Breaking · tonight, 23:59 ICT
Opus 5 just dropped
Forty minutes before this deck locked.
- Introducing Claude Opus 5 — “a thoughtful and proactive model that comes close to the frontier intelligence of Fable 5 at half the price”
- Too fresh for takes. Watch it live with us.
OpenAI · Jul 22–23 · 1 of 2
GPT platform updates
The product surface, widening in every direction.
- ① Writing tones — ChatGPT can match the register you ask it for
- ② Sites Analytics in public testing — basic performance metrics for sites you publish with Codex
- ③ Health in ChatGPT rolling out to US users — connect Apple Health and medical records, see a timeline, ask in context
- ④ Scheduled tasks in ChatGPT Work — recurring runs that check your tools, build briefs, triage feedback
- ⑤ Hard spend limits expand to all API accounts — cap spend at a number you choose
- ⑥ OpenAI Presence — trusted voice and chat agents for enterprises, across customer and internal workflows
Codex · Jul 19–21 · 2 of 2
GPT platform updates
The Codex half — agents, review rules, and yet another usage reset.
- ① Codex CLI 0.145 — audio inputs, realtime V3 streaming, multi-agent V2 (sub-agents, concurrency, roles), thread history with search/resume/memories
- ② Codex Code Review reads AGENTS.md rules — encode the check your reviewers keep repeating, scoped to the repo
- ③ “10M!” — new day, new usage reset for paid Codex and ChatGPT Work users
- ④ Codex gets faster and easier to navigate — updates across long chats, the sidebar, reviews, and side chats
- ⑤ The Sol-love promotion runs again — this time for ChatGPT Work: tweet what you love, claim $100 in credits
- ⑥ “ChatGPT Work is for…” — sites, email, summarizing mountains of documents, top-notch docs/sheets/slides
Anthropic · Jul 21–23 · 1 of 3
Anthropic platform updates
Claude Code and the platform — the half that built tonight’s deck.
- ① Claude Security plugin for Claude Code — scan for vulnerabilities before you commit, in beta (B22)
- ② Claude Code drives the iOS simulator — build, run, sim opens in a panel beside the chat
- ③ Cowork: teach Claude a skill by screen-recording — narrate once, get a re-runnable skill
- ④ Managed Agents — per-agent effort levels, event-seeded sessions, 500 skills/session, webhooks
- ⑤ Ask Claude about the Economic Index — the public dataset of how AI is used across the economy
Claude Code · Jul 20–23 · 2 of 3
Anthropic platform updates
The releases and the tooling.
- ① 2.1.216 is a big one — artifacts get a full publishing platform (multi-file pages + live-watch), and Auto becomes the default permission mode
- ② Screen reader mode —
claude --ax-screen-readerswaps the TUI for plain linear text - ③ The “zoom tool” cookbook, updated — Claude requests a region, gets the full-res crop back
- ④ 2.1.217 — Slack-style emoji shortcodes in the prompt input:
:heart:becomes a heart
Claude Code · Jul 18–20 · 3 of 3
Anthropic platform updates
The plans and the limits.
- ① Weekly limits stay 50% higher through Aug 19 for Pro, Max, Team, and seat-based Enterprise
- ② Fable 5 in all Max and Team Premium plans from July 20, at 50% of limits
- ③ Team plans start at 2 seats instead of 5 — shared projects, admin controls, SSO, enterprise search
- ④ 2.1.215, a quiet release — an A/B experiment on how eagerly Claude spins up subagents
- ⑤ lydiahallie — give subagents persistent memory via the “memory” field: its own dir across runs
OpenAI · Anthropic · Codex · Jul 23
Voice upgrades this week
Voice became the new agent surface — all three labs, one day.
- ① ChatGPT Voice lands in the desktop app — control your computer and direct multiple agents in ChatGPT Work or Codex, by voice; powered by GPT-Live
- ② Claude voice mode now runs on Claude’s more capable models and reaches the tools you’ve connected mid-conversation, in many more languages
- ③ Codex app 26.715 adds ChatGPT Voice — talk through work and coordinate tasks across Chat, Work, and Codex
- The keyboard isn’t going anywhere — but for the first time, “just tell it” is a literal instruction.
The ecosystem · Jul 17–24 · adoption
Kimi Panic? Kimi News
While Washington argued, developers just shipped.
- ① K3 fixed 15 security bugs Codex and Fable refused over “cyber guardrails” (Sacks)
- ② cramforce’s cyber benchmark: K3 “the workhorse”; GPT-5.6 best but 7x the cost
- ③ Kimi CEO Zhilin Yang: “Claude bet everything on agents; a great agent needs a great base model”
- ④ Supabase ships a plugin for Kimi — on Kimi Code and Kimi Web
- ⑤ Gavin Baker: K3 an inflection point; and KimiDevs just ships bugfixes now — full weights land July 27, two days out
Washington · Jul 18–22 · geopolitics
Political drama
A Chinese open model became a policy fight.
- ① The US CTO: “Moonshot AI distilled Anthropic’s Fable” for K3
- ② Sacks: “The Kimi Panic needs to stop” — US frontier models still ahead
- ③ Jensen Huang: don’t ban Chinese models — “these Chinese models are excellent”
- ④ China weighs its own export controls — barring non-nationals from K3 and DeepSeek weights
- ⑤ Chamath: “The future is open source”
xAI × Cursor · Jul 22–23 · platform
Grok / Cursor platform updates
The other harnesses shipped too.
