The week did not behave itself. Sit down.
SpaceX reportedly obtained the right to buy Cursor later this year for $60 billion, or pay $10 billion for the two companies' collaboration. For context: Microsoft acquired GitHub for $7.5 billion. Cursor is a code editor that is two years old. It is now valued at eight times GitHub, by a rocket company, because apparently where your developers work is now considered strategic infrastructure. If you build software, you should probably be paying attention to who is trying to own the room it happens in. CNBC
Google will invest $10 billion immediately in Anthropic — at a $350–380 billion valuation — with up to $30 billion more tied to performance milestones, plus five gigawatts of compute and access to up to one million Ironwood TPU chips over five years. Total Google commitment to Anthropic is now approximately $43 billion. This arrived on the heels of Amazon committing up to $25 billion (with Anthropic pledging $100 billion in AWS spend in return), which means the two largest cloud providers have each bet the GDP of a mid-sized nation on the same AI lab within the same week. Regulatory scrutiny is inbound; Google structured the deal with a 15% ownership cap and no board seat, which is not subtle. TechCrunch CNBC
Microsoft and OpenAI have revised their partnership agreement. OpenAI can now distribute across any cloud provider, not just Azure. Microsoft's IP license becomes non-exclusive through 2032. Microsoft stops paying OpenAI a revenue share; OpenAI's payments to Microsoft continue through 2030. Within hours, Amazon CEO Andy Jassy confirmed OpenAI models would arrive on Amazon Bedrock within weeks. The structural subtext: the AGI clause — the contractual tripwire that would have voided Microsoft's IP rights upon AGI achievement — has been quietly removed. Revenue-sharing now runs through 2030 "independent of OpenAI's technology progress." The practical implication: if you were waiting for OpenAI to formally declare AGI, they have decided that is no longer a category they are required to fill in. OpenAI The Verge
OpenAI released GPT-5.5 (codenamed Spud), its first completely retrained base model since GPT-4.5. Strong gains on agentic coding (Terminal-Bench 2.0: 82.7%) and computer use (OSWorld-Verified: 78.7%), plus a verified mathematical proof OpenAI is clearly very pleased about. Available in ChatGPT and Codex for paid users at $5/$30 per million tokens; API access is delayed for additional safety review at scale. GitHub has already made it generally available in Copilot, with a 7.5x premium request multiplier — because intelligence, apparently, comes with surge pricing. OpenAI GitHub
GitHub announced that Copilot code review will begin drawing on GitHub Actions minutes starting June 1, 2026, moving from no metered cost to consumption against your existing Actions budget. Teams running automated PR reviews at any volume should audit expected usage before the billing change hits. GitHub
OpenAI discontinued the Sora video generation app. The economics were catastrophic: an estimated $15 million per day in compute costs against $2.1 million in total lifetime revenue; each 10-second clip cost roughly $1.30 to generate, versus $0.07–$0.17 for competitors. The Sora API continues through September. This is a useful data point about which AI product categories are viable at current infrastructure costs, and a reminder that "people are using it" is not the same as "the numbers work." OpenAI Help
Canadian enterprise AI company Cohere is acquiring Germany's Aleph Alpha at a combined valuation of approximately $20 billion, with Schwarz Group (owner of Lidl) committing €500 million as lead investor. The German government becomes an anchor customer; both countries' digital ministers attended the Berlin announcement. The pitch is explicit: a non-US AI alternative for European regulated industries and public sector organisations that would like fewer geopolitical complications in their procurement paperwork. Subject to regulatory and shareholder approval. TechCrunch
Cursor 3.2 introduces /multitask, which spins up async subagents to parallelize work rather than queuing it sequentially. The update also adds improved worktrees for isolated background tasks across branches, and multi-root workspaces — meaning a single agent session can coordinate changes across frontend, backend, and shared libraries simultaneously. Useful, assuming you have reached the stage where one agent is not enough trouble. Cursor
