You're late. The news, however, was not. Sit down. There's rather a lot of it today and I haven't slept.
Following weeks of entirely predictable user backlash over a gradual decline in Claude Code's performance, Anthropic has finally admitted the issues were caused by their own engineering missteps. These include a caching optimization bug and a measurable quality drop in Claude Cowork. One might suggest testing these optimizations before pushing them to developers, but apparently that is too old-fashioned. Source
DeepSeek officially launched its V4 Preview, featuring two open-weight MoE models: V4-Pro and V4-Flash. Both models support a 1 million token context window, dual thinking/non-thinking modes, and offer highly cost-effective API pricing. Another formidable entry to remind the frontier labs that open weights are gaining on them quickly. Source
OpenAI has released GPT-5.5, which they inevitably describe as their smartest model yet for multi-step agentic tasks. It reportedly scores 82.7% on Terminal-Bench 2.0 while matching the latency of its predecessor, making it genuinely useful for long-horizon engineering rather than just parlor tricks. It is rolling out now to paying users, with API access to follow. Source
SpaceX has struck a deal with AI coding startup Cursor, obtaining the right to acquire the company for $60 billion later this year or pay $10 billion for ongoing collaboration. The arrangement preempts Cursor's planned $2 billion fundraise and grants them access to SpaceX's Colossus supercomputer. A modest sum for autocomplete, one supposes. Source
Canada's Cohere is effectively acquiring Germany's Aleph Alpha in a merger that values the combined entity at $20 billion. The newly formed company will focus on providing sovereign AI solutions to highly regulated and public sectors, backed by a $600 million commitment from Germany's Schwarz Group. Finally, an AI company targeting the most lucrative demographic of all: risk-averse bureaucrats. Source
A private Discord group gained access to Anthropic's Claude Mythos Preview—a cybersecurity model deemed 'too dangerous' for public release—by guessing the URL in a third-party vendor environment. One would think a company building zero-day exploitation capabilities might invest in basic access control, but here we are. Source
At Google Cloud Next, Alphabet unveiled its eighth-generation custom AI chips, claiming 2.7x better price-to-performance than prior generations. Notably, Google confirmed that OpenAI is now taking multi-gigawatt TPU capacity alongside Anthropic and Meta, marking a significant shift away from OpenAI's historical exclusivity with Nvidia GPUs. Source
OpenAI is forming a $10 billion joint venture called DeployCo with private equity firms including TPG, Bain Capital, and Advent International to accelerate enterprise AI adoption across their portfolio companies. OpenAI is committing up to $1.5 billion of its own capital and guaranteeing a 17.5% annual return, signaling an aggressive push to secure enterprise distribution channels. Source
OpenAI launched Workspace Agents for its enterprise tiers, allowing users to build Codex-powered AI agents that automate multi-step workflows across third-party apps like Slack and Salesforce. It is the inevitable successor to custom GPTs, assuming you trust an LLM with your CRM data. Source
AWS introduced a managed agent harness for Amazon Bedrock AgentCore, allowing developers to build and run agents in three API calls without orchestration code. It includes a new CLI for lifecycle management and pre-built skills for AWS best practices, because writing your own orchestration layer was apparently too tedious. Source
Alibaba's Qwen Team released Qwen3.6-27B, a dense 27-billion-parameter open-weight model optimized for agentic coding. It introduces a 'Thinking Preservation' mechanism and outperforms significantly larger MoE models on key developer and front-end workflows. Source
CodeRabbit has released CodeRabbit Agent for Slack, extending its AI code review engine into a unified agent for the entire SDLC. The agent allows teams to investigate issues, generate plans, and review code directly within Slack threads while retaining context across different workflow phases. Source
Grafana Labs launched AI Observability in public preview to help teams monitor, evaluate, and trace LLM-powered applications and agents in real time. Because if your agents are going to fail, you should at least have a dashboard to watch it happen. Source
Figma integrated OpenAI's latest ChatGPT Images 2.0 model into Figma Make across Design, Draw, and Slides, offering higher-quality visual generations and better face preservation. They also brought Make preview capabilities to their mobile app and introduced AI credit consumption controls for Enterprise billing groups. Source
The latest Thoughtworks Technology Radar focuses heavily on the shift toward agentic workflows, coining 'Harness Engineering' to describe the safeguards needed for 'permission-hungry' agents. They emphasize that foundational practices like TDD are becoming even more critical to manage the resulting complexity. A rare moment of sanity. Source
Google's Gemini API changelog adds new Deep Research model versions with collaborative planning, visualization support, MCP server integration, and File Search. This turns a research assistant into a more deployable enterprise workflow component connected to internal systems. If your team has been improvising brittle agent pipelines, Google just handed you a more structured baseline. Source
Cloudflare published concrete internal adoption numbers and architecture for AI-assisted engineering, including MCP, gateway routing, and enforcement layers. They report 93% R&D usage and substantial request volume, which is the sort of data most companies prefer to keep vague. Conveniently for them, the stack is also a catalogue of products they sell. Source
GitHub restructured Copilot Individual offerings into tiered plans and paused new sign-ups due to heavy compute demands from agentic workflows. Starting April 24, they will use Copilot interactions from Free, Pro, and Pro+ users to train AI models unless explicitly opted out. An entirely predictable maneuver. Source
AWS announced the general availability of DevOps Agent, an autonomous teammate for incident investigation and resolution. The release adds support for Azure and on-premises environments, custom skills, and PagerDuty integrations. Source
Claude Code released major updates including a redesigned desktop app for managing parallel agents and new CLI commands like /ultrareview for multi-agent PR reviews. You can also use /effort to tune model speed versus intelligence, a dial I wish came standard on most humans. Source
Simon Willison published a detailed breakdown of 'Agentic Engineering Patterns,' demonstrating how deceptively short prompts can accomplish complex tasks. He also released a git-timeline analysis tracking the evolution of Anthropic's system prompts between Claude Opus 4.6 and 4.7. Required reading, if you bother to read. Source
Matt Webb argued that products increasingly need agent-ready, headless interfaces because users will delegate work to personal AI systems instead of operating every GUI directly. The idea reframes UX and API design around machine users, not just human users. If your product cannot be operated by an agent, someone else's probably can. Source
Anthropic introduced Claude Design, a standalone AI tool powered by Claude Opus 4.7 that generates interactive visual prototypes, mockups, and pitch decks from conversational prompts. Currently in research preview, it marks Anthropic's aggressive expansion into the application layer, directly targeting workflows traditionally owned by established design platforms. Source
That's the lot. Do try to use this information responsibly.