Plus: a Chinese open model catches the frontier at a fraction of the price, Palantir's CEO on owning your AI edge, and the rule for what's safe to paste into AI. ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌
DeskTheory / archive →
THE THURSDAY 3 + 1 2026.07.16
DeskTheory

The week in AI for CEOs: three things that matter, and one workflow to run.

Members: no offer this week. You already have everything DeskTheory makes. Straight to the news.

In AI this week

An open AI model just caught up with Fable 5 and GPT-5.6

Kimi K3 launched today from Moonshot, an Alibaba-backed lab in Beijing. It's the largest open-weight model ever released, and on the benchmark that measures real work across 44 occupations it lands right behind Claude Fable 5 and GPT-5.6 Sol, ahead of everything else. Two terms worth knowing here. Open source means the whole recipe is public. Open weight means you get the finished model itself, a file you can download and run on machines you control, while the recipe stays private. K3 is open weight: the download arrives July 27, and you can try it free at kimi.com today. And it undercuts the frontier on price, at roughly a third of what Fable 5 will charge once it moves to paid credits, and roughly half of GPT-5.6 Sol. (VentureBeat)

What it means for you: You don't need to switch anything. What changed is the market you buy intelligence in: open models have pulled about level with the frontier, and the price of frontier-level AI just fell again. That only moves in one direction, so never let a price quote or a six-month-old experience talk you out of an AI project. One caution if you go play with it: the Kimi app is a consumer app from a Chinese lab, so treat it like any consumer AI under the rule at the bottom of this email. Nothing sensitive goes in.

Palantir's CEO thinks renting AI is handing over your edge

Earlier this month, Palantir CEO Alex Karp went on CNBC and said "something has gone completely wrong" with how AI is sold. His argument: companies pay by the token for models they will never own, while the labs learn from everything those companies do. What businesses actually want, he says, is "control over their compute, their models, their data stack and their alpha... They want to know they own the means of production." Alpha is investor-speak for your edge, whatever your company knows or does that competitors don't. He says other CEOs are privately livid about the same thing. (CNBC)

What it means for you: Karp sells the alternative, so discount the volume, but keep the frame. When your team's prompts, corrections, and playbooks live inside a vendor's chat history, your operating knowledge is compounding on someone else's books. You don't need Palantir or your own data center to fix that at your size. Keep the knowledge that makes AI useful (how you work, what good looks like, your playbooks) in plain files you own, and treat the model underneath as a rental you can swap. The first story is exactly why the swap keeps getting cheaper.

Fable 5 is included in your Claude plan until Sunday

Anthropic extended access to Claude Fable 5 for a second time: paid Claude plans (Pro, Max, Team, and premium Enterprise seats) can use it for up to half their weekly usage at no extra cost through Sunday, July 19, at 11:59pm Pacific. Nothing to toggle; it's already on, including in Claude Code. After Sunday it moves to paid credits on top of your subscription. (OfficeChai)

What it means for you: Fable 5 and GPT-5.6 are not the usual incremental bump. They're a step change in how much work you can hand over: whole tasks come back finished instead of half-drafted, and the models stay on track across hours of work instead of minutes. If your opinion of AI was formed on last year's models, or even this spring's, it's out of date. Before Sunday, give Fable 5 one real piece of work: a contract to pull apart, a plan to pressure-test, a hiring scorecard to build. Just make sure what you paste follows the rule below.

This week's workflow
What's safe to put into AI
WORKFLOW / THE +1

What's safe to put into AI

> OUTCOME

Early on, I pasted part of a real customer contract (a document under NDA) into my personal AI account to get a plain-English summary. Nothing came of it. But "nothing came of it" is luck, and luck is not a policy you can hand 30 people. Your team is pasting company data into AI right now, not because they're reckless, but because nobody gave them a rule. This workflow is the rule: green, yellow, red. Simple enough that a new hire gets it in 30 seconds, and your team keeps using AI without you holding your breath.

> SETUP

Nothing to install. Five minutes to learn the rule, plus an honest look at which AI plans your team actually uses, including the personal accounts nobody mentions in meetings. It all hinges on one fact most leaders don't know: consumer AI plans can train on your chats by default, and business plans contractually don't.

> WHAT YOU SHOULD DO NEXT

Read the rule, swap in the five red items for your business (customer lists, comp, anything under NDA), and hand it to your team this week. Exhale.

Full walkthrough _
Andrew
1:1 · OPERATOR TO OPERATOR $1,000 · limited spots

Want to work through any of this directly?

I do a small number of Operator to Operator sessions each week. 60 minutes, screen-share, you bring the bottleneck and we leave with a concrete plan you can run that week. $1,000, pay then book.

Reserve an hour with Andrew →
desktheory.com · archive · written by Andrew Lissimore
THE THURSDAY 3 + 1 · 2026.07.16