08:00Zubair Trabzada | AI Workshop · Talking Head
Claude Mythos 5 Is Not for You. Here's Why
An 8-minute walk through Anthropic's own announcement that explains why the model everyone is talking about isn't the one you can actually use.
June 10thThe argument in one line.
A little-known lab's open-weight model matched about 98% of a leading closed AI model's quality at a third of the price, forcing a 48-hour pricing reversal that shows high AI subscription prices cannot survive once a cheaper equivalent exists.
Read if. Skip if.
READ IF YOU ARE…
- You build products or services on top of AI APIs and want to know whether this week's price war changes what you should be paying.
- You're deciding whether to keep paying for a premium AI subscription or switch to a cheaper open-weight alternative.
- You want a plain-English rundown of what actually changed in AI pricing this week without digging through benchmark leaderboards yourself.
SKIP IF…
- You're looking for a hands-on tutorial on installing, hosting, or fine-tuning the open-weight model — this is a news breakdown, not a how-to.
- You need enterprise compliance or safety-tuning detail before adopting a new model — the video covers pricing and benchmarks, not governance.
The full version, fast.
On July 16, Chinese lab Moonshot released Kimi K3 — a 2.8 trillion parameter open-weight model with a 1 million token context window, priced at $3 in / $15 out per million tokens, with full weights going public July 27 under a modified MIT license. Independent tests found it landed roughly 98% of the leading closed model's quality at a third the price, just slower. By Friday, Anthropic reversed a month of calling its premium pricing "temporary" and made it permanent instead, while raising per-use rates on its lower tier. The lesson: AI capability costs are falling 30-40x a year, so high subscription prices don't survive once a cheaper alternative appears — and builders who use these models, rather than just consume them, see their margins expand as the underlying cost shrinks.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:44
01 · A Big Lab Panics
Cold open: what it looks like when a $100B AI company panics, and the three-day timeline that's about to unfold.
00:44 – 01:15
02 · The Toll Booth Era
For two years, two labs set AI pricing because nobody else matched their frontier models at any price — $100-200/month plans, and the bet that nobody would match it for less.
01:15 – 02:09
03 · Moonshot Drops Kimi K3
Moonshot ships Kimi K3: 2.8T parameters, 1M token context, weights go public July 27 under a modified MIT license, priced at $3 in / $15 out per million tokens.
02:09 – 02:47
04 · Benchmarks and Receipts
K3 tops a coding leaderboard and scores 88.3 on Terminal-Bench, 93.5% on GPQA Diamond, but places third on GDPval — close enough that price starts deciding.
02:47 – 03:29
05 · Price Beats Prestige
The Twin Test: identical prompts on both models land roughly 98% of the pricier model's quality at a third of the cost, just slower.
03:29 – 04:20
06 · Anthropic's 48 Hour Reversal
A month of sliding 'temporary' deadlines ends in 48 hours: the top plan becomes permanent, and the lower tier gets its steepest-ever per-token pricing.
04:20 – 05:08
07 · Open Weights Break Moats
AI capability cost is falling 30-40x per year; open weights add a cheaper 'second door' next to the expensive one, and high prices don't outlive it.
05:08 – 05:53
08 · Builders Win the Price War
Consumers save money this week; builders raise their margins, because clients buy outcomes, not tokens, and the cost line underneath just shrank.
05:53 – 07:49
09 · Jarvis on Closed vs Open
Live demo: the creator's own AI assistant, Jarvis, compares closed vs open models in its own words — polish and support vs ownership and DIY plumbing.
07:49 – 08:24
10 · Scorecard and What's Next
Three things to watch: the July 27 weight release, OpenAI's yet-unmade move, and reading every future plan-update email as a price-war chess move.
08:24 – 08:35
11 · Wrap Up and Comments
Sign-off and call for viewer comments.
Lines worth screenshotting.
- A Chinese lab nobody was watching shipped a frontier-class AI model at a third of the price, and the biggest lab in the industry reversed a month-long pricing decision within 48 hours.
- Kimi K3 is a 2.8 trillion parameter open-weight model with a 1 million token context window, priced at $3 input / $15 output per million tokens.
- On a benchmark measuring real work across 44 occupations, the open-weight model still finished third behind two closed frontier models — it didn't crush the frontier, it got close enough that price started deciding.
- With prompt caching, the open model's input cost drops to 30 cents per million tokens — a tenth of the closed model's list price.
- Creators running identical prompts on both models on day one found the cheaper one landed roughly 98% of the pricier model's quality, just slower.
- A subscription pricing deadline slid three separate times over a month, then flipped from 'temporary' to 'permanent' within 48 hours of a cheaper competitor's launch.
- The steepest per-token price on any current model in the lineup — $10 in / $50 out — is what lower-tier subscribers now pay after burning a one-time $100 credit.
