AI News Brief | 2026-05-26

AI Industry Overview

The AI landscape today is dominated by blockbuster funding and product launches. Anthropic is closing a massive $900 billion round while OpenAI has officially filed for an IPO. Google DeepMind’s AlphaProof Nexus solved decades-old math problems for a few hundred dollars, and researchers used Claude Code to discover novel AI scaling algorithms that humans likely wouldn’t have designed. Gemini 3.5 was formally released, bringing “frontier intelligence with action” to Google’s model lineup. DeepSeek made its 75% price cut permanent, setting output tokens at 34× below GPT‑5.5. Meanwhile, the “TrapDoor” supply‑chain attack exploited AI coding assistants by injecting malicious instructions into .cursorrules and CLAUDE.md files — the first coordinated weaponization of agentic tooling. OpenAI’s GPT‑5.6 (codename iris-alpha) was spotted in Codex logs, promising 150M token context window and a likely June launch. Apple is reportedly using a custom 1.2T‑parameter Google model to power the next Siri. On the hardware side, Huawei unveiled the “韬定律” (LogicFolding) to boost chip performance without new lithography. Pope Leo XIV published the encyclical “Magnifica Humanitas”, urging humanity to guard against unchecked AI influence.

---

ChatGPT

OpenAI is embedding ChatGPT directly into Microsoft PowerPoint via a new beta add-in, allowing users to generate slides and refine content without leaving the app. Separately, ChatGPT now supports flight and package tracking, a file library, and faster visual answers. A notable security move: all images generated by ChatGPT will carry invisible AI markers detectable by anyone, addressing disclosure concerns. PowerPoint add-in | Flight tracking & markers | Invisible markers

Claude & Claude Code

Claude’s long‑rumored “Mythos” model is progressing toward public release, with Anthropic focusing on safety and enterprise controls. Claude Security entered public beta, giving dev teams granular data access policies and audit logs. Claude Code v2.1.147 turns agent workflows into reusable infrastructure, while the Week 20 changelog (May 11–15) details better multi‑tool orchestration. In a groundbreaking research paper, Claude Code was allowed to autonomously discover novel scaling algorithms — a glimpse into AI‑driven AI research. Claude Mythos | Claude Security beta | Claude Code 2.1.147 | Scaling algorithms

Gemini

Gemini 3.5 is now live, described as “frontier intelligence with action” — combining reasoning with the ability to perform real‑world tasks via extensions. The Gemini app becomes more agentic, offering proactive, 24/7 help. Rumors indicate Apple is using a custom 1.2T‑parameter Google model (far larger than Gemini 3.5 Flash) to rebuild Siri, with simple queries running on‑device. Gemini 3.5 blog | App agentic | Apple rumor

DeepSeek

DeepSeek made permanent its 75% discount on the flagship V4‑Pro model, pricing output tokens at least 34× below GPT‑5.5. Analysts see this as a sustained price war that pressures Western competitors. The move builds on DeepSeek’s strategy of low‑cost, high‑performance models and has already caused API traffic to surge. 75% permanent cut | Price war | China Daily

Cursor

Cursor was named a Leader in the 2026 Gartner Magic Quadrant for Enterprise AI Coding Agents, a major validation for the startup. The team also launched Composer 2.5, a model optimized for long‑running AI coding tasks at a cheaper cost, and improved its automations feature for more reliable multi‑step workflows. Gartner Leader | Composer 2.5 | Automations

Devin

Cognition expanded Devin with several enterprise‑grade upgrades: support for Windows development, automated testing, and incident management. Devin Auto‑Triage now monitors production incidents and automatically fixes them. Additionally, Devin now supports Android app development, broadening its mobile capabilities. Windows & testing | Auto-Triage | Android

GitHub Copilot

GitHub Copilot now features auto‑model selection in VS Code, routing tasks to the best model based on the type of problem. A new semantic issue search capability lets developers find issues using natural language directly in Copilot Chat. The base model for Copilot Business and Enterprise has been upgraded to GPT‑5.3‑Codex, enhancing coding accuracy. Auto model | Semantic issue | GPT-5.3-Codex

Qwen

Alibaba Qwen released Qwen3.7‑Max, an agent‑first model for coding with an implicit cache feature that automatically speeds up repeated queries for free. In a striking demonstration, Qwen agents ran autonomously for 35 hours to optimize code for Alibaba’s own custom chip. Alibaba Cloud also unveiled an advanced agentic AI ecosystem for global customers. Qwen3.7‑Max | 35‑hour agent | Agent ecosystem

Kimi

Moonshot AI’s Kimi K2.6 is now being served by Cerebras at record speeds, achieving the fastest inference throughput for a trillion‑parameter model. Cerebras is bringing Kimi K2.6 inference to enterprise customers, promising low latency. The K2.6 tech blog highlights open‑source coding advances. Cerebras speeds | Enterprise inference | Tech blog

Kling

Kling AI continues its push into professional filmmaking. At Cannes, Kling launched a Filmmaker Initiative and partnered with creators to produce the series David’s House using AI‑generated scenes — claiming industry firsts like native 4K output and motion‑control preserving real actors’ emotions. Meanwhile, Kuaishou is evaluating a restructuring plan for Kling’s assets, potentially involving external financing. Cannes initiative | Filmmaker article | Restructuring

Perplexity

Perplexity is open‑sourcing Bumblebee, its internal AI model, likely to foster community‑driven improvements. The Comet AI browser for iOS received eight major upgrades, including better multimodal search and context retention. Perplexity continues to differentiate itself with real‑time, citation‑rich answers. Bumblebee open‑source | Comet iOS upgrades

