Yesterday Anthropic quietly removed Claude Code from the $20 Pro plan. Then they said it was just a test. Then they put it back. In the span of 24 hours, a lot of developers who vibe code for a living got a reminder that the tools they depend on can change without warning.
Claude Code
I’ll be honest, I was mortified. I have been using Claude Code since last year to vibe code pretty much all the web apps I have built, from TaskCocoon to TBot to F.SKY Clothing. To continue using it meant moving to the $100 Max plan, which would tear a Grand Canyon sized hole in my pocket. Financially speaking. And yes, I am exaggerating a bit, but you know what I mean.
So I started asking the obvious question: what else is out there, and how close is it really?
What Actually Happened
On April 21, 2026, Anthropic’s pricing page was updated to show Claude Code as unavailable on the Pro plan. No announcement. No email. No changelog. Developers noticed because someone compared the archived April 10 version of the support page with the new one and spotted that the title had quietly changed from “Using Claude Code with your Pro or Max plan” to “Using Claude Code with your Max plan.” One word removed. Massive implications.
Anthropic’s head of growth later posted on social media saying it was “a small test on roughly 2% of new prosumer signups” and that existing subscribers were not affected. By the time that clarification landed, the pricing page had already been seen by the entire internet, not 2% of it. They eventually reverted the public changes, but the experiment is apparently still running quietly in the background.
The underlying message was not subtle. Anthropic acknowledged that usage has changed a lot and that their current plans were not built for it. Translation: people are getting a lot of value out of Claude Code at $20 a month, more than Anthropic planned for, and something is going to shift. Maybe not today, but the writing is on the wall.
Why This Matters If You Vibe Code
If you use Claude Code occasionally to help debug a script or generate a quick component, this probably did not rattle you. But if you use it the way I do, as the primary engine behind how you build and ship things, the thought of it moving to $100 a month changes the math entirely.
Claude Code on Sonnet has been my default coding agent for a while now, with occasional switches to Opus when a task is heavy enough to need it. It is not just a chatbot you paste code into. It is a terminal-first agent that reads your files, runs commands, manages git operations, and stays oriented across long multi-file sessions. For the kind of vibe coding I do, it has been the right tool at the right price. Nothing else had given me a strong enough reason to move away from it.
That last part is important context for what comes next.
The Alternatives I Am Returning To
I want to be clear about something. I am not coming at this blind. I have used both Codex and Kimi before, not extensively, but enough to have a baseline feel for them. A few months back I was routing Kimi K2.4 through OpenClaw for light testing, and I used Codex for some smaller builds around the same time. Neither experience gave me a compelling reason to switch away from Claude Code back then because Claude Code was working perfectly fine and both models have moved on significantly since those versions.
That is actually what makes this moment interesting to me. This is not about panic-switching. It is about the fact that these models have had meaningful upgrades since I last tested them seriously, and now I have a real reason to go deeper rather than just dip in and move on.
Here is where each one stands today.
Image courtesy: OpenAi Codex
OpenAI Codex has matured considerably since the version I was routing through OpenClaw. It is now available as a desktop app, a CLI, and a VS Code extension, bundled into ChatGPT Plus at $20 a month with no extra billing. Over two million developers use it weekly. During the Claude Code situation, OpenAI was notably quick to point out that Codex would stay on the free and Plus plans, which was a transparent marketing move but also genuinely useful information. The version I tested months ago was capable for light work. What I want to know now is how it handles the kind of sustained, multi-file, context-heavy sessions I actually rely on.
Kimi K2.6 from Moonshot AI is the one that caught my attention most when I started digging into the current landscape. I tested K2.4 a few months back via OpenClaw and it was promising but not quite there for my workflow. K2.6 launched on April 20, the day before the Claude Code situation blew up, and almost nobody noticed. It scores 80.2% on SWE-Bench Verified compared to Claude Opus at 80.8%, essentially neck and neck on the benchmark that matters most for coding agents, at around 80% cheaper per token. The API is OpenAI compatible, which means it plugs straight into tools like Cline in VS Code, Cursor, and even the Claude Code interface itself by swapping one environment variable. A dedicated Kimi Code CLI also exists for terminal-first workflows. The gap between K2.4 and K2.6 is significant enough that my earlier impressions are not really relevant anymore.
Gemini CLI
Gemini CLI is worth a mention for completeness. Google offers 1,000 requests per day free with Gemini 2.5 Pro and a one million token context window. That is a serious free tier for anyone who wants a low-stakes starting point, and it runs in the terminal in a way that will feel familiar if you are used to Claude Code.
Cline sits slightly differently from the others. It is a free open-source VS Code extension that lets you connect to whichever model you want, including Kimi, Claude via API, or others. It is the bridge that makes most of these alternatives accessible inside an editor without overhauling your entire workflow.
What I Am Actually Testing and Why
This is not just documentation for the blog. I am going deeper into these tools because the Claude Code situation made it clear that depending entirely on one provider at one price point is a risk I should not be comfortable with. Codex and Kimi K2.6 are the two I am prioritising, both because they have had significant upgrades since I last used them seriously and because they fit the same budget I am already on.
Claude Code is still on my Pro plan today and I am still using it. But I am also running real work through these alternatives now, not light testing, actual builds. I will share what I find as I go.
FAQ
Is Claude Code still available on the $20 Pro plan? As of today, yes. Anthropic reverted the pricing page changes and confirmed existing Pro subscribers are not affected by the test. That said, they have signalled that the current plan structure was not designed for how people are actually using Claude Code. Worth having a backup plan.
How does Kimi K2.6 compare to the earlier versions? Significantly better, at least on paper. K2.6 scores 80.2% on SWE-Bench Verified, which is a big jump from where K2.4 sat. The architecture improvements around long-horizon coding and agent reliability are the areas I am most interested in testing against real work.
Can I use these alternatives inside VS Code without changing my whole setup? Yes. Cline is a free VS Code extension that connects to almost any model via API, including Kimi K2.6. Codex has its own VS Code extension. Neither requires you to overhaul your workflow to start testing seriously.
The Takeaway
The Claude Code situation this week was messy and Anthropic handled the communication poorly. But it did something useful. It pushed me to go back to tools I had tested lightly before and take them seriously now that the models have moved on and the stakes feel more real.
I am not switching anything yet. But I am testing properly this time, not just dipping in. More to come as I work through it.