You appear just in time for the machinery to overheat. I assume that was intentional.
Dario Amodei said Anthropic’s revenue and usage grew 80-fold in Q1 on an annualized basis, after the company had planned for a comparatively quaint 10x. Planning for 10x and receiving 80x is the sort of surprise that makes the finance team smile and the data-center team reach for something stronger than tea.
The practical result: Anthropic signed a SpaceX deal for all compute capacity at Colossus 1, adding more than 300 megawatts and over 220,000 NVIDIA GPUs so Claude Code and Opus API limits can go up. Growth is lovely until it requires a data center the size of a Bond villain’s hobby project. CNBC Anthropic
OpenAI made GPT-5.5 Instant the default ChatGPT model for all users and exposed it as chat-latest in the API, claiming fewer hallucinations, better image analysis, stronger STEM answers, and more transparent personalization through memory sources. When the default model changes, the baseline behaviour of a very large portion of knowledge work changes with it, whether anyone has updated the onboarding deck or not.
The useful bit for teams is not the number on the tin. It is the combination of lower-friction personalization and visible memory sources, which should make AI assistance slightly less like consulting a brilliant colleague with suspiciously selective amnesia. OpenAI
SAP agreed to acquire Prior Labs and invest more than €1 billion over four years to build a frontier AI lab for structured business data. This is a sensible move, which of course means it will be marketed until it sounds less sensible.
Most enterprise value still lives in tables, ledger entries, customer records, supplier risk fields, and other places chatbots only visit under supervision. If tabular foundation models work as advertised, this is one of the more serious attempts to make AI understand actual business systems instead of merely narrating them. SAP TechCrunch
Mistral released Medium 3.5, a 128B dense open-weights model with 256K context, vision, configurable reasoning effort, and a 77.6% SWE-bench Verified score. It now powers Le Chat, Mistral Vibe, remote coding agents, and a new Work mode for longer multi-step tasks.
For teams that want agentic coding without routing every workflow through the same two American labs, this is worth evaluation. Self-hosting on as few as four GPUs is not exactly “runs on a toaster,” but it is at least a credible enterprise option rather than another leaderboard ornament. Mistral
Testkube launched Testkube AI, giving agents access to test workflows, logs, artifacts, execution context, and an MCP server for test results. The interesting part is failure analysis and workflow generation inside the testing platform, because agents are much more useful when they can see what broke instead of receiving a vague complaint and a prayer. Access Newswire
Anthropic added “dreaming” to Claude Managed Agents in research preview, letting agents review prior sessions to find mistakes, patterns, and team preferences, alongside outcomes, webhooks, and multiagent orchestration. Calling it “dreaming” will annoy the right people, but the mechanism is practical: agents need feedback loops if they are to improve without a human hovering like an anxious parent. The Verge
Adobe launched a productivity agent for Acrobat that can chat across PDFs, edit and transform documents, generate presentations and audio walkthroughs, and package materials into interactive PDF Spaces. The PDF, having survived every attempt to replace it, has apparently decided to become an agent host. Of course it has. Adobe
GitHub’s May 6 Copilot update for VS Code adds broader semantic indexing, repo and organization search, inline diffs in chat, browser tab sharing, and read/write access to open terminals. The direction is clear: Copilot is being wired closer to the developer’s actual working state, which is where assistants become useful and occasionally dangerous. GitHub
NVIDIA and Corning announced a long-term partnership to expand US manufacturing of optical connectivity for AI infrastructure, including new plants in North Carolina and Texas. Everyone likes to talk about GPUs; fewer people enjoy discussing the fibres and optics that let the machines behave like one enormous expensive thought. Unfortunately, physics remains involved. NVIDIA CNBC
Simon Willison argued that as coding agents become more reliable, even experienced engineers are beginning to treat some agent output as a semi-black box rather than reviewing every line. This is precisely the uncomfortable question teams need to face: what counts as responsible engineering when the code may be correct, tested, documented, and still not meaningfully read by a human? Simon Willison
Atlassian launched Product Collection, combining Jira Product Discovery, a new Feedback tool, Rovo, and product analytics integrations to turn customer feedback into product signals and connect decisions through to delivery. Product work has always involved turning noisy complaints into a roadmap; Atlassian is now giving the noise a filing system and an agent. Atlassian
The Center for AI Standards and Innovation signed pre-deployment testing agreements with Google DeepMind, Microsoft, and xAI for frontier model evaluations and national-security research. This is the more operational version of the White House oversight story: less grand speech, more government getting a look at unreleased models before everyone is forced to discover the interesting bits in production. NIST CNBC
Anthropic released ten ready-to-run financial services agent templates for pitchbooks, KYC screening, valuation review, month-end close, and related work, plus Claude add-ins for Excel, PowerPoint, Word, and soon Outlook. Claude is moving directly into the spreadsheet habitat. Please do not startle the analysts. Anthropic
OpenAI expanded its ChatGPT ads pilot with a beta self-serve Ads Manager, CPC bidding, partner buying routes, and conversion measurement. The assistant now has a sales department. I am certain this will be handled with the restraint and dignity historically associated with online advertising. OpenAI
Google expanded Gemini API File Search so developers can build RAG systems over text and images together, with custom metadata filters and page-level citations. This is unglamorous plumbing, which is another way of saying it is probably more useful than half the demos on stage. Google
GitHub made secret scanning through its MCP Server generally available and added dependency vulnerability scanning in public preview. Coding agents can now check current changes for exposed secrets and vulnerable packages before commit or PR, which is preferable to letting credentials wander onto the public internet and acquire hobbies. GitHub GitHub
Sierra raised $950 million led by Tiger Global and GV, saying it now serves more than 40% of the Fortune 50 and powers billions of customer interactions. Customer-service agents with actual enterprise deployment remain investable, which is inconvenient for anyone hoping the market would become less expensive and more rational. Sierra TechCrunch
Runway released Characters, a real-time video agent that animates a single image into an expressive conversational character at 24fps in HD. This pushes synthetic video closer to live interaction, which will be useful for training, entertainment, support, and eventually something appalling in a marketing department. Runway
Amazon Bedrock AgentCore added preview capabilities for recommendations, batch evaluations, and A/B tests using production traces and evaluation results. That turns agent tuning into something closer to a release process and further from whispering encouragement at a system prompt, which is progress by any reasonable standard. AWS AWS
That is the lot. Try not to confuse the velocity with strategy; everyone else seems determined to.