One sentence reality check

The playbook has always been: get them addicted, dependent, subsidize their costs, lock out competition, and then, when there’s no other option available, milk them for every cent they've got.

A little back story …

Startups have followed a similar pattern since the dawn of the SaaS (Software as a Service) age. They provide an amazing and novel tool to anyone willing to sign up. They make entry simple, straightforward, and once you’re in the product and user experience keeps you coming back for me. It took roughly 20+ years to perfect the model, starting with Salesforce (https://www.bigcommerce.com/blog/history-of-saas/). Of course, there are other, more obscure examples, but Salesforce is what you would consider the first true SaaS. Since then, almost all new software startups have embodied SaaS paradigms, allowing you to use your browser to access services and tools that you would have otherwise had to install on your computer. More importantly, they made payment easy: from freeware to freemium to license payments to subscriptions and pay-per-use. 

The age of AI

When ChatGPT launched in 2022, we were all firmly entrenched in the idea that we could pay for software: QuickBooks, Google Workspace, Adobe Suite, etc., without needing to install a thing. We were comfortable with the fixed monthly costs and tiers. While there were and have been services that allowed you to pay per use, AI is turning that perspective and expectation on its head. No doubt, free tiers have made us adopt an, honestly, very versatile and useful technology at an unprecedented rate. We now use AI to help us draft emails, summarize content, research and without our space, code and ship software. 

Why this is important now

No doubt you’ve heard of or have tried Vibe Coding. If not, then vibe coding is using natural language with an AI LLM (Large Language Model) to build software. Yep, that’s right: plain language conversations allow you to build that quirky idea, launch a new marketing website, or just test out things you’ve been thinking of for the past 10 years. It’s been incredibly easy to do this and even more so today with some of the most powerful and capable models yet. We have our own perspective on vibe coding, which you can find here, here, and here. 

Remember earlier when we said that once adoption is reached, they eventually milk you for everything you’ve got? Well, that step happened much quicker than expected. If you’re using AI for development, the past few months would have shown you that you’re starting to see ever smaller returns in usage for what you pay for. AI is expensive and using it for something as context-heavy (the amount of information needed) is not sustainable when everything is subsidized. Many companies and founders who shifted to vibe coding and AI-first coding have noticed an explosion in their monthly bills due to pay per usage versus flat-rate subscriptions. Even those still on flat rate subscriptions have seen stricter limits. Technology should become cheaper as it matures; however, AI seems to be bucking that trend.

TLDR: If you have depended on vibe coding to get your software to market or to get your development done, then your costs are going to balloon in the coming months, possibly even eclipsing the cost of hiring a human engineer.

How does this affect you?

Quick checklist (yes/no)

• Are you using vibe coding to develop your solution?
• Have you been able to deliver an initial version?
• Have you started struggling with making adjustments, stabilizing, or adding new features?
• Are you aware that there are more than 10 development specific services you can use for your development work?

How can we help?

Can we be honest with you? At NITM, we are at the forefront of experiencing these issues daily. When our bill went from < $1,500 to almost 8k in one month, we realized this was a real problem that was going to make AI even less accessible. Not to mention recent movements by the US government to seemingly restrict certain AI models.

We’ve learned how to diversify our AI usage effectively while keeping costs down to even cheaper than what we had previously, with similar results. We stake this on two tenets: Stronger Harnesses for our AI development and Diversified providers.

A better harness

Harnesses in AI-assisted development and vibe coding refer to the structure you put around your AI development efforts. If you’re simply telling ChatGPT or Claude to “Build, be something amazing”, they will, but it’ll be mostly out of your control. Most coding tools, such as Claude Code, OpenAI Codex, GitHub Copilot, OpenCode, Cursor, and more, have a very strong structure for developing software. Each has its own internal system prompts, guidance, and commands, and more that make up their harness.

Because there are so many options, at NITM we have settled on maintaining and making our own support harness that augments these tools and allows you as much freedom as you need. We use these tools on a daily basis and constantly improve it as we identify gaps, improvements and new ways of working both through our discoveries and other ecosystems.

In real world efforts, we saw one of our clients struggle to understand how to develop what they needed with Claude Code, to being ready to ship their product to production within 1 month and a half of using our toolkit.

Diversification and Cost Management

Did you know that you’re not limited to JUST ChatGPT and Claude for development and planning? In fact, there are many solutions which offer even more competitive price and performance compared to the household names. At NITM we have diversified to use these tools complimentarily in a cohesive toolchain. Here’s how we keep our costs down while getting the most out of our spend.

Before: We spent $750-$1,500/month/developer on AI assisted internal development for our internal projects. Keep in mind that this amount saves us time and allows us to iterate faster.

Now, we recommend something similar to:

There are alternatives to the above, for example, using Claude or Codex $100 plan; however, depending on your velocity and complexity, you may find yourself being rate-limited more frequently. Utilizing cheaper implementation AI models, you get the benefit of well thought out plans and documentation using the strongest Anthropic or OpenAI models, delegate the implementation to more cost-effective models that are actually comparable to the most expensive frontier models, and review it for accuracy and completeness on the GitHub platform.

Conclusion

The world is changing and changing quickly. Being able to manage your costs while actually delivering the results you want is even more of an art with AI. Our recommendation here is to maximize tool diversification and do not get tied to any single ecosystem. At the end of the day these are all tools and we should use them in the way that benefits us the most. If you need help aligning your AI tool usage to your budget and workflows don’t hesitate to reach out to schedule a call:

Additional Resources

• https://ninjasitm.com/posts/ai-early-stage-founders-build-fast-validate-faster 
• https://ninjasitm.com/posts/not-just-developer-building-with-ai-agents-more-than-engineer
• https://ninjasitm.com/posts/ai-assisted-ai-agent-vibe-coding-non-technical-founders 
• https://github.com/ninjasitm/ai-assisted-dev-toolkit 
• https://www.anthropic.com/news/fable-mythos-access 
• Claude Code
• Opencode
• Codex