by Luis Rodrigues
What I’m building with AI + one hot take + five links worth your time.
The Take
This weekend I tried the new Kimi K3 to save some money. It just hit the Arena.ai coding leaderboard and costs 1/3 of Fable.
Kimi costs $3 and $15 per million tokens, vs. $10 and $50 per million tokens for Fable.
Top of the table, 1/3 of the cost. That was the theory before I started.
Source: Arena.ai - Tops the board, 30% of the price
Then I ran it. In my test, it took 3 minutes, where Fable takes 45 seconds. It spent about 12,000 tokens, roughly 70% of them on reasoning.
It kept saying "wait, but...". The model was arguing with itself, and I was paying in tokens.
Artificial Analysis has K3 at $0.94/task, cheaper than Fable at $2.75, roughly the same as Sol's $1.04.
So it's cheaper than Fable per task, but against Sol, half the sticker price buys the same real cost. The savings went into reasoning.
I understand people are excited about an Open Fable-level model, but for now, K3 has reasoning set to Max, and you can't lower it, so it questions everything.
Last week I wrote that the winner will be the cheapest / more capable model. This week, a small correction: cheaper per token is not necessarily the cheapest. What matters is the cost to get the job done and how long it takes to finish.
Why does this matter if you're building or getting ready to start? Run your own tasks in the models you have. The price table doesn't show you token usage.
“AI is cheaper" and "AI is cheaper for me" are two very different things.
The Build Log
One of our clients in thoughtled.ai wrote a nice post last week. Then she needed an image: opened Canva, drew it, exported the image, back to Thoughtled, uploaded and published.
This is how employee advocacy dies. It's not because people don't want to post. It's because every time we force them to leave the tool, someone quits.
Imagine, she moves to Canva, has to log in. In the meantime, a Slack notification pops up, and she forgets about the post.
We needed an image editor that is as simple to use as Canva. Build vs buy. An image editor is a product in itself, and we're two, so the decision is easy.
After searching the web, Polotno seems like the best solution; it supports everything and has very good licensing terms.
It took two days to integrate. The hardest part (which they've already solved) was that our code is React 19 and theirs was React 18.
Time saved is not the real story. With Polotno, we have a new editor that behaves like Canva. It's not an exact copy but has the same conventions, and conventions are what people learn.
Nobody needs to explain to our clients where the text tool is. If we had built our own, it would have taken months and probably wouldn't work as well.
Here we have the boring stack argument from backend to frontend again. I picked Django because it's well-documented and Claude can write APIs one-shot.
Polotno is the same bet, but the training data is in the user's head because they already know how all those buttons work.
Our product is not an image editor. We're building a voice transfer and post-coordination tool.
Those are the features we need to focus on. Everything else we need to find the best solution on the market and buy it.
The list of features that represent your product is smaller than you think. Everything out of that list is someone else's software. Focus on what you know well and buy the rest.
If you're not building yet, the same point holds: focus on the tasks only you can do. Buy the rest.
On My Radar
Cognition swapped Opus 4.8 for Fable 5, and Devin's cost went down despite Fable costing 2x per token.
It uploads entire repos to a GCS bucket completely unredacted. Another example of why you need to check what these tools do with your data.
Their solution to RAG is to use a subagent that plans and tries several sources and returns structured "I don't know" instead of a wrong answer.
Vercel AI gateway data is another sign that open models are starting to bite into closed models usage. Cost per task beats trust-me-bro benchmark.
In 8 days, an agent improved the harness that the team had hand-tuned for 2 years.
I’m opening 2 consulting slots for July
Interested? https://cal.com/lfrodrigues/consulting
Know someone who can benefit from this email? Forward it.
New here? Someone forwarded this to you? → Subscribe to Build What Matters: https://luisrodrigues.ai/
How was this issue?
