The Coding Model Wars Have Begun
The Weekend Leverage, June 28th
Happy new model release week to all those who celebrate it. My fingers itch. My mind burns. I yearn to make stuff on my computer constantly. It is the most exciting time in technology history, and, theoretically, it’ll only get faster from here. Yay! Oh no!
How have you all been using Fable and Sol? Send me an email with cool projects you have built, I’ll share any standouts in the next edition of the newsletter.
Today we will be discussing why this week was so momentous, but first, this edition is brought to you by Working Smarter, a podcast from Dropbox about making AI work for you.
AI is a wonderful, magical technology that has the unfortunate habit of not having the full picture of what you need or what actually matters to you. Pointing the model at a problem and saying “please solve” is like using a flamethrower to cook eggs. It is just too overpowered. Working Smarter, the podcast from Dropbox, features Dropbox engineers who share how they’re building context-aware intelligence that connects to all the tools your team uses for work—so you get AI that works wherever you do.
What I find interesting about the pod is that this is the gritty work that doesn’t usually get the spotlight, but is absolutely necessary for AI to actually have a positive ROI. Working Smarter puts the builders on the mic to talk about what it takes to build the AI tools behind inference, security, and multimodal search that surfaces what you need fast. If you care about how AI gets built, this is the podcast to listen to.
With episodes on context engineering, multimodal search, agentic AI, security, and more, Working Smarter goes under the hood to show you what it takes to build AI that actually understands you—and how it can help you work smarter, too.
Listen to Working Smarter wherever you get your podcasts, or visit workingsmarter.ai.
Why can’t software make you cry? Over the last few months I’ve been grieving through a run of medical emergencies among people I love, so I took the raw weights of GPT-2, fed it heaps of public domain poetry (roughly half of it Whitman, because Whitman rules), and used Fable to build a product about my feelings. The end result is a poetry bot that helped me to understand why software has been so artistically barren, and what the future of creativity with AI may look like. You can go summon a poem yourself. Read here.
Bending Spoons just IPO’d at $25 billion. Is the whole thing kinda stupid? These are the people who buy dead software (Evernote, Vimeo, Eventbrite, AOL), lay off nearly everyone, raise prices on the customers who remain, and vow to hold it forever. Two years ago they were worth $2.8 billion. They closed their first day of trading at $25.7 billion. But if you believe, like I do, that AI is collapsing the cost of building software, why is a company whose entire identity is “we make software more expensive” trading at such a crazy premium? Read here.
Bots just became the majority of the internet. What happens to the rest of us? Cloudflare’s CEO announced that bot traffic passed human traffic for the first time ever, 57.4% to 42.6%, which makes us the digital minority. The whole economy of the web assumes humans are the ones looking at screens and getting sold to. In this video I walk through what breaks when AI agents become the ones reading, searching, and paying instead. Watch here (my most popular video yet!)
The Coding Model Wars have begun. The coding use case for LLMs will likely produce over $75B in revenue next year—an amazing feat for something that didn’t exist five years ago. This week, four major coding-capable models shipped within 48 hours, each representing a different strategic bet on how to get that revenue.
OpenAI took the premium position with GPT-5.6 Sol at $5/$30 per million tokens. It leads GPT-5.5 on Terminal-Bench 2.1 but trails Claude Fable 5 on SWE-Bench Pro. To translate this nerdy benchmarks stuff, these new models are powerful enough to be multidimensional. Different models lead on different kinds of work.
Meta attacked the economics with Muse Spark 1.1 at $1.25/$4.25 and a one-million-token context window. Sol is four times more expensive on input and roughly seven times more expensive on output, so Meta is betting that near-frontier capability at far lower prices will win high-volume workloads.
Grok 4.5 is the data-and-distribution play. SpaceXAI and Cursor trained it using trillions of tokens of Cursor interaction data, launched it throughout Cursor’s products, and priced it at $2/$6. Cognition (remember them?) made the specialization bet with SWE-1.7, applying reinforcement learning to an open-weight model and optimizing it for long-horizon work inside Devin.
Four serious releases in 48 hours is not a normal product cycle! Let me try to explain, simply and clearly, what is happening here.
The first thing to note with these launches is that frontier models can still charge several times as much as competing products that appear nearly as capable. They can do so because coding capability has nonlinear value. For simple tasks, a cheaper model that is slightly less reliable is often the rational choice. But there are relatively few of those in coding. Most coding projects consist of dependent steps, and failure at any one step can ruin the entire run. A small improvement in reliability can determine whether the agent finishes the project or leaves your poetry robot half-built in a broken repository (ask me how I know.)
When the alternative is several hours of engineer time, repeated retries, or a failed deployment, another $20 of inference cost is almost irrelevant. Yes, frontier models are “smarter” but the actual product they are selling is access to a wider set of tasks that can be completed reliably. Each model release that cross the reliability threshold creates the premium pricing power.
From there, we can use theory to understand why every firm in this market is all trying to get to the bleeding edge of model capability. The first is Schumpeterian competition. Companies invest heavily to create a technological leap, collect temporary profits from the advantage, and then watch those profits erode when the next breakthrough arrives.
