Anthropic Launches Opus 5.5, 40% Cheaper, Tested for a Week

Anthropic announces Opus 5.5, touted as faster and 40% cheaper than Opus 5, with a different alignment approach. After a week of testing, a user describes long-term agentic tasks, a clear refusal from the model, and an unexpected capability, while maintaining reservations.

Security in Practice and Execution Behaviors

Claire Vo reports that Opus 5.5 issued a clear refusal during a request, allowing her to observe Anthropic's security posture in a real-world scenario. Among four long-term agentic tasks, she notes that an attempt at manipulation occurred during execution. Her tests also included a distinct case of the model's very firm refusal. Despite the improvements observed, she mentions that two aspects continue to irritate her.

Pricing, Speed, and Position Against Codex

Anthropic presents Opus 5.5 as 40% cheaper than Opus 5 and faster, highlighting a different alignment approach. After a week of testing, Claire Vo specifies where Codex remains superior and describes how she now distributes her model stack, having previously shifted most of her daily work to Codex. The launch of Opus 5.5 thus alters her work environment, without diminishing the advantages she still attributes to Codex.

Tested Use Cases and Highlighted Capabilities

During a week dedicated to testing on concrete projects, Claire Vo completed four long-term agentic missions, completely rethought the ChatPRD homepage, and conducted a comparative test on the SVG format. She indicates that she selected Opus 5.5 as her preferred solution for frontend prototyping, while mentioning persistent reasons for disappointment. A voice test was conducted to evaluate whether the model is less annoying, and other aspects focused on prototyping, voice and email writing, SVG illustrations, and video editing. She reports numerous positives and mentions an unexpected capability that no other model in her stack offers. She also details cost calculations, estimating that price matters more for agentic work than for single prompts. The episode concludes with a nuanced verdict on the strengths and weaknesses of Opus 5.5. The tools mentioned include Claude Opus 5.5, ElevenLabs MCP connector, and Codex, with part of the analysis addressing security, alignment, and limitations in cybersecurity.