I Spent 3 Hours Watching AI Rewrite My Code. What I Found Made Me Rethink Everything.

Here’s what kept me up last night.

I dropped a massive code refactor into Claude Opus 5. Three hours later, the agent had finished. I felt good. K3 and GPT-5.6 Sol gave it a clean bill of health. Clean code. No bugs. I was ready to ship.

Then I ran it through Opus 5 one more time, just to be sure.

My blood ran cold.

I found a vulnerability that would have cascaded through my entire system. A cache penetration attack vector that would have let someone drain my server dry. I have PTSD from seeing this exact pattern two weeks ago, when my bill got nuked by a coordinated attack.

This is the terrifying promise of the new AI era. The models are getting smarter. Scary smart. But they’re also revealing a hard truth: the bottleneck isn’t the model’s intelligence anymore. It’s the chaotic, contradictory environment we force it to operate inside.

—

Anthropic just dropped Claude Opus 5. The benchmarks are insane. On ARC-AGI-3, the gold standard for abstract reasoning and novel task generalization, Opus 5 scored 30.2%. GPT-5.6 Sol got 7.8%. Opus 4.8 got 1.5%. Opus 5 is nearly four times better than the next closest competitor.

But here’s the thing that matters: it costs half the price of Fable 5. $5 per million input tokens, $25 per million output tokens. Same price as Opus 4.8. You can even enable a Fast Mode that runs 2.5x faster for double the price.

Is it a Fable 5 killer? No. Not yet. But it’s the first real middle ground that actually works. I’ve been running 50% Fable 5 + 50% Opus 5, and the results are shockingly good.

—

I rewrote my entire API, RSS, and Skill system in three hours. The old version was rushed. It sucked. I needed to fix the user experience and reduce server load.

My Fable 5 quota was gone. So I did something hacky. I used the web version of GPT-5.6 Sol Pro — a trick that gives you deeper review depth without burning your Codex quota. Then I threw the solution at GPT-5.6 Sol, Kimi K3, and Opus 4.8 for deep review. All three said: looks good. Minor issues at best.

Then I set GPT-5.6 Sol Ultra loose with a goal: “Fix this. Make it right.” Three hours. I don’t know how many agents it spawned. It finished.

K3 and GPT-5.6 Sol did adversarial review. Found a few minor bugs. I thought: we’re good.

That’s when Opus 5 found the thing that would have broken my entire system.

—

This is the paradox of the new AI era. As models get exponentially smarter, your job as a developer doesn’t get easier. It gets harder in a different way.

Anthropic’s own research team identified six shifts that define the Claude 5 era. Read them carefully:

• Past: Give rules. Now: Give judgment space.
• Past: Give tons of examples. Now: Design clean interfaces.
• Past: Stuff everything upfront. Now: Unfold as needed.
• Past: Repeat reminders everywhere. Now: Write simple tool descriptions.
• Past: Cram memory into CLAUDE.md. Now: Let it auto-memorize.
• Past: Write a spec. Now: Provide code, tests, HTML prototypes, and full reference material.

Do you see the pattern? We’re moving from writing manuals to designing environments. You’re no longer a prompt engineer. You’re a harness engineer.

A great harness builds state, permissions, constraints, and feedback into the environment itself. You leave the goals and judgment to the model. Stop telling it what to do. Start designing the conditions under which it can figure it out.

—

Most developers are still over-prompting. They’re cluttering context with redundant instructions, conflicting rules, and memory dumps. Here’s what happens: the System Prompt says “keep relevant documentation.” The Skill says “never add comments.” The user says “match the old version.” The model can understand all of this. But it has to spend precious thinking tokens resolving the conflict. Which rule wins? What does “old version” even mean?

The longer your context, the more of these micro-conflicts pile up. The model gets slower. Dumber. More prone to catastrophic errors.

Anthropic is so aware of this problem that they recommend running /doctor in Claude Code to clear out garbage. The best thing you can do for your AI agent right now is to simplify its workspace.

I’ve been running almost bare. A few Skills. Some MCPs. Official plugins. Standardized docs. That’s it. My experience has been exceptional because I haven’t suffocated the model with noise.

—

I also tested Opus 5 on UI design. Claude Design + Opus 5 turned out a version that was actually normal. The previous version from GPT-5.6 Sol was so bad I wanted to scream. Opus 4.8 was better but not great. Opus 5 finally got it right.

Writing? Don’t bother. Opus 4.6 is still the king of prose on Claude.

—

Here’s my current workflow, as of today:

• Large-scale refactoring planning: Claude Fable 5
• Design + Review: Claude Opus 5
• Code execution + Bug fixing: Opus 5 or GPT-5.6 Sol (Claude if I’m at my desk, GPT if I’m mobile)
• Large code review: GPT-5.6 Sol Pro on web
• UI design: Claude Design + Opus 5
• Frontend quick builds: Kimi K3

This list will rotate faster than you think. The pace of model evolution has crossed some kind of singularity. Grok acquired Cursor and their capabilities jumped overnight. We’re looking at bi-weekly updates. Every morning, there’s a new model. Every day, our capability boundaries expand.

There is no more exciting time to be building than right now. But you need to adapt. Simplify your agent’s environment. Stop writing rules. Start designing harnesses. The models are ready. The question is: is your workspace?