We do have some idea. Kimi K2 is a relatively high performing open source model. People have it running at 24 tokens/second on a pair of Mac Studios, which costs 20k. This setup requires less than a KW of power, so the $0.8-0.15 being spent there is negligible compared to a developer. This might be the cheapest setup to run locally, but it's almost certain that the cost per token is far cheaper with specialized hardware at scale.
In other words, a near-frontier model is running at a cost that a (somewhat wealthy) hobbyist can afford. And it's hard to imagine that the hardware costs don't come down quite a bit. I don't doubt that tokens are heavily subsidized but I think this might be overblown [1].
[1] training models is still extraordinarily expensive and that is certainly being subsidized, but you can amortize that cost over a lot of inference, especially once we reach a plateau for ideas and stop running training runs as frequently.
Is Kimi K2 near-frontier though? At least when run in an agent harness, and for general coding questions, it seems pretty far from it. I know what the benchmarks say, they always say it's great and close to frontier models, but is this other's impression in practice? Maybe my prompting style works best with GPT-type models, but I'm just not seeing that for the type of engineering work I do, which is fairly typical stuff.
I’ve been pretty active in the open model space and 2 years ago you would have had to pay 20k to run models that were nowhere near as powerful. It wouldn’t surprise me if in two more years we continue to see more powerful open models on even cheaper hardware.
Now, these models are a bit weaker, but they're in the realm of Claude Sonnet to Claude Opus 4. 6-12 months behind SOTA on something that's well within a personal hobby budget.
I haven't tried Minimax M2.5 yet. How do its capabilities compare to Qwen3 Coder Next in your testing?
I'm working on getting a good agentic coding workflow going with OpenCode and I had some issues with the Qwen model getting stuck in a tool calling loop.
Same with unsloth/gpt-oss-120b-GGUF:F16 gets 25 tps and gpt-oss20b gets 195 tps!!!
The advantage is that you can use the APU for booting, and pass through the GPU to a VM, and have nice safer VMs for agents at the same time while using DDR4 IMHO.
You might be correct when you say the global 1%, but that's still 83 million people.
have you paid any attention to the hardware situation over the last year?
this week they've bought up the 2026 supply of disks
$20,000 is a lot to drop on a hobby. We're probably talking less than 10%, maybe less than 5% of all hobbyists could afford that.
If you can't write requirements an engineering team can use, you won't be able to write requirements for the robots either.
this is marketing not reality.
Get a few lines of code and it becomes unusable.
The naivete on here is crazy tbh.
I walked into that room expecting to learn from people who were
further ahead. People who’d cracked the code on how to adopt AI at scale,
how to restructure teams around it, how to make it work. Some of the
sharpest minds in the software industry were sitting around those tables.
And nobody has it all figured out.
But I’d separate that from the programmer-level reality: a lot is already figured out in the small. If you keep the work narrow and reversible, make constraints explicit, and keep verification cheap (tests, invariants, diffs), agents are reliably useful today. The uncertainty is less “does this work?” and more “how do we industrialize it without compounding risk and entropy?”
I wrote up that “calm adoption without FOMO, via delegation + constraints + verification” framing here, in case it helps the thread: https://thomasvilhena.com/2026/02/craftsmanship-coding-five-...
This is one of the most interesting questions right now I think.
I've been taking on much more significant challenges in areas like frontend development and ops and automation and even UI design now that LLMs mean I can be much more of a generalist.
Assuming this works out for more people, what does this mean for the shape of our profession?
FOSS meant that the cost of building on reusable components was nearly zero. Large public clouds meant the cost of running code was negligible. And now the model providers (Anthropic, Google, OpenAI) means that the cost of producing the code is relatively small. When the marginal cost of producing code approaches zero, we start optimizing for all the things around it. Code is now like steel. It's somewhat valuable by itself, but we don't need the town blacksmith to make us things anymore.
What is still valuable is the intuition to know what to build, and when to build it. That's the je ne sais quoi still left in our profession.