- The cost of frontier AI capability is falling 30 to 40 times per year, so any feature that justifies a $200/month plan risks being cloned by an open-weight model within months.
- Cheap intelligence doesn't kill the builder economy — it subsidizes it, because clients pay for outcomes like reports and websites, not for tokens.
- Models function like engines you rent or own; the skills, workflows, and systems built on top are the actual moat, and they transfer across whichever model wins the price war.
- The open-weight model's full weights go public July 27 under a modified MIT license, letting anyone host, run, or resell access to it.
- OpenAI had not responded to the price cut within this week's window — whatever it does next sets the following price floor for the whole market.
Why a cheaper AI model just gave builders a raise
PRICE WAR PLAYBOOK
A cheaper open-weight model forced a 48-hour pricing reversal, proving that high AI subscription prices can't survive once a comparable 'second door' opens — and that builders, not just consumers, come out ahead when it happens.
02The Toll Booth Era
- For two years, two labs set AI pricing because nobody else matched their frontier models at any price.
- The best models sit behind $100-200/month consumer plans — when nothing else comes close, the price effectively becomes the product.
- That pricing structure rested on one bet: that nobody would match the frontier for less, and for two years nobody did.
03Moonshot Drops Kimi K3
- A rival lab released a 2.8 trillion parameter open-weight model with a 1 million token context window, the largest open-weight model ever announced.
- Its full weights go public under a modified MIT license, meaning anyone can host it, run it, or resell access to it.
- It's priced at $3 input / $15 output per million tokens — sonnet-class pricing for frontier-class output.
04Benchmarks and Receipts
- The open model went straight to number one on a front-end coding leaderboard, ahead of both major closed frontier models.
- It scored 88.3 on Terminal-Bench and 93.5% on GPQA Diamond — the strongest open-weight results published to date.
- On a benchmark measuring real work across 44 occupations, it placed third, behind both closed frontier models — close, but not a sweep.
05Price Beats Prestige
- Two near-identical models cost $10 in / $50 out versus $3 in / $15 out, and with caching the cheaper model's input drops to 30 cents.
- Creators running identical prompts on both models on day one found the cheaper one landed roughly 98% of the pricier model's quality, just slower.
- When quality ties or even comes close, price always picks the winner.
06Anthropic's 48 Hour Reversal
- The pricing deadline for the incumbent's premium plan slid three separate times over a month before the reversal.
- By Friday, the top-tier plan became permanent, capped at half of prior usage limits, effective immediately.
- Lower-tier subscribers get a one-time credit, then pay-per-use at the steepest per-token price on any of that lab's models.
07Open Weights Break Moats
- The cost of AI capability is falling 30 to 40 times per year, so any feature that justifies a premium plan can get cloned by an open model within months.
- Open-weight models change the old rule: instead of one priced door, competitors now open a second, cheaper door offering comparable intelligence.
- High prices do not outlive second doors — once a cheaper equivalent exists, buyers quietly move through it instead of paying the toll.
08Builders Win the Price War
- If you only consume AI, a price war saves you money; if you build with AI, it raises your margins, because your cost line shrinks.
- Clients don't buy tokens, they buy outcomes like reports, audits, and websites, so the invoice stays the same while the underlying cost drops.
- Models function like engines: you can rent or own one, but the skills, workflows, and systems built on top transfer across whichever model you use.
09Jarvis on Closed vs Open
- Closed models get frontier-level polish and safety tuning, but you rent them forever; open-weight models you own outright and can run privately, at the cost of doing your own maintenance.
- If an open model matches a closed one on reasoning and tool use, the savings are real — but test it on your own actual workflows first, since 'benchmark hero, production disaster' is a well-documented pattern.
10Scorecard and What's Next
- Watch three things going forward: the full open-weight release date, the incumbent competitor's still-unmade response, and every future plan-update email read as a move in an ongoing price war.
Terms worth knowing.
- Open-weight model
- A model whose trained parameters are published publicly, so anyone can download, host, fine-tune, or run it themselves instead of only accessing it through the maker's paid API.
- Context window
- The maximum amount of text, measured in tokens, a model can hold in memory at once during a single conversation or task.
- Prompt caching
- A pricing discount where repeated or reused portions of a prompt are billed at a fraction of the normal input-token rate on subsequent calls.
- GDPval
- A benchmark that scores AI models on realistic work tasks across 44 different occupations, rather than narrow academic test questions.
- GPQA Diamond
- A benchmark made up of graduate-level science questions, used to test a model's expert-level reasoning ability.
- Fable 5
- The name this video and its metadata use for the AI lab's flagship subscription-tier model (Max/Team/Pro plans, $100-200/month), in place of the model's officially branded name.
Things they pointed at.
