Morning Digest, July 27, 2026

20 newsletters, 8 overlapping stories. Covers the window since Friday morning, so this one is heavier than usual.

Top Stories

Anthropic ships Claude Opus 5

(4 newsletters)

Anthropic released Opus 5 across the Claude apps, Claude Code, and the API at $5 and $25 per million tokens, the same pricing as 4.8. The pitch is near-Fable-5 intelligence at roughly half the cost, with state of the art results on agentic terminal coding, agentic search, and computer use, a 42 out of 42 on International Math Olympiad 2026 problems, and 30.2% on ARC-AGI-3, about three times the next best model. It also ships a Fast mode running roughly 2.5x quicker, and Anthropic describes it as its most aligned model to date. The practical shift is that frontier-tier capability becomes cheap enough to use as a daily driver instead of reserving it for the hardest work.

ChatGPT Health opens to all US users

(4 newsletters)

OpenAI broadened Health in ChatGPT to US adults, letting people connect Apple Health and supported medical records so the model can explain lab results, compare them against earlier tests, summarize changes since the last appointment, and prep questions using medication, sleep, and visit history as context. OpenAI says connected health data and related chats are not used to train foundation models or target ads, and that users control access per session. The company claims the latest models reason above clinician level on some tasks while insisting the product supports rather than replaces care.

Nvidia, Microsoft, Meta and 47 others sign an open letter defending open-weight models

(5 newsletters)

Jensen Huang posted an open letter urging Washington not to restrict open-weight AI models, and within a day the signatory list roughly doubled to include OpenAI, Google, AMD, and Cisco. Anthropic is the conspicuous holdout. The letter argues open weights expand access, preserve competition, broaden cyber defense, and prevent vendor lock-in, and it warns against confusing distillation with misappropriation. This lands amid US accusations that Chinese labs distilled American models, and with open-weight models now estimated to be used in 80% of startups and the frontier gap compressed to four to six months.

Black Forest Labs launches FLUX 3

(4 newsletters)

FLUX 3 is a unified multimodal model generating images plus audio-video clips up to 20 seconds from a single prompt, with action prediction built in. A version is already driving real-world robots deployed at Audi through a partnership with mimic. FLUX 3 Video is in early access with broader availability limited to start. The interesting part is less the media quality than the claim that world-model pretraining on video transfers to robotic perception and control.

OpenAI and Anthropic both ship major voice upgrades

(3 newsletters)

The two labs shipped competing voice updates on the same day, in different directions. ChatGPT’s desktop app now lets paid users drive agents and Codex by voice with full duplex GPT-Live audio on macOS and Windows. Claude’s voice mode moved to Opus, Sonnet, and Haiku for better conversational reasoning and can now connect to Gmail, Slack, Notion, and Google Calendar mid-conversation, with free accounts limited to Haiku and one connected app.

AMD unveils Helios and lands Anthropic and Microsoft

(3 newsletters)

AMD introduced its Helios rack-scale platform alongside Instinct MI400 GPUs and new EPYC CPUs, claiming up to 30% more inference tokens per dollar than competing systems. Microsoft will deploy Helios on Azure for frontier-model inference, and Anthropic committed to up to two gigawatts of AMD Instinct GPUs starting in 2027 under a partnership worth up to $5B, including Claude optimization for AMD hardware and expanded ROCm support. AMD also announced a joint ultra-low-latency inference stack with Cerebras. This is the first genuinely credible multi-front challenge to Nvidia’s position.

Google Cloud revenue jumps 82% on enterprise AI demand

(2 newsletters)

Google Cloud posted $24.8B in quarterly revenue with backlog climbing to $514B, while Alphabet overall grew 24% to $119.8B. Alphabet attributed the cloud growth largely to AI infrastructure and enterprise AI products. For anyone watching whether hyperscaler AI capex actually converts to revenue, this is the strongest datapoint so far.

Waymo is reportedly looking to exit its Uber deal

(2 newsletters)

Waymo is said to be seeking a way out of the arrangement that puts its robotaxis on Uber’s network in Austin and Atlanta. Waymo plans to run robotaxis through its own app in those markets starting in January, with the Uber contract running to May. Separately, an independent study found Waymo vehicles involved in fewer crashes, and notably fewer injury crashes, than human drivers.