“Ideas that surfaced: code as ‘just another projection’ of intended behaviour. Tests as an alternative projection. Domain models as the thing that endures. One group posed the provocative question: what would have to be true for us to ‘check English into the repository’ instead of code?
The implications are significant. If code is disposable and regenerable, then what we review, what we version-control, and what we protect all need rethinking.”
Absolutely. Also crucial is what's possible to build. That takes a great deal of knowledge and experience, and is something that changes all the time.
Second there’s a world of difference still between a developer with taste using AI with care and the slop cannons out there churning out garbage for others to suffer through. I’m betting there is value in the former in the long run.
In the past 6 months, all my code has been written by claude code and gemini cli. I have written code backend, frontend, infrastructure and iOS. Considering my career trajectory all of this was impossible a couple of years ago.
But the technical debt has been enormous. And I'll be honest, my understanding of these technologies hasn't been 'expert' level. I'm 100% sure any experienced dev could go through my code and may think it's a load of crap requiring serious re-architecture.
It works (that's great!) but the 'software engineering' side of things is still subpar.
We’ve been trying to build well engineered, robust, scalable systems because software had to be written to serve other users.
But LLMs change that. I have a bunch of vibe coded command lines tools that exactly solve my problems, but very likely would make terrible software. The thing is, this program only needs to run on my machine the way I like to use it.
In a growing class of cases bespoke tools are superior to generalized software. This historically was not the case because it took too much time and energy to maintain these things. But today if my vibe coded solution breaks, I can rebuild it almost instantly (because I understand the architecture). It takes less time today to build a bespoke tool that solved your problem than it does to learn how to use existing software.
There’s still plenty of software that cannot be replaced with bespoke tools, but that list is shrinking.
Claude Code is producing working useful GUIs for me using Qt via pyside6. They work well but I have no doubt that a dev with real experience with Qt would shudder. Nonetheless, because it does work, I am content to accept that this code isn't meant to be maintained by people so I don't really care if it's ugly.
If you want to get/stay good at debugging--again IMO--it's more important to be involved in operations, where shit goes wrong in the real world because you're dealing with real invalid data that causes problems like poison pill messages stuck in a message queue, real hardware failures causing services to crash, real network problems like latency and timeouts that cause services which work in the happy path to crumble under pressure. Not only does this instil a more methodical mentality in you, it also makes you a better developer because you think about more classes of potential problems and how to handle them.
Expert generalists are also almost impossible to distinguish from bullshitters. It’s why we get along so well with LLMs. ;)
I don't think you can find that level of ego anywhere in the software industry or any other industry for that matter. Respetc.
The text is actually about the Thoughtworks Future of Software Development retreat.
[0] Which is not even enough, these are the ones with truly excess money to burn.
Are you assuming tech debt has no financial cost?
I do like the idea that "all code is tech debt", and we shouldn't want to produce more of it than we need. But it's also worth remembering that debt is not bad per se, buying a house with a mortgage is also debt and can be a good choice for many reasons.
I suggest something like "Tidbits from the Thoughtworks Future of Software Development Retreat" (from the first sentence, captures the content reasonably well.)
Now producing code is _cheap_. You can write and run code in an automated way _on demand_. But if you do that, you have essentially traded upfront cost for run time cost. It's really only worth it if the work is A) high value and B) intermittent.
There is probably a formula you can write to figure out where this trade off makes sense and when it doesn't.
I'm working on a system where we can just chuck out autonomous agents onto our platform with a plain text description, and one thing I have been thinking about is tracking those token costs and figuring out how to turn agentic workflows into just normal code.
I've been thinking about running an agent that watches the other agents for cost and reads their logs ono a schedule to see if any of what the agents are doing can be codified and turned into a normal workflow, and possibly even _writing that workflow itself_.
It would be analogous to the JVM optimizing hot-path functions... ---
What I do know is that what we are doing for a living will be near unrecognizable in a year or two.
Here’s a free idea I’ve had that I have no idea how to implement. I hope somebody much smarter than me will come along, think it’s a great idea, and steal it. I highly encourage you to do so, and I wish you well.
