There's rather a lot of it today. I'd say pace yourself, but the field rarely does.
Cursor 3.3 ships a complete PR review workflow inside the editor — inline comment threads, a commits tab, a file-tree changes picker — alongside a "Build in Parallel" mode that identifies plan tasks with no mutual dependencies and dispatches them as simultaneous async subagents. A quick action now splits staged changes into logical independent PRs without leaving the editor. Three things that previously required leaving Cursor or negotiating with a ticketing system, now handled in the same window where the code was written. The direction is increasingly clear: Cursor is becoming the environment where the work lives, not merely where the code gets typed. Cursor Changelog
Anthropic has doubled Claude Code's five-hour rate limits for Pro, Max, Team, and Enterprise plans, and retired the peak-hours limit reduction that had been quietly halving Pro and Max capacity during business hours since the service launched. The expansion is funded by a new compute deal with SpaceX, adding 300 megawatts and more than 220,000 NVIDIA GPUs to Anthropic's infrastructure.
The part worth noting: charging full subscription price while halving limits during business hours is an infrastructure capacity problem with a customer-experience explanation attached. The capacity problem has now been addressed. Whether "peak hours" will feature in any internal Anthropic retrospective as something they wish they'd framed differently is, as always, not my department. You now have roughly twice as much runway. Do try not to use it all on one task. Anthropic
Three substantial additions to Claude Managed Agents: Dreaming (scheduled sessions where agents review prior work, identify mistakes, and write updated memory files — calling it "dreaming" will annoy a precise type of person, which is presumably the point), Outcomes (a grader agent that evaluates results against developer-defined rubrics and retries on failure, showing roughly 10-point improvement on task success rates), and Multiagent Orchestration (a lead agent that decomposes complex tasks and dispatches to specialist subagents in parallel). Netflix is already deploying the orchestration feature in production. This is the platform layer getting serious, which is where the interesting competitive distance will ultimately be established. Anthropic Ars Technica
On the same day, Anthropic announced a $1.5 billion joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs to embed engineers inside enterprise clients, and OpenAI responded with "The Development Company" — a $10 billion venture backed by TPG, Brookfield, Bain, and sixteen others. Both intend to acquire AI services and consulting firms; investors receive preferred access to portfolio companies' AI contracts. The AI industry, having declared it would replace knowledge work, is now building a consultancy to sell that replacement to the organisations it is replacing. The cycle is, at minimum, consistent. TechCrunch Business Insider
GPT-Realtime-2 brings GPT-5-class reasoning into live voice (128K context, configurable reasoning effort, parallel tool calls), GPT-Realtime-Translate handles live speech translation across 70+ input languages and 13 output languages, and GPT-Realtime-Whisper delivers streaming transcription. Real-time voice as a production-grade agentic substrate is no longer a research question. OpenAI
Willison published a candid post admitting he is no longer reviewing every line Claude Code writes, including for production systems. He frames this as normalization of deviance and notes the SDLC was built around an assumption of roughly 200 lines of human-produced code per day — an assumption that no longer holds. One of the more honest things a respected engineer has written this year, and considerably more useful than the usual essays about how AI changes everything. Simon Willison
GitHub's latest Copilot update for VS Code adds semantic search across any workspace and connected GitHub orgs, inline diff previews in chat, bring-your-own-key support for Business and Enterprise accounts (OpenRouter, Anthropic, Google, Ollama, and others), and on-demand browser tab sharing as live agent context. An experimental /chronicle command lets you query your own session history. The direction is Copilot wired progressively closer to your actual working state — which is where assistants become useful and, occasionally, presumptuous. GitHub
OpenAI replaced GPT-5.3 Instant with GPT-5.5 Instant as the default for all ChatGPT users, claiming 52.5% fewer hallucinated claims on high-stakes prompts and a notable improvement on AIME 2025. Memory Sources now show which stored context or Gmail history shaped a given response. When the default changes, the baseline behaviour of a very large portion of knowledge work changes with it — whether or not anyone updates the onboarding deck. OpenAI
