September 15, 2025: the Day the Industry Admitted AI Subscriptions Don't Work

Github + VSCode will probably eat Kiro's lunch pretty soon with this: https://github.blog/ai-and-ml/generative-ai/spec-driven-deve...

This reminds me of S07E01 of the TV show Black Mirror, which tells a story about a free brain tumor operation, followed by a monthly fee for "remote mind hosting". The standard plan goes worse and worse; your brain needs to serve as a computational resource while sleeping, and they keep pushing you to upgrade to Plus or Lux plans.

I don't understand how tech companies get away with using terms like "unlimited" or "lifetime"

At this point I don't see the point of using a coding tool not provided directly by one of the big foundation model companies. They all clearly see the market of agentic coding tools is one of the first real markets willing to pay, and they are all working on first party tools that are tightly integrated with their models.

Is anyone surprised about this? Unlike pretty much every other fixed-fee subscription, AI has non near zero usage costs, or has usage caps (1Tb storage, etc).

Everything else that has usage costs, storage, compute, etc, is all pay per use (AWS, etc).

Kind of glad that I missed the Cursor hype then rapid slide into enshittification, since I didn't think its goal of replacing an IDE would be a winning strategy and instead went for IDE plugins/extensions AI assistants like Augment Code, GitHub Copilot and Gemini Code Assist as they work in both Rider/Intellij, VS Code and NeoVim.

I'm assuming I don't use AI as much as most devs as I'm still comfortably within Augment Code's 600 user messages/mo dev plan limit which IMO is great value for being able to use its Context Engine. If I ever exceeded that I'd likely temporarily switch to Qwen Coder's effectively unlimited free tier [1] until the next quota refill.

Not sure if Augment's going to increase their prices or downgrade their quotas since they're still at the whims of Claude Sonnet / GPT 5 API pricing, but if they do I'd be looking at moving to an open model solution like Open Code / Roo Code so I can easily switch to the best value coding models of the day, between Qwen3 coder, Grok Code Fast, GLM, Kimi K2, DeepSeek, etc we're spoilt for choice [2].

[1] https://github.com/QwenLM/qwen-code

[2] https://openrouter.ai/rankings#categories

Wish they explicitly introduced Openrouter in the article. Aider (my choice) and Kilocode uses it to connect to any models and switches providers for a given model (for example Sonnet is hosted by Anthropic, Google and AWS and OSS models have even more hosts). I try models 20x cheaper than Sonnet.

I think we should differentiate between the AI industry and "the AI coder" industry.

The problem for most (maybe all) AI coder companies is that they are a fundamentally bad business: they are reselling a commodity (tokens in/out) without a moat around the value they add.

The majority of these companies aren't even adding value, they're simply reselling the the underlying commodity.

I mean, if I were to use an analogy, I'd use coffee crop/beans/grounds/drinks:

The base product in the value chain (the beans) have zero value to the end-user (coffee drinkers). Each step between "coffee farmer" and "coffee drinker" adds a little bit of value to the output of that step.

Roasted beans have more value than raw beans. Ground beans have more value than Roasted beans. Instant coffee has more value than ground coffee. A cup of instant coffee purchased from a street vendor has more value than a bag of instant coffee purchased at the supermarket.

Where, exactly in all of the AI token reselling, is the value being added? A token purchased from Cursor, Kiro, etc has exactly the same value to the end user as a token purchased at the supplier (Anthropic, OpenAI, etc).

How exactly were they planning to monetise, other than by training their own model on every request/response going past them? There's no avenue to value-add here other than by using the actual requests/responses to create a "better" model!

And even then, the "better" model may only be insignificantly better, with almost-zero value add, so how are they going to capture that insignificant value-add?

Looking at all of this, even MBAs, widely denounced for running tech companies into the ground due to sheer incompetence, might actually do better here!

>Welcome to the honest side of AI coding.

How's this different from the same playbook they're accusing other parties to play?

1. The Hook: we're honest!!!!

Come on.

Ad

brave statement the day after Codex CLI got a revamp

Bait and switch is a very old game. And venture capital is not a charity.

Did anyone really believe that companies would burn through billions in funding without trying to recover it and then some?

Any thoughts on how long before Zed Pro starts down this path? Or maybe they already are?

I switched from Cursor to Zed about five months ago. It’s a tradeoff of simplicity vs too many features that I’ve mostly been happy with. The last few days though, the “suggested” sonnet 4 model has gotten really slow.