After previously being disappointed by the AI tool Codex, the author revisited it and found its capabilities considerably improved, altering his initial negative perception. The article explores the dynamic nature of AI development, exemplified by Google’s progression from the flawed Bard to successful projects within its experimental labs. The author uses this as a foundation to reflect on his revised opinion of Codex, particularly in comparison to its competitor, Claude.
Regarding AI tools, the author highlights a significant challenge regarding computational costs which affect user accessibility due to usage limits. Tech companies must balance these costs against user demands and Codex has optimized this balance better than most. Despite being on both the $20 Codex tier and the $100 Claude Max tier, the author finds Codex offers more practical usability without quickly exhausting allocated resources.
Previously, the author criticized Codex for its lack of interactive querying compared to Claude Code, which engages more with users before executing tasks. However, on revisiting Codex, the author appreciates its direct approach which seems more efficient when the user has a clear directive and doesn’t require preliminary Q&A sessions. This revised view acknowledges that the necessity for upfront interaction is situational based on the user’s needs and task clarity.
Codex not only provides high computational efficiency but also integrates innovative features enhancing user experience. These enhancements include an ability to navigate apps on macOS, manage multiple tasks simultaneously through parallel agents, and encapsulate web interactions through a built-in browser. Furthermore, Codex introduced a /goal slash command, which allows users to set long-term objectives for Codex to autonomously manage without constant supervision, shifting the AI tool toward a more autonomous role.
A particularly intriguing forthcoming feature is Chronicle, currently available in a limited preview. It aims to reduce repetitive interactions by maintaining contextual awareness through persistent memories created from screen captures. These memories assist Codex in recalling past project details, thereby streamlining ongoing work without the need for continual reiterations by the user.
In terms of performance, Codex powered by GPT-5.5 has demonstrated substantial capability, particularly in coding-related tasks, evidenced by scoring 82.7% on Terminal-Bench 2.0, surpassing other models significantly. This contrasts with Claude Code, which while being preferable for frontend tasks due to the superior quality of visual code, does not match Codex in other coding capacities as per the assessments.
The author concludes by reflecting on the importance of choosing tools based on performance and functionality rather than brand loyalty or initial impressions. He expresses a nuanced perspective that while Codex has improved dramatically, making it a suitable choice for many backend tasks, Claude remains optimal for frontend applications where visual design and polish are critical. This evolution in the author’s approach toward AI tools underscores the significance of giving technology a second chance, allowing for consideration of advancements and improvements over time. The author’s experience invites readers to engage in the discussion, emphasizing a respectful exchange of views in the evolving narrative of AI development and application.
Read the full post on xda-developers.com
