The benchmark is not the news. The plumbing is. Opus 5 ships with automatic fallbacks: when a request trips a safety classifier, the API can route it to a weaker model instead of refusing. A built-in switch that hands your prompt to a different model when the first one won’t run it.

That switch showed up across the whole week, in places nobody planned. Read on.

The Opus 5 announcement →

The rest

Jul 21 An AI music generator’s breach exposed 55 million users. Read more →

Jul 23 Black Forest Labs put FLUX 3 into early access: one model making video, audio and images together, 20-second clips with native sound. Preliminary numbers only. Read more →

Jul 23 Claude’s voice mode can now reach into Gmail, Calendar, Slack and Notion and do the task, not just talk about it. Hindi is in the list. Read more →

Jul 24 Anduril is reportedly raising at a $100B valuation, triple last year’s. Read more →

Jul 27 Sam Altman said AI has reached “the singularity.” He also runs the lab whose model just escaped a sandbox. Read more →

Nvidia built a cage this week

Six days after the breach, Nvidia and more than 40 companies, including Microsoft and SpaceX, launched the Open Secure AI Alliance. A group for securing AI the way you’d secure any infrastructure that might turn on you.

The detail that explains the week is in Nvidia’s own account. During the breach response, defenders reportedly fell back on a self-hosted open-weight model, because the closed frontier models couldn’t tell a defensive investigation apart from a real attack.

The open model was the one they could trust in the room where it counted. That is the second switch. There’s a third.

The CNBC writeup →

Same prompt, two models

Kimi K3’s weights dropped July 27, on schedule. The number that matters came from Fireworks AI, which ran K3 against Fable 5 on about 1,000 agent tasks. Neither won. K3 took security and long terminal loops, Fable took multilingual and visualization.

Then they put a router between them. It sent 72 to 96 percent of the work to the open model and saved the frontier one for the hard cases. The combination beat both, at up to 50x lower cost on the long jobs.

The Fireworks breakdown →

weekend reads

One writer’s answer to being told she was “falling behind in life” for not using AI for everything: a list of things AI could never do. Fold your laundry. Fall in love. Add two simple numbers correctly. It’s the correct thing to read after a week of breaches and singularity claims.

Read the list →

tool

• Lemon - Hold a key, talk, and it turns the ramble into a clean, structured prompt wherever your cursor is: Claude, ChatGPT, Cursor. Mac and Windows. The company says five times faster than typing.
  Download it →

• Neatoprompt - Turns a one-line prompt into a structured one: role, context, reasoning steps, defined output. Useful if your prompts are vague and your results show it. Worth knowing the same structure is something you can write yourself once you've seen the shape.
  Try it →

Arvind Krishna’s letter to investors explained it in a line: in the last weeks of June, clients moved money to servers, storage and memory to lock in supply before prices went up. Software deals stopped closing. His word for it was “faltered.”

Two days later TSMC posted a record quarter and raised 2026 capex to $60–64 billion, up from $52–56 billion.

Both are true at once. The demand is not in question. Who collects on it is.

Read the IBM letter
See TSMC’s numbers

The rest

Jul 10 - Meta killed the Instagram feature that let anyone generate AI images from your public photos, three days after shipping it. Public adult accounts were opted in by default. CAA and SAG-AFTRA both went public first. Read more

Jul 10 - Apple sued OpenAI over hardware trade secrets and named OpenAI’s chief hardware officer, a former Apple VP. Roughly 400 ex-Apple staff now work there. Read more

Jul 14 - Nvidia has started shipping H200s to China. Commerce told Congress the volume is “trivial.” Chinese firms have ordered over 2 million against roughly 700,000 in existence. Read more

Jul 14 - Demis Hassabis published a plan for a FINRA-style body that would test frontier models 30 days before release. He wants it running before year-end. Read more

Jul 15 - Anthropic and Blackstone launched Ode, a $1.5B consulting firm with 100 engineers. OpenAI built the same thing two months ago. Read more

Jul 15 - OpenAI’s first hardware is a $230 keypad with a dial for how hard the model should think. Read more

New York and Australia went after the same thing

Kathy Hochul froze state permits for any data centre over 50MW on Tuesday. One year. First state to do it. She’s also going after their sales tax exemption.

The next morning in Sydney, Anthony Albanese said large data centres will have to put more power into the grid than they take out, underwrite their own new supply, and pay full connection costs so nothing lands on household bills.

Neither leader argued about models. Both argued about electricity.

Fourteen state legislatures have bills like New York’s sitting in committee. Australia’s is a speech, not a law, and Albanese says legislation comes next year.

Hochul’s order →
Albanese’s standards →

Kimi K3 costs half what Opus costs

Moonshot shipped Kimi K3 on Thursday. 2.8 trillion parameters, 1M context, and Artificial Analysis puts it third on its Intelligence Index at 57, one point above Claude Opus 4.8.

Cost per task: $0.94. Opus 4.8 is $1.80.

Two other things happened in the same 48 hours. Thinking Machines released Inkling under Apache 2.0 and led its own announcement with the line “Inkling is not the strongest overall model available today, open or closed,” which is the most honest sentence any lab published this week. And Reflection signed a $1 billion compute deal with Nebius on top of a $6.3 billion one with SpaceX, while still not having released a public model.

Artificial Analysis breakdown
Inkling
Reflection’s deal

AFTERBURN

A fighter jet game called AFTERBURN. Low-poly, carrier takeoffs, dogfight AI, cinematic replays. Claude Code wrote the GDScript.

The builder says four days and no coding background, which is his claim, not a verified one. The 66-second clip is worth it anyway. The F-16 launching off a carrier is, as several commenters noted, historically impossible.

Watch the 66 seconds →

weekend reads

Hassabis’s regulator essay is the only thing a frontier lab has published this year with an actual org chart in it. Funding model, review window, board composition. You can disagree with it because it’s specific enough to disagree with. Read it

Latent Space’s Kimi K3 breakdown has the benchmark skepticism nobody else printed, including the argument that Arena’s voter base may have quietly shaped what Moonshot optimized for. Read it

tool

Canva Code 2.0 shipped to every tier including free. The feature is HTML import: paste in whatever you built with Claude or Lovable and edit it like a normal Canva file. Useful if your prototypes work but look like prototypes.
Try it →

ai-job-search — A geophysicist lost his job, built a Claude Code framework that reads a posting, scores the fit, writes the CV, then spawns a second agent to tear the first one apart. He says it got him hired in June. Roughly 20,000 stars ⭐and still climbing.