Best Claude Code Alternatives: What to Use Now That Anthropic Nerfed Opus

Anthropic has a long history of being anti-consumer and a generally misanthropic company that has produced some of the most misaligned models to date. But their recent steps have gone way too far.

I've been an Anthropic Max user for a while. I've been using the Claude Max $200 or $100 plan depending on how heavily I'm using it that month and overall have really enjoyed my experience. Recently it's become next to unusable. A couple of weeks ago Anthropic quietly slashed the rate limits so aggressively that tasks I used to breeze through were just no longer possible. Of course they went on Twitter and gaslighted us about it, claiming they didn't change anything, or that it was just minor adjustments that would only affect the very highest usage users. Completely false.

Now recently they've nerfed the model Opus within Claude Code, as well as generally deteriorated the performance of the Claude Code harness to the point where it's next to unusable. I'm asking it to do basic tasks, giving it simple instructions, and it's just unable to follow them. That leaves me wondering: why am I paying $100 a month for something that barely works? Not to mention their constant outages and downtime.

So that leads to the alternatives, right? What am I going to use instead? I mean I've been a Claude Code power user for a while. I do love it but it's time to seriously take a look at the alternatives.

Codex from OpenAI: The Obvious Competitor

The most direct replacement is Codex from OpenAI. I really like the Codex app. It's polished, it actually works (unlike the Claude desktop app), and the GPT models behind it do something Claude consistently fails at: they search your codebase and build things in a sustainable, forward-thinking way. Claude Code and Opus just output slop. Features that occasionally work but pile on so much tech debt you can't even imagine.

That said, GPT models are not perfect. They're strange to work with. They require very specific prompting styles. They talk in this bizarre corporate tone and have an infuriating usage of the English language. And whatever you do, do not use them for UI work. The kind of UI GPT produces is offensive to look at. It uses internal technical-only terms as marketing jargon in the actual interface. Just mind-blowingly bad. But for backend work? Great.

OpenCode: Open Source and Almost Great

OpenCode is one of the most interesting free Claude Code alternatives out there. It's open source and built on top of a TUI (terminal user interface), which lets them do a lot of GUI-like features right in the terminal. It feels modern and capable.

Unfortunately, the scrolling within the TUI is really bad. You can scroll, technically, but it feels terrible and is always wrong. Every time I scroll up through output, the experience is so janky that I find myself not wanting to use it as much. That's a shame because the actual AI capabilities are solid. If they fix the scrolling, this becomes a top-tier option overnight.

Droid from Factory: Solid with Some Quirks

Droid from Factory is another strong contender. It's got a bring-your-own-key setup that's a little annoying to configure but really not that bad once you get through it. Overall it's solid.

The standout feature is Droid Missions, which are basically long-running background tasks. You kick off a mission, it works on it asynchronously, and you come back to results. If you're doing the kind of work where you want to set something going and check on it later, Droid handles that well.

The Secret Weapon: FirePass from Fireworks AI

Here's what I've actually been using the most. The FirePass from Fireworks AI is $7 a week for unlimited developer tokens through Kimi K2 Turbo. Seven dollars. I've put millions of tokens through this thing and the model is inferencing so fast it's kind of game-changing.

The best part: it supports bring-your-own-key. I plugged it into Droid, OpenCode, and even Hermes from Nous Research (which is a competitor to OpenClaw, and it actually works). Unlimited fast tokens across every tool in my stack for the cost of a coffee.

I even tried using it in Claude Code. Because Claude Code is such a sloppy piece of software built by the slop engineers over at Anthropic using their sloppy AI model Opus, it kind of works but doesn't really. It auto-completes after one turn and then stops. Classic Anthropic quality.

The Magic That Keeps Me Coming Back (For Now)

Something you'll notice across all these alternatives: they're all technically better than Claude Code. Claude Code is so buggy and so obviously vibe-coded. It's sloppy. But there is a certain magic to it that keeps me coming back.

Even though it's so poorly made that it actively hurts code quality, even though it bugs out frequently, it's genuinely a joy to use when it works. The UX, the personality of the model, the way it used to understand what you meant even when you weren't precise with your instructions. That last part is past tense now though. Since the nerf, the model needs way more babysitting and handholding than it used to. I wrote about this tension in my Claude Code vs Codex vs Cursor comparison.

I don't use any other Anthropic software though. The Claude desktop app is much too buggy. Same with Claude Cowork. It's the CLI or nothing.

What You Should Actually Do

Don't get too addicted to cheap tokens or any single harness. I mean it. You never know when a company is going to pull the plug or decide to destroy their own model for apparently no other reason than anti-user sentiment. This is exactly what Anthropic just did.

Always be experimenting with new AI coding tools. Things move so quickly that getting stuck on one tool is a liability, not a strategy. If you want to set up your projects so they work well across multiple AI agents (and you should), I covered exactly how to do that in How I Structure Projects for AI Agent Collaboration.

My current stack right now: Claude Code when it cooperates, Codex for the heavy lifting, FirePass tokens piped into OpenCode and Droid for everything else. That's four tools, and I'm probably going to add a fifth next week because that's how fast this space moves.

The real lesson from the Opus nerf isn't that Anthropic is bad (though they are). It's that depending on any single AI provider is a risk. Diversify your tools the same way you'd diversify anything else. Build your project architecture so any agent can pick up the work, not just Claude.

Frequently Asked Questions