A realistic comparison of Opus and Codex

Codex 5.3 comes out ahead for day-to-day software work—especially when tasks involve real-world complexity like migrations, PR reviews, and “make it correct” engineering. The tradeoff is speed and vibe: Opus 4.6 often gets to a working UI or a usable first draft faster, but it more frequently leaves behind sloppy details, misses edge cases, or introduces security and correctness issues that later require cleanup.

Pricing and access set the stage for why this comparison is messy. Opus 4.6 has published API rates ($25/M tokens out, $5/M tokens in, with fast mode costing 2–3x more and 6x higher expense). Codex 5.3 isn’t broadly available over the API yet, limiting direct benchmarking; the best guess is that Codex 5.3 pricing will resemble earlier Codex tiers (e.g., Codex 5.2’s $1.75/M in and $14/M out). Even so, the creator’s practical takeaway is that Codex tends to be cheaper per token, while Opus can look better per “run” when Codex burns more tokens reasoning through correctness. In subscription terms, Codex also appears more generous in usage quotas, with the creator reporting heavy Codex usage while staying far from limits.

On “intelligence” and hard problem solving, Codex repeatedly wins—sometimes by a lot. In a difficult migration of an old codebase (Round/ping.gg, built on an early T3 stack), Codex 5.3 was the first model to succeed. Its approach: bump what it needs, patch temporarily to unblock, then remove patches once the dependency chain is stabilized—an iterative strategy that avoids the cascade failure pattern where a model upgrades everything and breaks the rest.

Opus’s strengths show up when the goal is to unblock quickly or produce front-end work that looks good. The creator describes Opus as a faster “get it working” partner, particularly for UI design and for “computer-adjacent” tasks like editing dotfiles, SSH-ing to machines, and configuring systems. Opus also sometimes succeeds where Codex fails—such as getting a migration to complete without triggering the kind of deep “fix everything” loop that can trap thorough models.

The biggest fault line is diligence versus shortcuts. Codex is portrayed as “measure twice, cut once”: it tends to notice missing details, handle blockers directly, and avoid leaving insecure or inconsistent code behind. Opus is portrayed as “measure less, ship sooner”: it may ignore blockers by trimming scope, and it can produce working code that later turns out to be wrong or insecure. The creator cites examples involving environment variable handling, database schema/type safety gaps, and even a security-relevant bug where Opus made user association nullable in image generation.

Finally, the comparison isn’t just about model IQ—it’s about platform behavior and trust. The creator prefers Codex for codebase safety and security work, but prefers Opus for the day-to-day experience: faster iteration, more pleasant interaction, and better front-end polish. They also criticize harness quirks (especially Cloud Code) and highlight that Codex’s thoroughness can sometimes become counterproductive in long-running tasks.

Bottom line: if forced to pick one model for serious engineering, Codex is the safer default. If the priority is speed, UI aesthetics, and a more enjoyable workflow, Opus remains compelling—often best used as a complementary tool rather than a replacement.