Notion AI

Notion launched its Developer Platform (v3.5), effectively turning the workspace into a hub for AI agents. Third‑party developers can now build and deploy agents directly inside Notion, leveraging Notion’s database and document APIs. The release also benefits Claude users who can integrate notional workflows more seamlessly. Developer platform | TechCrunch | Claude integration

Midjourney

Midjourney V8.1 is out, offering HD image generation at 3× faster speeds. Weekly info drops highlight ongoing improvements to style consistency and image composition. The community continues to push creative boundaries with the latest version. V8.1 review | Info 15

Pika

Pika has raised $135 M in total funding. Version 2.2 introduces a Scene Director, sound design, and API updates, making AI video generation more controllable. The Pika MCP (Model Context Protocol) allows users to generate videos directly from Claude, bridging coding and video creation. Funding | Pika 2.2 | Pika MCP

Sora

OpenAI has officially discontinued Sora, its text‑to‑video model. The decision comes after months of limited public access and reports of high computational costs. Google quickly moved to fill the gap with upgraded video generation in Gemini and Veo. Meanwhile, Recursive invested $650 M in self‑improving AI, a bet that resonates with the Sora shutdown. Discontinuation | Rise and fall | Google fills gap

ElevenLabs

ElevenLabs is partnering with Splice to create next‑gen AI‑powered creative music tools. Spotify will use ElevenLabs for an AI audiobook creation tool, expanding voice synthesis into publishing. Developers can now build agents in ElevenLabs using Claude Code, enabling voice‑driven agentic workflows. Splice partnership | Spotify audiobook | Claude Code integration

Bolt.new

Bolt.new now runs Opus 4.6 (Anthropic’s latest model) for all users, offering better reasoning and coding quality. Bolt is also expanding into Microsoft Azure and M365 for enterprise AI development, making its rapid prototyping platform available inside corporate clouds. Opus 4.6 | Azure expansion

Flux

Flux (the AI electronics design platform) raised $37 M to simplify hardware development. Its Spring 2026 updates bring faster AI layout generation, better component sourcing, and smarter routing. The company aims to “take the hard out of hardware” with generative AI. Funding $37M | Spring updates

MetaSo AI

MetaSo Search V2 was released as an AI skill on Termo, providing contextual search capabilities. Reviews highlight its ability to re‑define search in the AI era, offering feature comparisons and pricing. Search V2 | Review | Redefining search

v0 by Vercel

v0 introduced permission modes (Auto, Ask, Full), giving users control over AI actions. A detailed blog post explains how Vercel made v0 an effective coding agent, emphasizing prompt engineering and feedback loops. The changelog also records ongoing refinements. Permission modes | How they made it | Changelog

Windsurf

Cognition (makers of Devin) has acquired the remaining Windsurf team and technology, deepening ties with Anthropic. The Windsurf Editor (formerly Codeium) continues to receive updates via its changelog, focusing on agent stability. CEO Jeff Wang appeared on the Neo Show to discuss the future of AI‑assisted development. Acquisition | Neo Show LIVE | Changelog

Aider

Aider pushed a security fix for urllib3 (CVE‑2026‑44431) and gained attention as developers compare it to Claude Code. An XDA article and a Medium post both highlight that for certain coding tasks, Aider’s open‑source, file‑based approach offers simpler and more reliable results than Anthropic’s tool. Security fix | XDA article | Medium comparison

ChatGLM (Zhipu)

Zhipu AI released GLM‑5.1, an MIT‑licensed open‑source model designed for long‑duration autonomous execution. It achieved a record 400 tokens/second inference speed via its high‑speed API. Long‑context performance reportedly surpasses Opus 4.6, making it a strong contender for agentic coding tasks. GLM‑5.1 launch | 400 tokens/s | June release details

Cline

Cline introduced the Cline SDK, an upgraded agent runtime. The CLI was rebuilt on top of this SDK, with versions 3.0.12 and 3.0.13 released, adding better error handling and performance improvements. The SDK aims to make agent development more modular. Cline SDK | CLI v3.0.13 | CLI v3.0.12

Cody (Sourcegraph)

CodeWords (not directly Cody, but related news) raised $9 M to let users turn words into proactive AI agents and automations. Meanwhile, Cody by Sourcegraph is reviewed as an AI that knows your whole codebase, leveraging Sourcegraph’s code intelligence for context‑aware assistance. CodeWords $9M | Cody review

Consensus

Consensus launched a Research Agent for efficient literature synthesis, part of its drive to become the “AI OS for Researchers”. The company also raised $30 M to reach the next 10 million researchers, underscoring the demand for AI‑powered academic search. Research Agent | $30M funding | Review

Continue

Continue released a “Quality control for your software factory” update, emphasizing its role in enforcing code quality standards. The platform also benefits from broader integration trends with Claude and other models. Continue quality control

CopilotKit

CopilotKit launched its Enterprise Intelligence Platform, providing a persistence layer for agentic applications. The company also published a blog explaining how it provides the complete frontend stack for AI agents, positioning itself as a key component in the modern agentic stack. Enterprise platform | Frontend stack | Agentic stack article

Doubao (ByteDance)

Doubao (ByteDance’s AI assistant) faces a lawsuit from a user over a failed 600 CNY refund promise. Meanwhile, its Pro model experienced a significant benchmark drop (main index down 18.4 points, code execution down 30.8 in one day), raising questions about model degradation. ByteDance is planning subscription plans for Doubao. Lawsuit | Benchmark drop | Subscriptions