Why You Should Be Paying for AI Subscriptions Right Now
Many people are afraid to spend $20 to $40 per month on AI subscriptions. My perspective is that for anyone working in tech, if you're not going to try these tools now, the cost of getting in later will be far too high.
This is actually backed by research you can find online. Right now, an average Pro user is paying around $20 per month for a subscription, but the actual cost that Anthropic has to absorb is estimated at anywhere from $100 to over $2,000 in computational costs per heavy user, depending on usage intensity. This is the same playbook many companies like Uber used in the past: get you in cheap, let you build your workflows around the product, then gradually increase prices as dependency grows.
If you don't try it now, the cost to get in might simply be too much later. Right now, while things are still relatively affordable, you can experiment freely. Try it for a month, switch to another platform, and figure out what actually works for you. Eventually, we'll be shifting toward local AI and local models. But until then, you never know when these subscriptions could reach hundreds or even thousands of dollars per month once Anthropic stops subsidizing the bill for all of our AI usage.
Why Claude Code Offers Great Value Right Now
We obviously know that Claude Code is a separate platform from Claude's text-based AI chat. Compared to all the other tools I've personally used like GitHub Copilot, Codex, Gemini CLI, Cursor, and others, all of these platforms are genuinely good and get the job done. However, what sets Claude Code apart is that it has built an infrastructure where you can give it custom skills and custom agent protocols. It runs in your CLI, and now that the Claude Code desktop app has launched, it comes with some really great features like remote sessions. You can kick off a session on one of your GitHub branches from your phone, and by the time you get home, the app is already scaffolded for you.
A recent example for me was video editing. I gave Claude Code a skill from Remotion, then simply uploaded a video. It built the entire workflow, edited the video, and handled the animations. The friction involved in creating videos or content is so much lower now, and that's exactly the kind of versatility Claude Code offers.
When to Use Codex vs. Claude Code
Platforms like Codex are excellent for deep problem-solving or code reviews, since those models are specialized for exactly those tasks. But if you're looking for a general-purpose solution with strong CLI integration, Claude Code is the better option. If the pricing stays where it is, it's worth maintaining subscriptions to both Codex and ChatGPT while you can still afford to experiment. Eventually, when costs become excessive, you can cut whichever one isn't pulling its weight.
What You're Actually Getting with Claude Code
I actually had Claude Code do research on its own capabilities, and here is what it found. As of March 2026, the infrastructure they've built goes way beyond just chatting with an AI. Here's a quick breakdown:
| Feature | What You Get |
|---|---|
| Custom Skills & Commands | Drop markdown files into your project and Claude auto-loads them as reusable skills. Slash commands for repetitive workflows. |
| MCP Servers | Connect Claude to any external tool — Slack, GitHub, databases, internal APIs. 300+ integrations available right now. |
| Hooks System | Run custom shell commands before or after Claude takes any action. Block dangerous operations, auto-run tests, enforce rules. |
| Remote Sessions | Kick off a session from your phone, come back to a finished scaffold on your machine. Headless mode for CI/CD pipelines. |
| Agent SDK | Build your own custom agents using the same infrastructure Claude Code runs on. Available in Python and TypeScript. |
| CLAUDE.md Memory | Project-level memory files that persist across sessions. Claude remembers your stack, conventions, and architecture. |
| IDE Integrations | Native extensions for VS Code and JetBrains. Sees your open files, highlighted code, and error panels automatically. |
| Multi-Agent Orchestration | Spin up parallel sub-agents that work independently on different parts of your codebase simultaneously. |
| Git Worktree Support | Isolated worktrees per session so multiple agents can work on the same repo without conflicts. |
| Sandboxed Execution | OS-level sandboxing on macOS and Linux. Docker microVMs for full isolation. Run agents unsupervised without risk. |
Let me break some of these down because the table alone doesn't do them justice.
Custom Skills are probably the feature that sold me. You literally drop a markdown file into your project, describe what you want Claude to know or do, and it just picks it up. No configuration, no API calls, no plugins to install. I gave it a Remotion skill and suddenly it could edit videos. That's the kind of extensibility most platforms charge extra for.
MCP is the one most people sleep on. Model Context Protocol lets you connect Claude to basically anything — your Slack, your database, your internal tools. The ecosystem has exploded to over 300 integrations, and it's an open standard so it's not locked to Anthropic. You're building on something portable.
Hooks give you real control. You can set up rules like "block any command that touches production" or "auto-run tests after every file edit." It's the difference between trusting an AI blindly and actually governing what it does in your codebase.
Remote Sessions and Headless Mode are where it gets wild. You can literally start a coding session from your phone on the train, tell it what to build, and by the time you sit down at your desk, the work is done. Or pipe it into your CI/CD pipeline so it handles code reviews automatically.
Multi-Agent Orchestration means you're not limited to one AI doing one thing at a time. You can have a lead agent coordinating work while sub-agents handle independent tasks in parallel — each with their own context window, their own worktree, doing actual concurrent work. This is what separates it from just being a chat interface.
The point isn't that every other tool is bad. Copilot, Cursor, Codex — they all have strengths. But none of them have built out this kind of infrastructure layer where the tool itself becomes a platform you can extend, automate, and orchestrate. That's the value proposition, and at current pricing, it's genuinely underpriced for what you're getting.