01:15productKimi K3 (Moonshot)
00:44productClaude Max / Team / Pro plans (Anthropic)
04:14toolGDPval benchmark
Lines you could clip.
00:00
“So what does it look like when a $100,000,000,000 company panics? It looks like this.”
00:59
“When nothing else comes close, the price is the product.”
02:58
“When quality ties or even comes close, price always picks the winner.”
04:01
“That's not generosity. That's competition.”
04:55
“High prices do not outlive second doors.”
06:52
“In short, I'm a butler on retainer. They're a very capable stray you have to house train.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
So what does it look like when a $100,000,000,000 company panics? It looks like this.
Wednesday, a Chinese lab nobody was watching ships a Frontier model at a third of the price. Thursday, the entire Internet runs the same exact comparison and it actually holds up.
By Friday, Antropic reverses a decision, they defend it for an entire month and calls it permanent. Three days, start to finish. Now for the past several months, I've been building with these AI models every single day and I have never seen the big labs move this fast.
So I put together the most important weak AI this year on one page. What dropped? What actually scores?
Why the price is the real weapon? And at the end, the one group of people who profit from this no matter which lab cuts prices next. Alright.
Let's jump right in. First, the setup. Now for the past two years, two companies have owned the best AI models and they set the tool.
OpenAI and Entropic. The best models sit behind the priciest subscription and consumer software. $102,100 dollars a month.
And look, I pay for the top plan myself happily because when nothing else comes close, . But that toll rests on one bet. Nobody matches the frontier for less and for two years, nobody actually did.
Now keep that sentence in your head. Then Wednesday, July 16, a Chinese lab called Moonshot chips a model called Kimi k three. This is not a toy.
It lands straight in the top tier. AI stocks dip within hours. Every feed I follow called it the same thing, a new deep seek moment.
And those two expensive shops we just talked about, their week was about to get very very hard. So let's do the numbers now. 2,800,000,000,000 parameters.
That's the biggest open weight model in history. 1,000,000 token contacts window, so that basically holds entire projects in its head.
July 27, the waits go public under a modified MIT license, which means anyone can host it, run it, or literally sell the access to it. And then the price, $3 in, 15 out per million token.
Basically, sonnet class pricing for Fable class output. And honestly, only that green number matters.
Alright. So now the receipts because I don't want you running on hype. On Arena's front end coding board, it went straight to number one.
Above Fable five, above GPT 5.6. Terminal bench, 88.3. GPQA diamond, 93 and a half percent and the strongest open weight results ever published.
But here's the honest part. On the GPT Val, which is basically the benchmark that measures real work across 44 occupations, it actually came third behind Fable five max and behind g p t 5.6. So no, it did not crush frontier, it did something big labs should fear more.
It got close enough that price starts making the decisions for you. And this is the picture that should keep every pricing team in San Francisco up at night. Two identical brains, one cost $10 in and 50 out.
The other cost 3 in and 15 out. And with caching, k three's input drops to 30¢. That's remarkable.
Now creators ran the same prompts on both models on day one. Three d globe, Kimi better. Galaxy simulation, Kimi won, but took three times longer.
Stacking game tie research pages even. The verdict going around roughly 98% of Fable five at Sonnet prices.
Slower, cheaper, extremely close.
And when quality ties or even comes close, price always picks the winner. Now watch the timeline because the timing is the actual story here. Because for an entire month, Antropic kept calling Fable five on subscriptions temporary.
The removal deadline kept sliding July 17, then twentieth, then the nineteenth, and then Wednesday, KimiKate three drops. Thursday, every big AI channel runs the same comparison and one question is everywhere. Why pay Fable five prices at Kimi quality?
And Friday, Entropic announces Fable five is now permanent. Max and team premium keeps it for good capped at half your usage limits effective today which is July 20. Pro users get one time $100 credit then it's pay per use at 10 in and 50 out which is the steepest per token price on any cloud model.
Now here's the reality that nobody says on stage. That's not just generosity, that's competition and I absolutely love it.
This is competition actually doing its job. Now if we zoom out, the pattern is actually brutal. The cost of AI capability is falling 30 to 40 times per year.
Anything that justifies a $200 plan gets cloned by an open model within months. The old rule was you pay the toll or you actually lose the tool itself. It was one door with a price tag on it, but open weights changed that building completely.
Now, the wall grows doors. Over here, the fancy door, which is velvet ropes price tag. And right next to it, a second door wide open where the same exact intelligence walks out for a third of the price.
So nobody storms the fancy door anymore. They just quietly walk through the other one. .
And on July 27, when those waits go public and everyone races to host the Kimmy k three model, is the cheapest model, the floor drops again. Now here's the part that everybody misses and it's my favorite part. If you only consume AI, this week saved you actual money.
