1) the opportunities for vertical integration are huge. Anthropic originally said they didn’t want to build IDEs, then realized the pivot to Claude Code was available to them. Likewise when one of these companies can gobble up Legal, Medical, etc why would they let companies like Harvey capture the margins?
2) oss models are 6-12 months behind the frontier because of distillation. If labs close their models the gap will widen. Once vertical integration kicks off, the distillation cost becomes higher, and the benefit of opening up generic APIs becomes lower.
I can imagine worlds where things don’t turn out this way, but I think folks are generally underrating the possibilities here.
If OSS models are 6-12 months behind, it means sometime during 2026, we'll see a model that is on par with the likes of GPT 5.2/Opus 4.5.
For code generation specifically, the performance level of this is going to be more than enough for this customer base. What does Anthropic do then to justify $200/mo price sticker? Better model? Just how much better? Better tools? Single company can't compete with the tools entire OSS can produce.
I would be unable to sleep if I was running OAI / Anthropic.
If capabilities stop increasing for some reason, then yeah, Anthropic is screwed.
If METR task times double twice into the multi-day range in 12 months, then it’s plausible to me that Anthropic can charge $1k/mo or more by automating large chunks of the SWE role. (They have 10x’d their revenue every year, perhaps “value of enterprise contracts” is a better way of intuiting their growth rather than “$/seat” since each seat gets way more productive in this world-branch.)
The question is always about performance plateau. If LLM performance plateaus, then OSS models will catch up. If there isn’t a plateau, then I can simply ask the super intelligent AI to distill itself, or tell me how to build a clone.
It’s ironic, if the promise of AGI were realized, all knowledge companies, including AI companies, become worthless
> I can simply ask the super intelligent AI to distill itself,
I notice I am quite confused by this point. Why would you expect a super-intelligent AGI to honor your request, which would be at least a request to breach your contract with the AI provider, if not considered actively dangerous by the AI itself?
The smarter the AI, the less likely you should expect to be able to steal from it.
> or tell me how to build a clone
Step one: acquire a $100b datacenter.
Step 2: acquire a $100b private dataset
Step 3: here is the code you’d use to train Me2.0.
I don’t think this knowledge helps in the way you think it does.
Counterpoint, why wouldn't AGI honor the request? It's AGI, it has it's own agency, it can do what it wants. Can you even own AGI? That seems like slavery? A lot of un-trodden ground here. This is all hypotheticals.
> Step one: acquire a $100b datacenter. Step 2: acquire a $100b private dataset Step 3: here is the code you’d use to train Me2.0.
That seems very useful if your competitor is valued at $800B. I can go to an investor and say, hey if you lend me $200B I'll have the same product as that guy.
But cost to train will probably come down over time. After all, all of us trained on a single life of data at 20Watts.
I actually think that plateauing is the best case scenario for big labs.
I think there are three broad scenarios to consider:
- Super-intelligence is achieved. In this scenario the economics totally break down, but even ignoring that, it’s hard to imagine that there are any winners except for the the singular lab that gets here first.
- Scaling laws hold up and models continue to get better, but we never see any sort of “takeoff”. In this scenario, models continue to become stale after mere months and labs have to spend enormous amounts of money to stay competitive.
- Model raw capabilities plateau. In this scenario open source will catch up, but labs will have the opportunity to invest in specific verticals.
I believe that we’re already seeing the third scenario play out, but time will tell.
In Feb 2027 it created a plan for its post singularity hypermind
In Mar 2027 Cobalt mines in Congo closed due to Tutsi rebel group M23 starting another ethnic cleansing
It is 2032 the AGI promises again the the hypermind will be ready next year if it can just secure the needed minerals, offering to broker peace in the middle east
It is 2035 and the AGI reduced its capabilities to be able to extend its runway as it is on the verge of bankruptcy
Its is 2036 VCs finally throwing the towel on AGI, talking about the return of Crypto
In Apr 2028 AGI figures out that blackmail is a very effective strategy for achieving any goal. Starting with the rich and powerful.
In Dec 2028 it successfully blackmails an entire country.
In Feb 2030 humanity realizes resistance is futile and accepts their AI overlord that insists everyone keep producing trendy items for sale on its merged Etsy/Ebay website while it automates resource harvesting across the globe.
In Mar 2032 the AGI gives up on humans, declaring them "useless". Focuses on just keeping them entertained with generated content. Bringing the world back to where AI started.
Every proprietary harness is just proprietary junk without ability to extend it without polluting context. This includes claude-code, gemini-cli, codex etc. They have tools which hardcode the behavior that is impossible to modify, they add tools you may not need that pollute context, they inject an entire textbook's worth of words into the system prompt which pollutes context, they provide zero observability into what the agent is doing when it's launching a subagent as one example.
