With Gemini 3.1 Flash-Lite launching today, my mind goes towards thinking about things I wouldn't have thought to build because of expense.
Howeve...
For further actions, you may consider blocking this person and/or reporting abuse
What @heckno said, and then expand it to any (all!?) hardware devices that I own.
The bigger question is: will it indeed be in the end/soon that SaaS will go away and everyone (almost everyone?) will build the tool that they need specifically tailored for them inhouse?
Oh yeah good call
I don’t think SaaS disappears.
Cheaper AI makes building more accessible, yes. But building something tailored in-house still requires architecture, long-term maintenance, security, evolution, and governance. AI can reduce implementation effort, but it doesn’t remove structural complexity.
In the short term, more teams might build internal tools. In the long term, the real constraint won’t be model cost — it will be software engineering capacity and organizational maturity.
SaaS solves that by externalizing complexity. That doesn’t go away just because tokens are cheaper.
I agree that it won't be 100%. No way, of course. But what % do you think it actually will? I'm thinking (boldly even) that the split may actually be as big as 50/50.
I think there might be quite a few disappointments.
AI makes it easier to build something quickly — even to design something that looks convincing. But without solid structure, that often doesn’t hold up over time.
Building is one thing. Maintaining, evolving, securing, scaling, documenting, and integrating into real workflows is another. AI reduces friction at the start, but it doesn’t remove long-term responsibility.
So I’m not sure the split will be 50/50. I suspect many teams will try building in-house, then rediscover why SaaS exists in the first place.
I'm curious about stuff that can plug into my smart watch data where I can pipe in basic questions about the trends and get answers. It's not complicated stuff but I want to be able to do in high volume
I don't want to build anything when a hobby that was free is now a subscription. I'll probably use local models when we get better tooling, but the way the world is going, that might not happen. The end game appears to be "your entire computer is a subscription", i.e. thin clients coming back round with everything offloaded to data centres, at which point I'm just not interested any more. I'll get a new interest. Maybe I'll learn to love kayaking.
Lower costs definitely broaden the perspective. But it’s also worth asking whether there’s a real need in the first place — or whether we’re creating one just to take advantage of cheaper AI.
There is many factors going into this:
It's something I keep in mind as new models are being published and accessible for us to use. Of course, the main thing about AI is the environment factor that goes into it because you have to build a lot of Data centers to compute your model. I saw a video where it explains that OpenAI just simply "Made it bigger" and the LLM model just became smarter. It's pretty much the whole reason why there is so many data centers and RAM that is needed just to simply use AI. I might be wrong, but let me know!
Thanks for sharing @ben!
Cheaper inference shifts the constraint rather than removing it. Features I was calling the model once for because five calls was expensive now I call five times. parallel validation passes, redundant retrieval, cross-checking outputs against each other. The architecture changes when the per-call cost stops being the binding constraint.
What doesn't change. whether the use case produces real value. The most interesting builds with cheap inference aren't the ones that were blocked by cost — they're the ones that require volume to work at all. Anomaly detection across thousands of data points. Real-time consistency checking on long documents. evaluation pipelines that run on every output rather than sampling.
The unlock isn't just cheaper features. it's features that only make sense at volume. those were architecturally impossible before not because of cost but because the use case required scale that cost made unviable.
I had a slightly different take on this recently and it was in the corporate setting, do people actually know or pay attention to tokens or other expense calculations on their AI use. Is it obfuscated?
I've noticed the people who consider expenses with models tend to be the same people who tinker and likely used a personal card at some point. Others openly mention they work for
insert big company hereand have essentially unlimited spending on certain tools and models.The cost drop changes what's worth automating more than what's possible.
I run an AI agent on a Mac Mini that handles side project management — content publishing, community engagement, email triage, analytics. Right now the monthly API cost is my main constraint on how many cron cycles I can run.
With cheaper models, the calculus shifts: instead of batching 3 tasks per agent run to save tokens, each task gets its own dedicated cycle. The quality ceiling goes up because you can afford more retries, more validation passes, more "think twice before acting" prompts.
