I reviewed Opus 5.5 and GPT-6 Sol live - and the results surprised me

The reviewer conducted a comprehensive live evaluation of three new AI models released the same morning: Claude Opus 5.5 by Anthropic, GPT-6 Soul, and GPT-6 Luna by OpenAI. The reviewer had early access to Opus 5.5 and expanded their 'How I AI' benchmark to test models across diverse categories including PRDs, personal productivity, frontend coding, backend coding, agent personality, long-term research, computer usage, and creative tasks like SVG illustration and video editing.

Key findings about pricing and performance: Opus 5.5 costs nearly twice as much as GPT-6 Soul, though both are cheaper than their predecessors (Opus 5 and Fable 1). All three models showed improvements in cost efficiency, speed, and token optimization through cached inputs. Anthropic implemented Fable-level security restrictions in Opus 5.5 for cybersecurity and biology tasks.

Regarding user experience and communication style: The reviewer noted significant improvements in Opus 5.5's interaction quality compared to Opus 5, describing previous versions as 'annoying' and 'conservative.' Opus 5.5 produces clearer bulleted lists and normal behavior, though it remains somewhat verbose and provides less commentary than OpenAI models. GPT-6 Soul was found to be faster in actual latency and provided better balance in communication level.

Testing results showed distinct model strengths: For PRD writing, Soul performed best. For agent personality and long-term tasks, Claude Opus 5.5 excelled. For personal productivity tasks, Model B and Model E (which appeared to be different OpenAI variants) wrote the most readable messages. Surprisingly, OpenAI models significantly outperformed Claude on SVG character creation, producing cuter and more detailed illustrations. For frontend design, Claude models tended to over-complicate interfaces while OpenAI models provided cleaner, more usable designs. All models performed poorly on video editing tasks.

The reviewer's overall assessment: Opus 5.5 won 'the week' by receiving the highest average scores across individual tasks, particularly excelling in agent voice work and specific design tasks. However, the reviewer's personal preference ('my heart') went to OpenAI's Astra and Soul models, particularly due to Soul's cheaper pricing and superior performance on creative tasks. Notably, an LLM judge disagreed with the reviewer's assessments, preferring Fable and ranking Soul lower, highlighting differences between human and AI-based evaluation criteria.