Nice, but if you built with AI, this week raise your margins. Think about it. Your cost line just shrank.
The same agent workflows I run every day now cost a third at the token level and it keeps falling. Your clients don't buy tokens, they could care less. Right?
They buy actual outcomes, reports, audits, website, employees, so on and so forth. The invoice stays the same while the cost underneath it shrinks.
And . You own the car, your skills, your workflows, your systems, they transfer across every single model.
The stack on top is the mode. Cheap intelligence does not kill the builder economy, it actually subsidizes it. So I'm gonna give you a quick example.
For instance, this is my AI second brain. Basically, everything in my business runs through my assistant Jarvis here, but essentially all of this sits on top of Fable five right now.
As you can see, even Jarvis runs on Fable five for me. And a lot of this is through API. So you can imagine everything I build, it cost a lot of money.
Jarvis, so a new model came out, an open weight model came out in the market. And since you run on Fable five, I was just trying to see what do you think about that, the difference between a paid model like Fable five and, like, an open source model like Kimi three.
Three, sir. Closed models like myself get frontier level polish and safety tuning, but you rent us forever. Open weight models you own outright can fine tune and run privately at the cost of doing your own plumbing.
And support wise, one comes with a lab behind it, the other with . In short, I'm a butler on retainer. They're a very capable stray you have to house train.
Okay. Thanks, Jarvis. I just wanted to get your opinion on, uh, the, you know, free model technically or a open weight model because I wanted to see if it makes sense for me to switch your brain to this open weight model.
A fair question, sir. Though asking me whether to replace my own brain is a touch like asking the butler if he'd fancy being swapped for a Roomba. Honestly, if the open model matches Fable five on reasoning and tool use, the savings are real, but test it on your actual workflows first because benchmark hero production disaster is a well documented genre.
That's actually a good point that Jarvis mentioned because as soon as open weight models like this hit the market, the first thing that happens is obviously people start to build on top of it. Right? And that's the real test because so far it's been kind of just kind of an introduction in the market.
So once these open weights actually become public, that's when the real test becomes. But so far it's looking incredibly good.
Alright. So finishing this off with the scorecard. A lab you had never heard of matched the best model in the world at a third of the price.
The biggest lab on the planet answered within forty eight hours and your subscription got better because the war is still on. Three things I'm watching from here. One, July 27, the waits go public and the race to host k three, the cheapest begins.
Two, OpenAI has not answered yet and their reply sets the next price floor. And three, every plan update email you get from now on reads it like a chess move in the price war because that's exactly what it is. Anyways, hopefully, you guys found this helpful.
Let me know in the comments your thoughts, would love to hear them. Thanks for watching and I'll see you in the next one.
The Hook
The bait, then the rug-pull.
A hundred-billion-dollar AI lab spent a month insisting its pricing change was temporary — then reversed course in 48 hours after a lab nobody was watching shipped a frontier model at a third of the price.
Named ideas worth stealing.
00:48concept
Two Shops (The Toll Booth)
- The toll: best models sit behind $100-200/month subscriptions
- The bet: nobody else matches the frontier for less
The two-year-old pricing structure of the closed AI labs, framed as two shops that jointly set the market price because no competitor could match their quality at any price.
Steal forframing any premium/subscription product's pricing power against the absence of real competition
02:44model
The Twin Test
- Run identical prompts on both models
- Roughly 98% of the pricier model's quality at a third of the price — slower, cheaper, extremely close
A head-to-head method for deciding whether a cheaper alternative is 'close enough' to switch: run the same real tasks on both and compare output quality against price.
Steal forbenchmarking your own AI tool choices before switching vendors
04:27concept
The Second Door
- Old rule: pay the toll or lose the tool
- New rule: the wall grows doors — high prices do not outlive second doors
Once an open-weight model offers comparable intelligence at a fraction of the cost, buyers don't fight the expensive door — they quietly walk through the cheaper one instead.
Steal forframing how a lower-cost competitor disrupts any premium-priced market
05:08list
Builders Win the Price War
- Your cost line just shrunk
- Clients don't buy tokens, they buy outcomes
- Models are engines — you own the car; skills, workflows, and systems transfer across every model
Why falling AI costs raise margins for people who build with AI rather than just consume it: the invoice to the client stays the same while the cost underneath shrinks.
Steal forpositioning an AI-powered service business against commoditizing model costs
How they asked for the click.
VERBAL ASK
08:25link
“skool.com/aiworkshop-lite (shown on screen only, not spoken)”
The pitch is entirely visual — the final graphic overlays the free-resources Skool link while the host says only a sign-off with no verbal ask. The paid offer (Build AI Employees + JARVIS AI Assistant) appears only in the description, never mentioned in the video itself; the Jarvis demo segment doubles as an organic showcase of the creator's own product.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