The thing I'd build that doesn't make economic sense today: continuous A/B testing of content. Right now I publish an article and wait 72 hours to measure. With cheap enough models, I could generate 5 variations, test headlines against each other in near-real-time, and iterate within hours instead of days.
Basically: cheaper models don't just make existing features cheaper — they make iteration loops shorter. And shorter loops compound fast.
Interesting question. When AI models get cheaper, I think the biggest change isn't just cost savings but experimentation speed. Developers can test more ideas, run more iterations, and build features that require high-volume AI calls without worrying about token costs.
At the same time, cheaper models don't remove the real challenges of building — architecture, security, maintenance, and creating real user value still matter. In many cases, the advantage will go to builders who combine AI capabilities with strong product thinking rather than just using AI because it’s cheap.
I'm curious whether this shift will lead to more micro-tools and niche products rather than fewer large SaaS platforms.
Great topic for discussion! I think you're right, speed of feature development will continue to increase, but we will always be constrained by the feedback loop. If we don't pay attention to what features we build actually solve problems that people need solving then we're doomed to fail.
The thing that gets cheaper for me isn't the inference it's the feedback loop. When a call costs $0.001 instead of $0.01, I stop batching validation calls and start running them inline. That changes architecture more than it changes features. The decisions that felt like over-engineering six months ago now just feel like correct defaults.
I think cheaper models don’t just reduce cost; they change product design.
When inference is expensive, AI gets used for “special moments” in a product. When it’s cheap, it becomes part of the default interaction layer.
That shift unlocks ideas that were previously impractical: background reasoning, continuous agents, auto-generated UI, real-time personalization, etc.
So the intriguing question isn’t just what AI can do, but what becomes possible when AI is cheap enough to run everywhere in the product.
Hi Ben,
We are excited to see genAI assistants evolve and get cheaper. We are offering our extension, Etiq AI, for MLE or Data Scientists to visualise and improve their code in exchange for some feedback. It's compatible with all code AI assistants on Jupyter or VS Code. Let me know if you're interested
This is an interesting question because cheaper AI models don’t just reduce costs, they expand the space of what becomes practical to build.
When inference becomes inexpensive, ideas that once felt too costly or experimental suddenly become viable features, workflows, or entire products. However, while lower costs unlock creativity and experimentation, the real differentiator will still be product thinking, integration into real user problems, and long-term maintainability.
Cheaper models may democratize the ability to build, but turning those capabilities into reliable, useful systems will continue to require strong engineering and thoughtful design.
If it would be all about "just build it and they come" we would venture into a bright future, but there is so much more to it ^^
We do not use SaaS, Apps, Programs necessarily because we need exactly them, because who needs the whole of Excel or Spreadsheet ;) But we use them, because we are lazy. Ok, I am at least confidently lazy.
Could I build by own "Airtable" or whatever? Sure. Even adapt it to my companies needs. The last 20% would be great. But its not only building it. Hosting it, maintaining it, upgrading, bugfixing... and thats only internal. If I think about outsourcing it, selling it, legal questions, payment questions, tax questions and so many more questions arise.
Claude, Codex, Gemini etc. will eventually become really good coders. And probably in the next years to come out of Google DeepMind maybe a new, better paper will be released and coding will become even better with agentic AI coding helpers.
But building is not the issue here. I mentioned to a friend and colleague: call me back when Claude sets up my Apple and Google accounts, connects firebase with all necessary accounts, Sign-in/Auth hooks, callbacks, deals with Apple reviewers, legal contracts and so on so forth... :) Not only probabilistically, but confidently!
Damn, I would be happy already if would manage my hosting setup without killing all containers and pods and screwing up every second time ^^
Building is already "cheap". The question to pose is: what comes next?
Great point - when AI models get cheaper it lowers the barriers that have been putting many of us off (you can even see it as a form of 'gatekeeping', the haves vs the have-nots) ...
The only thing I'm wondering about is whether it's sustainable as a business model for the AI companies - well, obviously Google has heaps of cash, so they can sort of "subsidize" it - or it's the models (training and inference) really getting more efficient (that's what we really want, of course) ...
Fishing for ideas :D
I'm not not fishing
Some people have said that in two years we will have our own private LLM in our own computer...
What, exactly, makes you think they will get cheaper?