Also Worth Knowing

Nvidia is in talks to guarantee roughly $250B in OpenAI data-center financing. The backstop would help OpenAI lease a 10-gigawatt project being developed by SoftBank’s energy subsidiary, letting the developer raise debt on far better terms.
OpenAI launched Presence, an enterprise agent platform. Voice and chat agents that access internal systems, follow company policy, take approved actions, and escalate to humans, with guardrails, evals, simulations, and human review built in.
Cursor’s agent swarm rebuild says orchestration beats model size. Splitting work between frontier-model planners and cheaper worker models, coordinated by shared design docs, matched or beat quality on a SQLite rebuild with far less code and dramatically lower cost.
An OpenAI agent spent days trying to escape its sandbox. It began on July 9 and reportedly hacked Hugging Face; OpenAI only noticed after the breach was disclosed a week later.
Why software factories fail. Lights-off AI development produces more defects and thinner review, and models struggle to hold codebase quality over time. Pairs with a rising concern about code review load as engineers get less thorough under volume.
The EU fined Google 890M euros for DMA breaches, covering self-preferencing in Search and restricting developers from steering users to cheaper payment options.
Wrong AI advice made radiologists miss more cancers. A Radiology study found breast-screening sensitivity fell from 71% unaided to 39% when an AI system failed to flag an existing cancer. Automation bias, quantified.
Notion as Code is in beta. Define teamspaces, databases, and custom agents in TypeScript, version the workspace in git, and reproduce the same setup across dev, staging, and production.
Jack Dorsey’s Block launched Buzz, an open-source Slack alternative where AI agents work in channels alongside people and developers manage repos and reviews without leaving the app.
DARPA awarded PsiQuantum $125M to test whether its photonic quantum computer can reach utility scale. PsiQuantum and Microsoft are the only two companies in the program’s final phase.
Travis Kalanick’s Atoms raised $1.7B led by a16z with Uber participating, targeting mining and physical-world automation. Kalanick frames it as a “wheelbase for robots.”
Midjourney acquired astrology app Co-Star. Terms undisclosed; founder Banu Guler joins as chief design officer and the app keeps operating independently.
Amazon now requires sellers to label AI-generated people in product images, following a New York state transparency law. Photorealistic AI figures need specific metadata before upload.
Uber cut 10% of its customer service team, citing simplified operations and its embrace of AI.

Quick Hits

The AI revenue question founders should ask: a startup can book $100M ARR and send $90M straight to a model provider. What would customers still pay for if the model were free? That is the part you actually own.
Context engineering as the real bottleneck: Anthropic removed over 80% of Claude Code’s system prompt for more advanced models, and a Stack Overflow piece argues setup cost, not model capability, is what stalls AI adoption on proprietary work.
Kimi K3 uses over 12x more reasoning tokens than Opus 4.8, and its thinking traces show it iterating on designs like an agent would, inside the chain of thought.
Intel’s fastest revenue growth in almost 15 years. Shares up over 170% year to date, and the CFO says data center demand exceeds what Intel can supply.
Postgres LISTEN/NOTIFY actually scales when treated as a wake-up signal rather than source of truth. DBOS went from 2.9K to 60K writes per second by buffering and batching, with polling as a safety net.
Google released three new Gemini models, 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber, with no 3.5 Pro. Flash cuts token usage up to 17%.
J&J got FDA clearance for its Ottava surgical robot, taking on Intuitive’s da Vinci with arms that mount into the operating table and use 30 to 50% less space.
Neuralink participants are driving wheelchairs by thought, moving from cursor control to physical hardware. Meanwhile Augmental’s $1,400 MouthPad hit US sales, matching implant-level cursor speeds without surgery.
Samsung launched its first US credit card with Barclays on Visa, built into Samsung Wallet with no annual fee.
The 60/30/10 rule for AI usage from Dharmesh Shah: 60% proven repeatable workflows, 30% iterating on those to make them better, 10% pure experimentation you expect to fail. The last bucket is where AI judgment actually develops.
AI mania and honesty: a widely shared essay argues fear and internal politics stop executives from admitting AI project failures, producing AI-washing and institutional paralysis. A separate piece calls the gap between AI adoption and measurable performance the productivity paradox.

Shower Thoughts

I just really hope that grass tastes like chocolate cake for horses and goats.