In partnership with
The Big Sip
Image: Kimi
The take: Moonshot built a top AI model for $4.6 million. OpenAI just raised $6.6 billion.
What happened: Moonshot AI released Kimi K2 Thinking on November 6th. It scored 44.9% on Humanity's Last Exam and 60.2% on BrowseComp. That beats GPT-5 and Claude Sonnet 4.5.
Why it matters: Everyone keeps saying you need billions of dollars to build good AI. Moonshot just proved you don't. Their model costs $2.50 per million tokens. OpenAI charges $15 for the same thing. Companies are noticing.
What to watch: How many businesses dump their expensive AI contracts for the cheaper one that works better? Also, watch how fast the big companies start saying, "Well, actually, benchmarks don't tell the whole story."
[Analysis] VentureBeat, 6 Nov
K2 achieved 60.2% on BrowseComp (GPT-5 managed 54.9%, Claude 4.5 achieved 24.1%), 71.3% on SWE-Bench Verified, and 44.9% on Humanity's Last Exam.
Silicon Valley VCs call it “disruptive.” Your wallet calls it “about bloody time.”
Sponsor Break
Before we jump into today’s brew, here are some words from today’s sponsor…
Shoppers are adding to cart for the holidays
Over the next year, Roku predicts that 100% of the streaming audience will see ads. For growth marketers in 2026, CTV will remain an important “safe space” as AI creates widespread disruption in the search and social channels. Plus, easier access to self-serve CTV ad buying tools and targeting options will lead to a surge in locally-targeted streaming campaigns.
Read our guide to find out why growth marketers should make sure CTV is part of their 2026 media mix.
Here’s Your Brew
Moonshot spent $4.6 million training its new AI model, K2 Thinking.
It beats GPT-5 on web research. It matches Claude in coding. Then they posted the whole thing on Hugging Face for free.
CNBC's source leaked the cost. Probably wanted good press, just like DeepSeek did earlier this year.
Here's the smart part: K2 Thinking has a total of one trillion parameters, but only uses 32 billion at a time. It's like having a giant toolbox, but only grabbing what you need for each job.
Way faster and way cheaper than using everything at once. Then they compressed it to run twice as fast without compromising its integrity.
OpenAI spends hundreds of millions per model. So does Anthropic. Both keep telling investors you need billions to win at AI.
But companies are ditching them for cheaper Chinese models.
Airbnb's CEO chose Alibaba's Qwen.
Why?
"Fast and cheap."
He said this out loud about a free model. That competes with his friend Sam Altman's paid product.
Hard to defend your expensive moat when someone beats your test scores and gives their model away for free.
Two Sides, One Mug
Image: Kimi in action
Pro: Open and affordable beats closed and expensive.
Con: You’re handing your data to a state-linked firm that doesn’t answer to GDPR. Good luck with that.
Our read: The price gap is too wide for due diligence. Enterprises will adopt first and audit the politics later.
Receipt of the Day
[Report] Kimi K2 Thinking on Hugging Face, 6 Nov
Full model release: 1T parameters, 32B active, 256K context, INT4 quantisation.
Modified MIT licence with one catch: hit 100M monthly users or $20M revenue, and you've got to display "Kimi K2" in your UI.
That's the restriction. Everything else is standard MIT.
Spit Take
"$4.6M training cost vs OpenAI’s $2B+ burn rate" — CNBC Reporting, 11 Nov 2025
Your Coffee Break Links (and water cooler chatter)
[Analysis] Artificial Analysis telecom test, Nov — Why: Independent lab measured 93% on customer service benchmark. Highest score they've verified. Not Moonshot's internal numbers.
[Report] Moonshot GitHub, Nov — Why: Tech docs show it handles 200-300 sequential tool calls without losing the thread. Most models drift after 30.
[Report] Fortune on Airbnb's choice, 21 Oct — Why: Chesky explains why he picked Qwen over his friend's ChatGPT. Includes the phrase "wasn't quite ready," which is PR-speak for "didn't work."
Join your team of caffeinated skeptics. ☕
Opinionated world news that respects your time.
One bold take, the best counter, and the receipt(s) that prove it (all in sixish minutes).
Mugshot Poll 📊
Would you deploy Chinese open-source AI to save 75%?
Before we wrap up today…
Does crypto/Web3 intrigue you?
Is worrying about losing money or getting scammed stopping you from investing?
This course is for you!
Click to share (2 referrals get you free access to your very own crypto mastery email course).
Everything you need to unlock crypto in 5 days (for beginners).
You can read yesterday’s newsletter on the UK’s alien camera here.
For the love of coffee, see you tomorrow!
Enjoy your Tuesday and stay caffeinated.
How did we do?
Thanks for reading!
Are you subscribing?
Be sure to get your daily curse and coffee fix by hitting the daily knowledge button below.