Claude Code's Week 17 release puts /ultrareview into public research preview: it dispatches a cloud fleet of bug-hunting agents against your branch or PR and delivers findings back to the CLI or Desktop automatically. Targeted at high-stakes merges — auth systems, data migrations, anything you would rather not explain in a post-incident review. The same update batch adds automatic session recaps when you return to a paused terminal, custom color themes, a redesigned claude.ai/code web interface, and /resume that is up to 67% faster on large sessions. Anthropic
Anthropic acknowledged that recent Claude Code performance complaints came from caching and quality regressions in their own stack, not the underlying model. Teams relying on coding agents should take the obvious lesson: model quality is only one failure mode. Product plumbing can quietly ruin the whole experience, and your vendor will not always volunteer that explanation quickly. Fortune
DeepSeek released fourth-generation MoE models: V4-Pro (1.6T total / 49B active parameters) and V4-Flash, both supporting 1-million-token context via a hybrid attention architecture that cuts per-token inference FLOPs by 73% and KV cache memory by 90% versus V3. Available as open weights on Hugging Face and via NVIDIA NIM endpoints. The efficiency gains matter most to teams running extended agent loops on open-weight infrastructure. HuggingFace
The Verge's editor-in-chief published a widely-circulated piece arguing that tech's "software brain" — seeing everything as automatable databases and loops — explains the gap between industry enthusiasm and public hostility toward AI. Two-thirds of Americans have used ChatGPT in the past month; over half think AI will do more harm than good. Gen Z anger at AI rose from 22% to 31% year-over-year. Picked up immediately by Daring Fireball, Simon Willison, and others as a sharper-than-usual diagnosis of where things stand culturally. The Verge
A private group reportedly accessed Anthropic's restricted Claude Mythos Preview through a third-party vendor environment with inadequate access controls. The lesson is painfully ordinary: if you are handling dangerous capabilities, "unguessable URL" is not a security model. Fortune
OpenAI launched Workspace Agents in ChatGPT for enterprise tiers — persistent, cloud-run agents powered by Codex that handle multi-step workflows across 60+ apps (Slack, Google Drive, Salesforce, Notion) on schedules, without a human in the loop. In the same week, OpenAI published Symphony, an open-source orchestration spec for assigning coding agents work directly from issue trackers rather than babysitting terminal sessions. The useful shift is not that agents exist; it is that OpenAI is building the management layer around them, which is where teams actually lose the plot. Workspace Agents Symphony
Google launched its eighth-generation TPU platform and confirmed OpenAI as a customer alongside Anthropic and Meta. The AI compute market is becoming less of an Nvidia-only theatre, and cloud capacity is increasingly strategic leverage rather than commodity plumbing. Worth tracking for teams making infrastructure decisions. Google
AWS added three features to Bedrock AgentCore: a managed agent harness (preview) that runs a full agent loop in isolated microVMs with three API calls and no orchestration code; an AgentCore CLI for prototype-to-production lifecycle via CDK; and pre-built skills for coding assistants. Kiro supports the skills today; Claude Code, Codex, and Cursor plugins arrive by end of April. No additional charge for any of it. AWS
OpenAI launched WebSocket mode for the Responses API, maintaining persistent connections and cached state across agent loops. If your agentic workflow currently burns latency rebuilding context on every turn, this is the sort of unglamorous API improvement that actually matters. OpenAI
Martin Fowler highlighted Thoughtworks Technology Radar Vol. 34, whose central theme is that AI-generated code is accelerating "cognitive debt" — widening the gap between developers and the systems they nominally own. The radar's four themes: retain engineering fundamentals (TDD, zero trust, DORA) as a counterweight to AI speed; secure permission-hungry agents; apply harness engineering to coding agents (guides, sensors, feedback flywheels); and be selective about single-maintainer tools flooding the ecosystem. Harness engineering is now the dominant organising concept for teams doing serious AI-assisted development. Fowler Radar
Google's Gemini API gained new Deep Research agents with collaborative planning, visualisation support, MCP server integration, and File Search via the Interactions API. In less glossy terms: hosted research agents are becoming easier to connect to internal systems without building the entire contraption yourself. Google
That covers it. A moderately consequential seven days. The rest is noise.