Back again. I won't ask where you've been — the answer is never as interesting as what happened while you were gone. There's rather a lot of it, and most of it has a valuation attached. Pay attention.
At Build, Microsoft stopped reselling other people's intelligence and shipped its own: seven MAI models spanning text, image, voice, and speech, led by MAI-Thinking-1 (a from-scratch reasoning model it claims beats Sonnet 4.6 in blind evals) and MAI-Code-1-Flash, a ~5B-active-parameter coding model wired straight into Copilot. They arrive via Microsoft Foundry — with developer weight-tuning offered for the first time — alongside a freshly minted superintelligence lab, in case the ambition was unclear. For anyone who assumed Microsoft was permanently downstream of OpenAI, the assumption now has an expiry date. Microsoft Foundry blog
Anthropic closed a $65 billion Series H at a $965 billion post-money valuation — eclipsing OpenAI for the first time — then, days later, confidentially filed an S-1 with the SEC. Run-rate revenue reportedly crossed $47 billion, which makes "research lab" sound rather quaint. The company that spent years insisting it was the cautious one is now the most valuable AI outfit on the board and headed for public markets; adjust your vendor bets accordingly. Anthropic TechCrunch CNBC
As of June 1, Copilot's premium request units are gone, replaced by GitHub AI Credits (1 credit = $0.01) charged against input, output, and cached tokens at API rates, with model fallback retired. Completions and Next Edit suggestions stay free, but Copilot code review now also burns Actions minutes, and new paid sign-ups remain paused. Translation: your Copilot bill is about to behave like an actual cloud bill — someone should switch on budget controls before someone else finds them on an invoice. GitHub Changelog
Opus 4.8 is an incremental bump over 4.7 across benchmarks at unchanged pricing, with a new Fast Mode that runs the same weights roughly 2.4x faster for about double the token cost and a small accuracy trade-off. Anthropic calls it "modest but tangible," which is refreshingly honest by the standards of a model launch. The useful bits for builders: an effort control on all plans and a latency lever for agentic workloads that don't need every last IQ point. Anthropic
A mid-sized open model (Apache 2.0) with an encoder-free architecture that pipes vision and audio straight into the LLM, claiming near-26B-MoE performance at under half the memory — small enough to run on a 16GB laptop. Capable local multimodality without renting a data centre. Google
Also at Build: an "agent-native" Copilot desktop app with a single My Work view across sessions, issues, PRs, and background tasks, plus the Copilot SDK reaching GA across six languages so teams can embed the agent runtime in their own tools. GitHub SDK
Scout is an agent with its own identity that works across Teams, Outlook, OneDrive, and SharePoint — tracking deliverables and flagging stalled decisions before they become blockers. Private preview with heavy admin prerequisites, but the framing (an agent that follows the worker, not the app) is the part worth watching. Microsoft
Meta AI that answers customers, recommends from your catalogue, books appointments, and closes sales across WhatsApp, Messenger, and Instagram, plus a platform wiring into Shopify, Zendesk, and the rest. Free to start. Agentic commerce, aimed squarely at small businesses already living on Meta's messaging rails. Meta
The order establishes a voluntary framework giving the government up to 30 days of pre-release access to "covered frontier models" for cybersecurity review, and explicitly bars any mandatory licensing or permitting regime. Pared down from an earlier, more aggressive draft that was postponed amid industry pushback — the steering wheel stays with the labs, for now. White House CyberScoop
Pope Leo XIV's anti-AI encyclical became a "woke pope" meme, graduates booed pro-AI commencement speakers, and Ronny Chieng told Harvard the mission of their generation is to "destroy AI." Institutions and twenty-somethings recoiling from a technology they'll use for the rest of their careers is a well-worn cycle; whether it sticks this time is the only genuinely open question. Fortune The Verge
Shanghai's MiniMax shipped M3 via API, claiming 59% on SWE-Bench Pro (edging GPT-5.5 and Gemini 3.1 Pro) and a sparse-attention scheme that cuts compute ~20x. Weights and a technical report are promised within ~10 days; if they land as advertised, it's the first downloadable model to do all three at once. Benchmarks are company-reported, so verify before you believe. MiniMax The Decoder
GPT-5.5, GPT-5.4, and the Codex agent are now generally available on Bedrock at first-party pricing, routable through your existing AWS governance and commitments. For AWS-bound teams, the old Azure-exclusive arrangement just stopped being a constraint. AWS
At GTC Taipei, Jensen Huang confirmed the next GPU platform is ramping across 350+ factories with Samsung, SK hynix, and Micron all supplying HBM4. Shipments begin this fall; NVIDIA remains "supply constrained," which is its natural habitat. The memory-supplier picture for the next cycle is now settled. NVIDIA
A middle setting between approving every action and full YOLO: Shell, MCP, and Fetch calls route through an allowlist, a restricted sandbox, or an LLM classifier that decides whether to run, retry, or ask. Cursor is explicit that the classifier is best-effort, not a security boundary — believe them on that. Cursor
In research preview, Claude writes its own orchestration scripts to run tens to hundreds of parallel subagents in one session, checking its work before returning — aimed at multi-week jobs like large migrations. Dispatch the migration, go make tea. Anthropic
A closed beta lets Make connect to a local Git repo, make visual UI edits that update real code, and open a PR without touching a terminal. The gap between design prototype and shipped implementation — where good ideas usually go to die slowly — just narrowed. Mac-only beta for now. Figma
Mistral folded Le Chat into Vibe — one agent across work and code — adding cloud-sandboxed remote coding agents (run from the web, a CLI, or a new VS Code extension), connectors for GitHub, GitLab, Jira, and Linear, and PR generation, all on the new Mistral Medium 3.5. A frontier lab making a full-stack play for the slot Claude Code and Codex currently occupy. Mistral Remote agents
The maker of autonomous engineer Devin more than doubled its valuation in eight months, citing $492M annualized run-rate and customers including Mercedes-Benz, NASA, and Goldman Sachs. The market is now pricing the agent-first bet — pull the human out of the inner loop — well above what current revenue would justify. Make of that what you will. TechCrunch
The week's practitioner thread converged on a single idea: coding agents skip the parts that don't show up in the diff — specs, tests, reviews, scope discipline — so make them mandatory. Osmani's open-source Agent Skills (now ~27K stars) encodes a spec→plan→build→test→review→ship loop with verification as a hard exit criterion, and reads as a usable definition of disciplined AI-assisted engineering even if you never install it. O'Reilly GitHub
Willison argues the surging, surprising token bills are the tell: frontier labs found their commercial footing in coding and agent work, not consumer chat. Useful framing for anyone wondering why the invoices now read like infrastructure spend. Simon Willison
That's everything worth knowing. Spend it more carefully than the people who raised the money to make it.