At Google's Tunix hackathon on Kaggle, developers were challenged to turn small foundation models into general-purpose reasoning engines using TPUs and limited compute. The winning team achieved this through a multi-stage post-training pipeline combining supervised fine-tuning (SFT) with advanced alignment techniques such as GRPO and SimPO. The results show that the community can successfully train highly capable structured-reasoning models using open-source resources.

Original text
Tomasz Tunguz's Blog

Security Transformation in the Age of AI Agents

Jonathan Jaffe, CISO of Lemonade, discusses new security challenges in the age of AI agents. He notes that AI is equally powerful for attackers and defenders, but the window of exploitable vulnerabilities is shrinking because AI can generate, review, and patch code faster. In response, security teams are shifting toward engineering teams; for example, Lemonade's security department is made up entirely of engineers, and the company has built an internal AI platform that includes agents. In addition, every agent (with 200 to 10,000 potentially running on a single endpoint) needs an identity, and policies must provide more sophisticated controls at the point of operation, beyond the capabilities of today's identity and access management systems.

Original text
Cloudflare BlogFirsthand

How we built Cloudflare's data platform and the AI agents on it.

Cloudflare built a unified analytics platform called Town Lake and used it to develop Skipper, an internal AI agent.

Original text

Finished reading the news? Let’s talk about your business.

Leave your contact details, and I’ll get in touch with you personally.

WECHAT

Add me on WeChat

Press and hold or select the WeChat ID below to copy it.