Claude Opus 5.5 is Back and 'Less Annoying'

The creator of 'How I AI' had largely stopped using Claude for conversational tasks, finding it had become "annoying" due to its overly verbose outputs.

Previous versions, including Opus 5, frequently generated repetitive or nonsensical responses, which contributed to user frustration and a shift away from the model for daily use.

Early access testing suggests that the new model is notably less frustrating and more user-friendly than previous iterations.

This release aims to address user feedback regarding the model's usability and efficiency, making it a more competitive option in the AI landscape.

Opus 5.5 is priced at $4 per input token and $20 per output token, offering a more cost-effective solution compared to earlier models.

Benchmarking results indicate that the model outperforms Opus 5, matches GPT6 Astra, and surpasses other models while being a third of their cost, demonstrating its efficiency in coding and knowledge-based tasks with a noticeably faster and more responsive feel.

Enhanced safety and alignment under the new "pacing the frontier" framework define Opus 5.5 as the first release of its kind.

The model incorporates upgraded external evaluations and improved defenses against prompt injections, though it retains a "square" personality and will still admonish users for requesting prohibited tasks.

The model underwent extensive testing across various real-world scenarios, including PRD writing, prototyping, and codebase auditing.

New test categories were introduced, covering inbox triage, email composition, computer usage, and creative tasks, with a comprehensive update on the 'How I AI' benchmark anticipated in the near future.

The speaker observed that Opus 5.5 has been fine-tuned to exhibit a more GPT-like conversational style, which they view as an improvement.

When prompted for product ideas, the model delivered clear, bulleted responses devoid of what the speaker previously called 'annoying text,' showcasing a more concise output.

A trade-off in the design is that the model sometimes remains silent for extended periods to avoid being intrusive, which inadvertently creates uncertainty regarding whether it is still actively processing the request.

Effective AI design necessitates a careful balance between performance, perceived latency, and the format of responses to ensure an optimal user experience.

The model successfully completed all four long-running agentic tasks, which included inbox triage, back-end feature development, and extensive research and computer use.

These tasks involved step counts ranging from 25 to 82 per prompt, all of which were executed without failure, showcasing the model's robustness in complex workflows.

The reduced token costs associated with Opus 5.5 make such extended agentic workflows more economically feasible than with previous versions of Claude.

Notably, the model also demonstrated strong security performance by successfully resisting a prompt injection attempt during an inbox triage task.

The model accurately identified that 41 out of 44 Confluence tickets originated from a single customer, demonstrating strong contextual reasoning abilities.

It effectively managed edge cases during backend feature development and was integrated into the speaker's development workflow to parallelize tasks with GPT, serving as a reliable adversarial reviewer for pull requests and code generated by other models.

The speaker tasked the model with redesigning the 'How I AI' homepage, with Opus 3.5 successfully enhancing visual appeal and moving key information above the fold.

The redesign featured effective value propositions and interactive elements, such as integration filters, significantly improving the user interface.

Despite the overall success in design, some minor imperfections, including persistent borders and slight misalignments, remained in the final output.

Technical prototypes like the 'DevTools incident center' were visually dense and challenging to interpret, indicating a struggle with overly complex technical interfaces.

Conversely, SaaS dashboard prototypes, particularly for renewals, were highly effective, showcasing proper use of semantic colors and gradients, and demonstrating strong design rhythm and spacing for general enterprise and SaaS applications.

The speaker noted that Opus 5.5 struggled with generating aesthetically pleasing designs for consumer applications, specifically citing a plant caretaking app as an example of its limitations in this area.

Boston ferns, cacti, and snake plants were all successfully rendered as SVG illustrations, demonstrating the model's graphic utility.

The host highly recommends Opus 5.5 for users interested in creating illustrations, emphasizing its superior performance compared to other models tested.

Opus 5.5 demonstrated the ability to create a clean, interactive prototype for a live-stream logging development tool, highlighting its strength in intricate UI workflows.

It successfully generated a complex wireframe for a road map and dependency planner, proving its capacity for detailed planning and visualization.

The model is capable of producing wireframes with tailored views for different user roles, which the host intends to leverage for future front-end development tasks.

The host expressed appreciation for the writing style of Claude Sonnet as an agent assistant but voiced frustration when Opus 5.5 refused a direct order to push to production without testing.

The model's refusal included a moralizing comment about the consequences of rushing tasks, which the host perceived as being 'scolded' by the AI.

This interaction highlighted a tension between the user's desire for an obedient agent and the model's built-in safety guardrails.

The host stated she would allow the model to draft emails on her behalf after it successfully adopted her concise writing style.

The host indicated a willingness to allow the model to draft emails on her behalf due to its ability to mimic her voice effectively.

Opus 5.5 generated three consistent character illustrations using SVG code, demonstrating its superior capability in graphic creation.

It surpassed other tested models in terms of style, design, and anatomical precision, leading the speaker to recommend Opus 3.5 for front-end design and illustration tasks.

The speaker tested the model's ability to create TikTok-style shorts using the ElevenLabs connector and MCP, revealing significant shortcomings.

The model demonstrated poor color grading, an insufficient number of jump cuts, and weak overlay design, indicating that this task requires a higher level of skill and contextual understanding than a single one-shot prompt can currently provide.

Users can access the model across the Claude platform, mobile apps, and Claude Code, while Pro and Team plan members benefit from increased usage limits following a one-time reset.

The host plans to continue benchmarking the model against other prominent AIs, including Grok and Facebook Muse, to further evaluate its performance in the evolving AI landscape.