OpenAI Codex Accelerates Connection and Fixes Cache on Day 19 of 28-Day Challenge
The OpenAI Codex team announced on Day 19 of the 28-day daily improvement challenge that the connection speed for dots calls has significantly increased, and switching inference settings between Codex and Work no longer invalidates the cache.
Codex lead Tibo (Thibault Sottiaux) stated that these are minor updates; previously, switching inference mid-session would clear the context cache, leading to repeated billing and interruptions in long conversations. Now, the cache can be retained, reducing usage consumption.
The team also updated the Codex CLI, adding reusable prompt presets, inline model and permission settings, enterprise MCP login, and fixing issues with long session recovery, while also adding support for Windows enterprise proxy authentication.
Tibo noted that CLI usage has doubled since DevDay; the CLI is open-source, supports cross-session prompt reuse and enterprise scenario authentication, attracting more developers to shift from web to command line for automated scheduling.
At the same time, developers compared Codex with Anthropic Claude Code (default Opus 5.5), which scored higher on benchmarks like Terminal-Bench, and developers often use them in combination: Claude Code handles complex design and architecture, while Codex manages batch or backend tasks.
The improvements focus on usage efficiency and enterprise usability rather than new model releases; the beneficiaries are heavy long-session and CLI users, while those still relying on web interfaces or facing usage limits are under pressure, as competition drives both sides to continuously optimize cache and calling experience.
Source: Public Information
ABAB AI Insight
OpenAI Codex has evolved from early code completion to a full-stack agent, launching reusable cloud environments and voice CLI at DevDay 2026, followed by the initiation of a 28-day daily improvement challenge to address usage limits and competition from Anthropic; it has previously reset usage multiple times and extended cache to 12 hours.
On the capital path, OpenAI is investing resources into optimizing subscription usage and CLI enterprise features, rather than just stacking model parameters; the motivation is to lock in professional developers through cache retention, connection acceleration, and doubling CLI growth, especially in light of Claude Code (Opus 5.5) leading in some benchmarks.
This is analogous to the early migration of IDE plugins to terminal agents, and the competition between GitHub Copilot and Cursor for embedded agents in editors; currently, the focus is shifting from model capability competition to engineering experience and usage efficiency, with both sides differentiating their models and toolchains.
Essentially, this represents a technological substitution and restructuring of the industry chain: coding agents are transitioning from single completions to continuous sessions and enterprise integration, where cache and connection optimization directly determine effective usage; the mechanism is that developer workflows are sensitive to continuity and cost, and daily iterations force platforms to respond quickly to pain points like cache invalidation and enterprise authentication, or users will turn to competitors or mix usage.
ABAB News · Cognitive Laws
Uninterrupted cache is the true usage for long tasks
Doubling CLI indicates professional users prefer control over web
Benchmark leadership does not equate to workflow lock-in