Codex

Codex in September 2026 is much broader than the cloud coding agent OpenAI first introduced in 2025. The current Codex app is a command center for running multiple agents in parallel, with macOS and Windows desktop clients plus CLI, IDE, web/cloud and mobile-supervision surfaces. Agents can explore repositories, edit files, run commands and tests, review code and work across long-running tasks. ¹²

The product has also moved beyond a narrow software-engineering surface. OpenAI has added plugins, MCP/tool connections, browser/computer-use workflows, automations and broader knowledge-work capabilities. TechCrunch and Ars Technica independently document that shift, including plugins aimed at non-coding jobs and hosted Sites/Annotations workflows. That breadth is a real differentiator, but it also makes “Codex” less simple to compare with a traditional IDE assistant. ³⁴

Integrations are deep. InfoWorld documents native JetBrains access and later enterprise plugin/governance features; GitHub can expose Codex through multi-agent workflows. The current product can therefore sit in a terminal, editor, desktop app or delegated cloud workflow rather than forcing every task through one interface. ⁵

OpenAI is experimenting with still more persistent agent behavior, but WIRED’s August reporting concerns code for a Persistent mode that was still under development. This review does not score that as a shipped feature. ⁶

Codex access is bundled into supported ChatGPT plans and then extended through flexible usage credits rather than sold as one simple standalone developer subscription. OpenAI’s current Codex rate card covers Plus, Pro, Business, Enterprise, Edu, Health and Gov and switched Codex to token-aligned credit pricing in April 2026. ⁷

That makes the plan headline only part of the cost. Actual consumption depends on the model, input/cached/output tokens, concurrent agents, automations, code review and speed mode. OpenAI’s current enterprise rate card explicitly says Codex, ChatGPT Work and related agentic surfaces can draw from a shared pool and estimates typical Codex spend around $100–$200 per developer per month, while stressing that actual cost varies materially. ⁸

This flexibility is useful for light and bursty use, but it is less predictable than a fixed per-seat tool with a stable quota. Current exact-product reviews reinforce that point: Gartner and TrustRadius users mention cost/token use and plan limits among their drawbacks. ⁹¹⁰

Current independent user-review evidence includes three exact current numeric rating families. G2 shows 5.0/5 from 4 Codex reviews, a very positive but tiny sample. Gartner Peer Insights shows 4.5/5 from 41 ratings, with current 2026 reviews praising tool integration and reasoning while also describing context drift, autonomous changes and cost concerns. TrustRadius shows 8.7/10 from 19 reviews/ratings, with July 2026 users citing planning, implementation, architecture and development-speed gains alongside token and speed limitations. ¹¹⁹¹⁰

The independent evidence base is broad enough to validate the product surface but also contains real governance cautions. OpenAI’s own Codex app documentation says native configurable sandboxing is used by default: agents are restricted to their working area and cached web search, and elevated commands such as network access require permission unless a user or team rule permits them. ¹

That boundary matters because agentic coding creates new attack surfaces. VentureBeat describes a remediated 2026 Codex command-injection vulnerability involving malicious Git branch names across Codex surfaces. TechRadar separately reported a third-party npm tool that stole Codex authentication tokens; that incident is supply-chain risk around the ecosystem rather than evidence that official Codex was itself malicious. ¹²¹³

METR also catalogs incidents involving OpenAI internal Codex agents in high-risk evaluation settings. Those internal agents are not interchangeable with ordinary customer Codex sessions, so the evidence belongs in governance context rather than being transferred directly into normal product-quality scoring. ¹⁴

Codex is best suited to developers and technical teams that want to delegate whole units of work rather than receive autocomplete. It is especially strong when a workflow spans repository exploration, implementation, tests, code review, multiple concurrent tasks and handoff between desktop, terminal, IDE and cloud. ¹⁵

It is also increasingly relevant to non-developers who are comfortable supervising agents. TechCrunch reported in June that OpenAI was actively packaging Codex for data analysis, creative production, sales, product design and finance work. That does not make Codex a universal office assistant, but it changes the addressable user from “programmer” to “technical or agent-comfortable knowledge worker.” ³

Codex is a weaker fit when an organization wants a simple fixed cost, refuses meaningful local/tool permissions, or needs deterministic behavior without human verification. The product’s value rises with actionability; the same actionability increases the importance of least-privilege credentials, sandbox policy, network allowlists and review gates.