November 11, 2025
I pay OpenAI, Anthropic, and Google directly for API access. No ChatGPT Plus. No Claude Pro. No Gemini Advanced. Just API keys and a client app that lets me use them.
People ask why I bother. Isn't it more complicated? Don't the consumer apps just work?
Here's the thing: consumer AI apps are optimized for the provider's revenue, not your productivity. The subscription model creates incentives that work against you as a user. Once you understand this, you can't unsee it.
Here are 5 reasons I use direct API access instead of consumer subscriptions.
1. You Pay for What You Use (Not What They Think You Might Use)
ChatGPT Plus costs $20/month. Claude Pro costs $20/month. Gemini Advanced costs $20/month.
That's $60/month if you want access to all three flagship models.
With API keys, I typically spend $5-15/month total across all three providers. And I use AI constantly. The math is simple: API pricing is based on actual token usage, not a flat fee designed to capture maximum revenue from light users.
The subscription model works like a gym membership. Providers profit most from people who pay but don't show up. They're betting you won't use enough AI to justify the subscription. If you're a light user, you're subsidizing heavy users. If you're a heavy user, you're hitting usage limits. Either way, you lose.
Pay-per-use pricing flips this. Use a lot? Pay more. Use a little? Pay less. No subsidizing anyone. No artificial limits to protect margins.
2. No Secret Model Routing
Here's something most people don't know: consumer AI apps don't always give you the model you think you're getting.
ChatGPT Plus claims you're using GPT-4o or whatever their current flagship is. But during peak times, or when you've used "too much," they quietly route you to cheaper, faster models. The interface looks the same. The responses get worse. You have no idea it's happening.
This isn't a bug. It's a feature. Subscription margins depend on managing server costs. If everyone who paid $20/month actually used flagship models for every request, the providers would lose money. So they route strategically.
With direct API access, you pick the model. You get that model. GPT-4o means GPT-4o. Claude Sonnet 4.5 means Claude Sonnet 4.5. No bait and switch. No mystery routing. You can even see the exact model in the API response metadata.
3. No Arbitrary Usage Limits
ChatGPT Plus limits you to a certain number of messages per time period with their best models. Claude Pro has similar limits. These limits change without notice based on "demand."
These aren't technical limitations. Server capacity isn't the issue. These are business decisions to protect subscription margins.
Think about it: if you could use unlimited flagship models for $20/month, power users would cost the provider far more than they pay. The limits exist to cap your value while keeping your subscription.
API access has no message limits. You pay per token, so there's no reason to restrict usage. Process 100 documents in an hour? Fine. Have a 500-message conversation? Go ahead. The only limit is your budget, and you control that.
4. Full Technical Control
Consumer apps hide parameters to keep things "simple." You can't adjust temperature (how creative vs. consistent the model is). You can't set max tokens. You can't tune frequency penalties. You get whatever defaults they decided were good enough for everyone.
This matters more than you might think. Different tasks need different settings:
- Creative writing benefits from higher temperature
- Code generation needs lower temperature for consistency
- Long-form content needs higher max tokens
- Research tasks benefit from adjusted penalties to reduce repetition
With API access, you control all of this. Same control developers have. You're not treated like someone who can't handle "advanced" settings.
Cumbersome exposes these parameters in a clean interface. Adjust temperature mid-conversation. Set system prompts. View raw API responses. Full transparency about what's actually happening.
5. One Fewer Party Seeing Your Data
Let's be clear: OpenAI, Anthropic, and Google still process your data when you use their APIs. Your prompts hit their servers. Their systems see your content. This is unavoidable with any cloud AI.
But with consumer apps, there's an additional layer. The app provider (ChatGPT, Claude.ai, Gemini) stores your conversation history, manages your account, and builds features around your usage patterns. That's two parties with access to your data instead of one.
With API keys in an app like Cumbersome, your conversations sync via your personal iCloud account, not through any third-party servers. The app developer never sees your data. You're still trusting the AI provider, but you've eliminated the middleman from the privacy equation.
It's not perfect privacy. But it's one fewer company with access to everything you type.
The Tradeoff: Rate Limits (That Fade Over Time)
I should be honest about the downsides. API access isn't all upside.
Each provider has rate limits on their APIs. These limit how many requests you can make per minute or how many tokens you can process per day. New accounts start with aggressive limits.
Here's the key difference from consumer app limits: API rate limits loosen over time. As you use the API and spend money, providers automatically increase your limits.
Google's tier system for API rate limits. Spend $250 and wait 30 days, you get Tier 2. Spend $1,000, you get Tier 3. Limits scale dramatically at each tier. See Google's rate limit docs for current details.
OpenAI's tier system works similarly. Higher spend and account age unlock higher limits. See OpenAI's rate limit docs.
Anthropic follows the same pattern. $5 gets you Tier 1, $40 gets Tier 2, and so on. See Anthropic's rate limit docs.
After a few weeks of normal usage, most individual users have limits high enough that they never hit them.
Consumer app limits work the opposite way. They exist permanently to protect subscription margins. You don't "earn" higher limits by being a good customer. You hit the same ceiling whether you've been paying for a month or a year.
So yes, your first week with API access might feel constrained if you're a heavy user. But the limits fade. Consumer app limits don't.
For more on navigating this, see the detailed guide to API access.
The Real Question: Who Are Consumer Apps Built For?
Consumer AI apps are built to maximize average revenue per user. Every design decision flows from this goal:
- Subscription pricing captures value from light users
- Usage limits protect margins from heavy users
- Hidden model routing reduces costs without users noticing
- Simple interfaces prevent users from accessing cheaper models directly
- Ecosystem lock-in (conversation history, integrations) increases switching costs
None of this is about giving you the best AI experience. It's about extracting maximum revenue while giving minimum access.
API keys flip the incentive structure. Providers want you to use more because you pay per token. There's no reason to limit you, route you to cheaper models, or hide capabilities. Your interests and theirs are aligned.
Getting Started Takes 10 Minutes
The main objection I hear: "Isn't this complicated?"
Not really. Here's what you do:
- Create developer accounts at OpenAI, Anthropic, and Google AI Studio
- Add a payment method to each
- Generate API keys
- Add them to an app like Cumbersome
That's it. Ten minutes of setup for permanently better AI access. The detailed walkthrough covers every step.
Who Should Stick With Subscriptions?
To be fair, consumer apps make sense for some people:
- You genuinely don't want to think about any of this
- You need specific features only available in consumer apps (like ChatGPT's browsing or plugins)
- Your employer pays for subscriptions anyway
But if you're paying your own money for AI access and you care about getting the most value, API keys are objectively better. Less money. Better models. No limits. Real privacy. Full control.
Try It
Cumbersome is free for iPhone, iPad, and Mac. Add your API keys and start paying providers directly instead of through subscription markup.
Or keep paying $20-60/month for artificial limits and mystery model routing. Your call.
Bless up! 🙏✨