← sources

We are in a pre-monopoly market and I don’t like it

source: reddit (reddit:ClaudeAI)published 2026-10-11 19:59ingested 2026-10-12 02:36

retained text

Here's my hot take: AI providers should pick a lane. Either they train and serve models, or they build and serve agents. Anthropic, OpenAI, Google, Meta, X: these companies are too vertically integrated, and that integration takes leverage away from users. They build the model, run the inference, and own the harness on top; so when something works brilliantly, it's hard to tell whether the model earned it or the scaffolding did. Most users don't know the difference, and the vendors have little reason to explain it.

Imagine a world where Anthropic and Claude Code were separate, partner teams, but distinct products. Claude Code would be built to be configured, extended, and swapped out piece by piece, while Anthropic would win or lose on the strength of the model alone. Sure, tools like Cursor and Cline already let you switch models but the in-house apps get the best pricing, the newest features, and the stickiest data, so the open option rarely competes on equal footing. If/when one gets bailed out and eats the others, we will be full /monopoly-mode. No others will stand a chance.

Take this week as an example: the recent Projects update (“improves” performance in projects by sharing context across sessions, not a replicable feature, not opt-in, eats context window) has people annoyed, and rightfully so. This is just one example of why I’m against the current way the consumer AI products are built.

Counter argument: they'd never make money that way.  
Rebuttal: the consumer apps aren't where the real money is anyway. The durable revenue comes from enterprise and API licensing, which is where model quality matters most. They should be competing on the model, not on chat UI.

I'll admit a full breakup is a blunt instrument, and the line between model and harness is blurrier than it used to be. Even so, three smaller rules would get us most of the way there:

**No self-preferencing.** A lab shouldn't give its own app better pricing, rate limits, or early model access than it gives everyone building on its API.

**Your stuff leaves with you.** Projects, memory, and history should export cleanly, so switching costs stop doing the competing for them. This should be a law - no gate-keeping my data.

**Neutral scorecards.** Benchmarks should run on a fixed, shared harness; then we can finally see what the model does on its own.

Separate the focus, and both the agent and the model get better. I’d expect this to also help local inference lovers, because agent harnesses would have more focus. Okay, that's all feel free to roast me about why I'm wrong.

document fetched 2026-10-12 00:21 · canonical https://reddit.com/r/ClaudeAI/comments/1x3hvpq/we_are_in_a_premonopoly_market_and_i_dont_like_it