The idea is to have some kind of substrate—like a superpowered AST—that is the true code: the thing that actually gets compiled and run. Humans never look at this directly. Instead, we look at a representation of this code, and we can toggle between different representations of it.
I’m borrowing ideas from topology in mathematics here: if I look at a shape one way, I should be able to transform it into a different shape, but isomorphically, everything is still the same. That would let me look at the same thing in different ways, understand it from different angles, critique it more easily, and maintain it more easily.
Gemini tell me that this idea has already been tried in the past? Projectional Editing? Intentional Programming?
A useful complement is the programmer-level shift: agents are great at narrow, reversible work when verification is cheap. Concretely, think small refactors behind golden tests, API adapters behind contract tests, and mechanical migrations with clear invariants. They fail fast in codebases with implicit coupling, fuzzy boundaries, or weak feedback loops, and they tend to amplify whatever hygiene you already have.
So the job moves from typing to making constraints explicit and building fast verification, while humans stay accountable for semantics and risk.
If useful, I expanded this “delegation + constraints + verification” angle here: https://thomasvilhena.com/2026/02/craftsmanship-coding-five-...
Personally, I'm more interested in whether software development has become more or less pay to win with LLMs?
When we have solid tests, the agent output is useful and we can trust it. When tests are thin or missing, the agents still ship a lot of code, but we spend way more time debugging and fixing subtle bugs.
I agree that AI tools are likely to amplify the importance of quick cycles and continuous delivery.
Local or self hosted LLMs will ultimately be the future. Start learning how to build up your own AI stack and use it day to day. Hopefully hardware catches up so eventually running LLMs on device is the norm.
> One large enterprise employee commented that they were deliberately slow with AI tech, keeping about a quarter behind the leading edge. “We’re not in the business of avoiding all risks, but we do need to manage them”.
I’m unclear how this pattern helps with security vis-à-vis LLMs. It makes sense when talking about software versions, in hoping that any critical bugs are patched, but prompt injection springs eternal.
We've experimented with rolling open source models on local hardware, but it's so easy to inject things into them that it's not really going anywhere. It's going to be a massive challenge, because if we don't provide the tools, employees are going to figure out how to do it on their own.
Yes, but some are mitigated when discoverd, and some more critical areas need to be isolated from the LLM so taking their time to provision LLM into their lifecycle is important, and they're happy to spend the time doing it right, rather than just throwing the latest edge tech into their system.
This isn't a case where you have specific code/capital you have borrowed and need to pay for its use or give it back. This is flat out putting liabilities into your assets that will have to be discovered and dealt, someday.
Chinese open source models are dirt cheap, you can buy $20 worth of kimi-k2.5 on opencode and spam it all week and barely make a dent.
Assuming we never got bigger models, but hardware keeps improving, we'll either be serviing current models for pennies, or at insane speeds, or both.
The only actual situation where tokens are being subsidized is free tiers on chat apps, which are largely irrelevant for any sort of useful economic activity.
I think this is often a mental excuse for continuing to avoid engaging with this tech, in the hope that it will all go away.
What people probably get messed up on as being the loss leader is likely generous usage limits on flat rate subscriptions.
For example GitHub Copilot Pro+ comes with 1500 premium requests a month. That's quite a lot and it's only $39.00. (Requests ~ Prompts).
For some time they were offering Opus 4.6 Fast at 9x billing (now raised to 30x).
That was upto 167 requests of around ~128k context for just $39. That ridiculous model costs $30/$150 Mtok so you can easily imagine the economics on this.
There's a difference between running inference and running a frontier model company.
https://www.theinformation.com/articles/anthropic-lowers-pro...
You're putting way too much faith in Dario's statements. It wasn't "abundantly clear" to me. In that interview, prior to explaining how inference profits work, he said, "These are stylized facts. These numbers are not exact. I'm just trying to make a toy model," followed shortly by "[this toy model's economics] are where we're projecting forward in a year or two."
Token costs are also non-trivial. Claude can exhaust a $20/month session limit with one difficult problem (didn't even write code, just planned). Each engineer needs at least the $200/mo plan - I have multiple plans from multiple providers.