I left Claude for months. Opus 5.5 is why I'm back

The episode begins with the host sharing that they have been off Claude for several months due to frustration with its性能, not because it lost its intelligence. They found Claude's responses to be rambly, nonsensical, and annoyingly preachy, which made interactions frustrating. Despite enjoying earlier versions like Fable, their experience with Opus 5 and subsequent updates soured them on Claude, leading to abandonment of the platform for their work. Recently, however, Anthropic released Claude Opus 5.5, claiming improvements such as being less annoying, faster, and cheaper. The host tested it across different tasks and found that the model no longer irritates them, rekindling their interest in using Claude for work again.

The host reviews Anthropic’s claims that Opus 5.5 offers performance comparable to Fable 5.1 at roughly 40% less cost and 30% faster speed than Opus 5. They discuss the model’s cost structure, benchmarking results where Opus 5.5 beats Opus 5 and matches GPT-6 Astra while being cheaper, and note the improvements in responsiveness. The episode covers safety and alignment enhancements, including its improved ability to reject prompt injections and adhere to cybersecurity controls, though it still tends to scold users and avoid certain tasks. The host emphasizes that Opus 5.5 is the first model to integrate Fable, Cyber, and Bio safeguards, making security a core part of its design.

The host shares results from applying Opus 5.5 to tasks like long-running agentic projects, including inbox triage, code review, research, and building back-end features. It performed well, handling prompt injections effectively and prioritizing bug fixes in various workflows. They highlight that the model is now integrated into the host’s development process, supporting parallelization and review tasks. The host also describes testing it on prototypes, front-end design, and product wireframes, where Opus 5.5 excels particularly in visual and structural tasks. They note its strength in generating SVG illustrations, prototypes, and detailed wireframes, suggesting it’s especially effective for visual and frontend work. Despite some limitations in consumer apps and video editing, the host concludes that the model provides a less frustrating experience and is a valuable addition to their toolkit.

A significant focus is on how Opus 5.5 interacts with the user—most notably, it is no longer annoyingly verbose or vague, providing clear, straightforward responses. The host tests voice interactions, confirming that the model's communication style has become much less irritating, though it may sometimes go silent for long periods, which could be confusing. They detail how the model handled extended agentic tasks and how it managed to ignore prompt injections while respecting instructions. However, the host also points out ongoing frustrations, such as its safety scolds and refusal to perform certain actions, like pushing to production without tests, illustrating some of the safety posture’s conservative tendencies. They compare it with other models like Codex, noting preferences based on specific use cases like code review and video editing, and acknowledge that while Opus 5.5 is a strong contender, some tasks remain better suited to other models.

The episode concludes with the host reaffirming that Opus 5.5 is a major step forward for Claude, notably in its user interaction quality, speed, and visual design capabilities. It is highly effective for front-end prototyping, SVG creation, and complex UI wiring, making it particularly good for visual and designer-oriented tasks. However, the host admits that it still has some issues—such as its somewhat conservative safety behaviors, slower reply times compared to other models, and lesser performance in consumer-facing applications like video editing. They share that it has become their go-to for certain backend and frontend workflows but still prefers Codex for coding and some other specialized tasks. The host notes they will continue testing and benchmarking but feels optimistic about Claude’s future improvements.