The second theory is the Red Queen effect. When every competitor is improving rapidly, standing still causes you to fall behind. Schumpeter explains why the prize is so valuable while the Red Queen explains why the pace is so relentless. Together, they explain why every company capable of entering this market feels compelled to do so.
To forecast what happens next, we have to go back to frameworks.
For that, the useful starting point is Abernathy and Utterback’s theory of industrial innovation. Many young industries begin in a fluid phase where products differ widely, technical progress is fast, and companies compete primarily through product innovation. As the architecture stabilizes, competition shifts toward incremental improvement, cost, scale, and process efficiency.
Coding AI is clearly still a fluid market. Models differ in cost, context, speed, reliability, and—perhaps most importantly—in the application architecture surrounding the model. The market has not agreed on whether the central product is an IDE assistant, a terminal agent, an asynchronous software engineer, or an army of cooperating agents. Shoot, we do not even know what the dominant interface will be.
The next phase begins when users broadly agree on what a coding agent should do and swapping one base model for another no longer changes the product very much.
That is when Clayton Christensen’s theory of modularity becomes important. While performance is not yet good enough, tightly integrated systems tend to win because the model, agent, tools, and training environment must be carefully coordinated. Once a component becomes good enough, however, it can be standardized and swapped out.The attractive profits then move toward whichever part of the system is still not good enough.
For coding AI, that could be the agent, proprietary interaction data, enterprise security, workflow integration, distribution, or ownership of the customer relationship. Again, a long-ass list, because no one actually knows what it’ll be yet. Every company is finding an open strategic lane, going all in, and praying to the God of Capitalism that its chosen advantage becomes the important one.
And as a reminder, one I often have to repeat to startup founders, having the best technology does not guarantee capturing the profits. Again, great technology DOES NOT WIN. David Teece’s theory of complementary assets argues that value often accrues not to the inventor, but to whoever controls the assets necessary to commercialize the invention.
Put the theories together and the likely sequence looks like this:
Frontier capability retains a premium while reliability remains the binding constraint and integrated systems materially outperform modular ones.
As base models become good enough and easier to substitute, competition shifts toward price, efficiency, and distribution.
Profits then migrate to whichever layer remains scarce, while feedback loops between users, data, product quality, and distribution may allow one or two platforms to pull away.
We will know that transition is underway when customers stop switching to every new frontier release, model routing becomes standard, changing the base model has little effect on task completion, and frontier providers begin losing pricing power. My bet is that doesn’t happen anytime soon.
1X just launched the best robot hand I’ve ever seen. The hand has 25 degrees of freedom, is just as strong as a human hand, and moves freaky fast. As I’ve been arguing forever around here, robotics is going through its GPT2 moment right now. The technology now has a proven path forward and will, in about 2-3 years, take over the world’s attention, just like ChatGPT did. (One fascinating sidebar here is that the launch website is beautiful and well designed while the copy is grade A, 100% Claude slop. The writing is dogwater, trite, bad. I would love it if the quality of the copy could match the quality of the technology!)
Character.ai figured out what slop is for. The world’s most popular chatbot company launched the (c.ai) series, 3 in-house microdramas with AI-generated animation. The series is pretty standard teen type content with a romance anime, a haunted party game, and a Hunger Games knockoff. Each series runs 10 episodes of under 2 minutes, with the first 8 free and the final 2 paywalled. The company then combines the AI generated micro-dramas, with chatbots by allowing users over 18 to chat with the characters, interrogate them about the plot, and roleplay alternate storylines. As someone who has talked to many users of this type of thing, it’ll most likely be roleplay fantasies :(
Microdramas are the perfect petri dish for AI video because the format already made peace with being cheap. These are vertical soap operas that optimize for cliffhangers, and these things are printing money. Short drama apps generated $2.98 billion in in-app purchases in 2025, up 115% year over year, and while users of one of the apps spend 35.7 minutes per day in it, beating Netflix mobile’s 24.8. You know it is bad when I’m deafening Netflix as a beacon of quality content.
In my piece this week, I argued that AI art gets interesting when it stops imitating older media and leans into the one thing only this medium can do. For AI, the signature could be that the character talks back. This attempt by Character is one of the better attempts I’ve seen to do that.
CEO Karandeep Anand told The Hollywood Reporter the goal is “not to create an AI slop machine for Gen Z.” I mean…fail right? This is clearly, spiritually, slop. The Sloppening’s core math says infinite supply makes generic content worthless, and Character.ai is the first studio to build a business model that agrees: give the slop away and charge for the LLM. If you want to make not slop, you have to find a way to make your product elevate your user’s tastes.
Jack White’s new album rips: White is 51 and is still pissed off at the world. I mean this as a compliment. The album is an angry, punchy, blues-guitar record that is my favorite of his since the White Stripes days. It was so good that I bought tickets to see his show in Boston on Friday. I will report back on it!
Go and be kind this week,
Evan
Sponsorships
We are now accepting sponsors for the Q3 ‘26. If you are interested in reaching my audience of 35K+ founders, investors, and senior tech executives, send me an email at team@gettheleverage.com.