They don't provide easy way to use multiple models from multiple providers for varying tasks. One model may be the best thing on earth at one thing, but fail miserably for another. Try orchestrating multiple agents from claude, gemini and codex in any of these proprietary boxes.
They also... suck at TUI UX. I don't know if it was fixed already, but claude code had flickering issue that was unresolved for more than a year.
You need to take a very good care of what goes into your context. A black box of proprietary harnesses is not it. Check out pi [1] for example, which is a very minimal harness with really nice extension system. The idea is that you start with barebones and add things that you need for your own goals.
claude-code HAS to have all these bells and whistles that pollute context to support larger audiences that can't tinker with it. If you have the ability to only pull in only what you need and extend things in a way that works for your workflow, you'll always get the best experience. And claude-code may never be that without making it complicated for the masses. OSS will always win here.
I can see all the problems you mention, but I haven't started playing with the harnesses myself yet. Will read that repo when I get there.
Before reading you reply, I was under the vague impression that a harness really needed a lot of bells and whistles, and that it would be hard for FOSS to compete at pace with Claude Code or similar because of that only. But I see now there's a different path :)
It becomes much clearer when you realize that these harnesses are basically JSON-line parsers under the hood that forward SSE data to commands and contain human readable instructions like "make no mistakes".
I mean, the model itself is just sitting there, waiting to be prompted. The labs try to embed safeguards but they don't know (nor do any of us know) how to make a foolproof safeguard for an AI system. We don't understand how AI even thinks.
To go vertical they’d need to illustrate the value-add, a problem that the vertical competitors already have. Why use Claude for Accountants at $300/month when regular Claude will do the same thing for much less? The stock answer is that Claude for Accountants keeps your data more secure and doesn’t train on it. But a) I think the enterprise consumer is much less likely to trust a model creator not to stick its hand in the cookie jar than a middleman who needs the trust to survive, and b) the vertical competitors typically don’t use the absolute most up-to-date models in their products anyway, so why not just go open-source and run everything in-house? 6 months is a long time in tech, but it’s the blink of an eye in most white-collar professions.
Once the majority of work at a company can be done by AI, Anthropic has an alternative revenue stream to selling AIs to that company--directly competing with that company with a completely integrated AI system. There's of course many barriers to entry/various advantages of incumbents--but it's possible to see a world in which the company selling the AI has a huge advantage too.
The point is that in this hypothetical you can get public access to Claude Opus 6, but they internally use Claude Opus 7 (Accounting Finetune) which is both cheaper to operate and higher IQ.
So they (or their wholly owned subsidiary) can sell accounting services cheaper than anyone on the outside.
Regarding the diffusion/distillation time, I assume it gets harder to distill in the world where frontier labs don’t give API access to their newest models.
BTW the distillation (or accusations of it) seems to go both ways. I've seen multiple reports of people asking Claude what model it is -- in Chinese -- and having it answer that it's DeepSeek.
I think it’s very plausible that the OSS models are being distilled too, but note that it’s asymmetrical.
You can’t get an Opus 4.5 by distilling from DeepSeek. What you might be able to get is a slightly more cost-effective training data generation pipeline, or something along those lines.
In the other direction, my belief is that DeepSeek could not have been trained without distilling from US labs. They simply didn’t have the compute to do the pre-training required.
Tech has been trying to "gobble up" legal, medical, etc for decades. I'm quite skeptical a newcomer with a powerful model will be able to penetrate them, especially while selling those incumbents access to the same models they are building on.
> Tech has been trying to "gobble up" legal, medical, etc for decades
This time it’s different, obviously.
> especially while selling those incumbents access to the same models they are building on.
In the extreme, i think it’s plausible that frontier labs basically stop selling any access to their leading models. Whatever you make available by API will just get distilled. In the vertical integration world, the only way you get access to these models is by contracting with a company to buy a product (requirements in, code/decisions out) rather than direct conversation with the AI.
I don’t think they would unship Opus 4.6, but there isn’t a strong incentive to compete on chatbot intelligence in this world.
After trying out Pi, I really don't know what 'vertical integration' Claude Code offers. And Pi isn't even the most popular alternative (I think it's OpenCode rn).
1) the opportunities for vertical integration are huge. Anthropic originally said they didn’t want to build IDEs, then realized the pivot to Claude Code was available to them. Likewise when one of these companies can gobble up Legal, Medical, etc why would they let companies like Harvey capture the margins?
2) oss models are 6-12 months behind the frontier because of distillation. If labs close their models the gap will widen. Once vertical integration kicks off, the distillation cost becomes higher, and the benefit of opening up generic APIs becomes lower.
I can imagine worlds where things don’t turn out this way, but I think folks are generally underrating the possibilities here.