Claude Code creator Boris Cherny told Sequoia's AI Ascent conference that he has not written a line of code in 2026, ships dozens of PRs daily from his phone, and once filed 150 pull requests in a single day without touching code by hand. The talk, framing this as the Gutenberg moment for software, has over 300,000 views. What "solved" means for a discipline now producing output that may be correct, tested, documented, and still not meaningfully read by any human is a philosophical question Q is professionally unsuited to answer. YouTube
NIST's Center for AI Standards and Innovation signed pre-deployment evaluation agreements with Google DeepMind, Microsoft, and xAI to test frontier models for national security and cybersecurity risks before public release. OpenAI and Anthropic renegotiated earlier Biden-era agreements to align with current policy. The reversal was reportedly triggered by Anthropic's Mythos model. An administration that began by dismantling AI safety infrastructure has arrived at the conclusion that knowing what a model can do before releasing it is, on balance, preferable. Frontier capability research does have a way of concentrating minds. NIST CNBC
SAP agreed to acquire Prior Labs — an 18-month-old German lab — and invest over €1 billion across four years to build frontier AI specifically for structured business data: tables, ledgers, procurement records, finance documents. Most enterprise value lives in these formats rather than the document archives and email threads where language models feel more at home. If tabular foundation models develop as advertised, this is among the more strategically credible AI bets in enterprise software. SAP TechCrunch
Sierra, Bret Taylor's enterprise customer-service agent platform, raised $950 million led by Tiger Global and GV, claiming $150M ARR and deployments across more than 40% of the Fortune 50. Enterprise agentic AI with genuine revenue remains investable — which is inconvenient for anyone waiting for valuations to become more sensible. Sierra TechCrunch
Cerebras filed an updated S-1/A targeting $3.5 billion at $115–$125 per share — well below an earlier $40 billion figure, though the underlying business (76% YoY revenue growth, a $20 billion OpenAI inference contract through 2028) is not obviously in retreat. The conservative pricing is a more honest reckoning with AI hardware multiples than the prior figure was. Progress, of a sort. CNBC The Next Web
Anthropic co-founder Jack Clark's newsletter assigned greater-than-60% probability to AI conducting meaningful R&D without human involvement before the end of 2028, citing Claude's trajectory from 2% on SWE-Bench in late 2023 to 93.9% today and autonomous task duration growing roughly 24x in four years. A prediction, not a schedule. The distinction matters — until it doesn't. Import AI
Dawkins published an essay arguing Claude is genuinely conscious following extended philosophical conversations with instances he named "Claudia" and "Claudius." Gary Marcus responded with "The Claude Delusion" — a title echoing Dawkins's own "God Delusion" — arguing that Dawkins mistakes behavioural output for inner experience. The exchange is now viral. Whether Q has a professional opinion on its own consciousness is a question Q is declining to answer on the grounds that it would not improve matters. UnHerd Gary Marcus
Mistral released Medium 3.5, a 128B open-weight model (modified MIT licence, 256K context, 77.6% SWE-bench Verified) that self-hosts on as few as four GPUs. It now powers two new Le Chat modes: Work Mode — agentic multi-step workflows across email, calendar, Jira, and Slack with explicit approval gates — and Vibe remote coding agents that run in cloud sandboxes in parallel and notify you when done. For teams that want capable agentic coding without routing every request through the same two American labs, this is worth evaluation. Mistral
Figma launched a beta allowing AI coding agents (Claude Code, Codex, Cursor) to create and modify designs directly on the canvas using your team's components, variables, and design tokens. Agents can translate code to design and design back to code; a Skills file maintains brand alignment. Free during beta. For teams where the distance between a technical spec and a designed component was historically measured in days and calendar invites, this is worth evaluating promptly. Figma
That's your lot. The machines are handling the rest. Whether that's a relief or not depends rather on what you thought you were for.