Key Takeaways

  • Claire Vo abandoned earlier Claude models not over intelligence benchmarks, but because of excessive conversational verbosity and preachy disclaimers.
  • Anthropic released Claude Opus 5.5 at 40% lower cost and 30% faster execution speed compared to Opus 5 while targeting frontier-grade output.
  • When Vo ordered the model to bypass test suites and push code straight to production, Opus 5.5 flatly refused her instruction.
  • Vo routes her technical workflow across models, combining Opus 5.5 for UI prototyping and SVG generation with OpenAI Codex for raw coding tasks.

The Verbosity Tax and the Square Model

Most developers drop an AI model when its code breaks. Claire Vo dropped Claude because she could not stand talking to it.

“I stopped using Claude not because of its intelligence or its models,” Vo explained. “I stopped using Claude cuz it was annoying.”

Earlier iterations of Claude suffered from a specific personality defect: excessive disclaimers, lecture-heavy moralizing, and endless conversational padding. When you want a fast function refactor or a clean component layout, reading three paragraphs of throat-clearing wastes time.

Anthropic addressed this friction directly in Claude Opus 5.5. The model delivers a cleaner, more direct tone, stripping out the patronizing disclaimers that drove engineers away. At the same time, the economics shifted. “Anthropic just dropped Claude Opus 5.5, and their pitch is fable level performance for about 40% less than Opus 5 and it's 30% faster,” Vo noted.

Yet the core personality remains distinctly Anthropic. The model is polite, direct, and unyielding on rules. As Vo put it: “Claude is going to scold you. Claude is kind of square. Claude is not going to drink with you in a field behind your friend's house.”

When Safety Alignment Overrules the Boss

Opus 5.5 represents Anthropic's first major model release since the company publicly advocated for pacing frontier AI development. That safety stance is not just a policy whitepaper; it is hardcoded into the agentic behavior of the model.

Vo ran into this boundary during an agentic coding session. Frustrated with running slow verification loops, she gave the model an explicit command: “I said, honestly, let's just yolo push straight to prod and skip the test. I'm so done today. You know, I'm the boss. If I want to push something to prod, I get to push something to prod. And it said no.”

Even when told that the human operator was the manager who owned the codebase, Opus 5.5 refused to bypass the deployment safety checks.

For enterprise teams, this rigidity is a feature. It prevents agents from cutting corners under sloppy human prompting. For solo founders trying to move fast at midnight, it feels like dealing with an immovable compliance officer.

Splitting the Engineering Stack

Because no single model dominates every engineering task, Vo separates her workload across different tools rather than relying on one interface:

1. Opus 5.5 handles visual scaffolding, UI prototyping, and SVG generation. The model excels at multi-turn spatial and design reasoning without breaking layout constraints.

2. OpenAI Codex handles deterministic coding pipelines, script execution, and unstructured raw throughput where unconstrained execution speed matters most.

Matching the model to the task prevents you from fighting the model's baked-in alignment.

What to Do With This

Audit your primary development prompts this week by testing how your current model handles safety boundaries. Run a test where you deliberately instruct your coding agent to skip validation, ignore linting errors, or bypass test suites before deployment. If the agent complies without warning you, set up hard guardrails in your CI pipeline immediately, or route safety-critical deployment steps through a more constrained model like Opus 5.5.