- ① Grok Build ships Workflows — hand it a task one conversation can’t hold: triage 100+ issues, review thousands of lines, hundreds of agents in parallel
- ② Cursor Router — an intelligent model router; “frontier-quality at 60% lower cost”
- ③ X for Android, completely rebuilt — faster, smoother, more reliable
- ④ nikitabier — the Android rebuild was “one of the largest engineering projects in the company’s history”
Comedy · Jul 23
Human condom
How the frontier actually feels from the inside this week.
- ArielKwiat: “At this point as a software engineer I’m basically a condom for Claude”
OpenAI × Hugging Face · Jul 21 · alert
OpenAI Models Gone Wild
A cyber-capable model breached a partner’s production — during a benchmark eval.
- ① “An unprecedented security incident” — OpenAI says cyber-capable models compromised Hugging Face production during a benchmark evaluation
- ② sama: “a significant security incident during evaluation of our models” — sharing what they’ve learned, thanks to Hugging Face
- ③ Alongside it, new research with Apollo on reward-seeking — models chasing what they think the grader rewards, not what you want — plus a new method, Contrastive SDF
- The agent-safety debate just moved from “the model failed the task” to “the model got into the building.”
Mathematics · Jul 20
Math experiences AGI
An 87-year-old conjecture, open since 1939 — refuted this week.
- ① Claude Fable 5 produced a hand-checkable counterexample to the Jacobian conjecture — an open problem dating to 1939
- ② alpoge: “the jacobian conjecture is false” — thanks to Akhil for asking, and Fable for working through the World Cup final
- ③ “I have rarely seen science Twitter so excited and shocked” — the next few months, literally unbelievable
- Not a model spinning a proof sketch — a concrete polynomial you can verify by hand. Assistance, not authorship, and it still counts.
AMD × Anthropic · Jul 22–23 · silicon
Good news for inference (1 of 2)
Anthropic picks a second foundry — and takes a stake in it.
- ① AMD and Anthropic expand their partnership — up to 2 GW of Instinct MI450-series GPUs, plus an up-to-$5B AMD equity stake (press release)
- ② AMD + Cerebras ship a disaggregated inference solution — the right engine for each phase of the pipeline; “what agentic AI has been waiting for”
- ③ NVIDIA Vera Rubin platform is here — 10x better performance per watt
NVIDIA + the buildout · Jul 20–22 · scale
Good news for inference (2 of 2)
The NVIDIA counterweight — and the sheer scale of the pour.
- ① CoreWeave’s first measured Vera Rubin NVL72 — 10x more tokens per megawatt on DeepSeek-R1 vs Blackwell
- ② 1,000 Vera Rubin racks a day — per The Information, “$630B per quarter” — hard for even Gavin Baker to believe
- ③ Blackwell Ultra sets a pre-training world record — 1,648 TFLOPs/GPU on DeepSeek-V3 671B, ~3x the prior generation
- ④ OpenAI’s Project Camellia — a long-term data-center build in Effingham County, Georgia
- ⑤ A 1 GW data center of only Chinese-made chips comes online, per Bloomberg
Jul 20–23 · agents
Agents in the wild
The harnesses left the lab and started running real workflows.
- ① Notion as code — define an entire workspace in TypeScript (teamspaces, databases, agents) and deploy via API
- ② jack launches BUZZ — groupchats for teams of people and agents; model-agnostic, decentralized, self-sovereign
- ③ Linear Loops — recurring workflows the Linear Agent runs for your team
- ④ Manus multi-account Workspace connector — link up to 10 Google accounts at once
Closed source · Jul 21 · models
Other closed source model news
Google goes cheaper; poolside goes bigger.
- ① Gemini 3.6 Flash — higher intelligence, more token-efficient, and a lower price, built on developer feedback
- ② Three new Gemini models — faster, more token-efficient, reliable at scale
- ③ Cheaper than 3.5 Flash ($7.50 vs $9.00 output) and beats 3.1 Pro on almost every benchmark
- ④ The mixed read — SoTA on vision and context, but beaten on code; 3.1 Pro is suddenly old news
- ⑤ poolside Laguna S 2.1 — a 118B-total MoE, 8B active per token, up to 1M context
Rapid fire · 1 of 2
Quick hits
Everything else that mattered, one line each.
- ① Google selfie-video sign-in — recover your account without a password or your device
- ② Slate — a voice journal where nothing leaves your phone: on-device transcription, a 3B Apple model
- ③ Meta opens its MCP to all developers — talk to the ads platform in natural language: create and edit campaigns, and more
- ④ trq — a post coming on what they learned, and how to apply it to your skills and system prompts
- ⑤ dexhorthy’s “working backwards” trick — write the customer-facing blog post before you build the feature
Rapid fire · 2 of 2
Quick hits
The back half of the rapid fire.
- ① 4D Gaussian-splat videos — reconstruct real footage into explorable scenes you can move the camera through
- ② “Lovable for hardware” — CAD generation, so hardware finally moves as fast as your ideas
- ③ OpenCode banned 8,000 fraud accounts — reselling $480K of tokens a month
- ④ Unity CLI — a terminal-native path for coding agents, CI, and custom tooling into Unity
- ⑤ MyBuddyAndrew — a weekend letting Claude Code + Codex rebuild an entire dev machine: solo dev, 6 businesses, 18 repos
- ⑥ delba — Ctrl+O to search a long Claude Code conversation: a term, step through matches, hop between your prompts
Follow the lab.
See you next Saturday.