Codex hits 90M installs

The story here is not really “90 million people suddenly started coding with AI.” The real story is that OpenAI upgraded Codex with GPT-5.5 on April 23, 2026, then kept shipping Codex updates into May, and that seems to have triggered a huge burst of package installs around the Codex tooling. (developers.openai.com) The catch is that the headline number floating around — 90 million installs in one week — does not appear in OpenAI’s own announcements. (openai.com) What OpenAI has clearly put on the record is the product change: bigger context, better coding performance, and a push to make Codex the default agent for real software work. ### What actually changed in Codex? OpenAI added GPT-5.5 to Codex on April 23 and called it the recommended choice for implementation, refactors, debugging, testing, validation, and knowledge-work artifacts. The Codex changelog also shows steady follow-on releases — including CLI updates on May 7 and May 8, plus a Chrome extension on May 7 — which tells you this was not a one-day splash but an active rollout. ### Why are people fixated on the install number? Because 90 million in seven days sounds like a platform shift. But turns out that figure is best read as package-install or download activity around `@openai/codex`, not 90 million distinct human users. npm-style numbers can include CI jobs, reinstalls, automated environments, and repeated fetches, so the signal is real demand mixed with noisy measurement. (openai.com) ### So what is solid and not hype? The context window. GPT-5.5 in Codex supports 400,000 tokens, while GPT-5.5 in the API supports up to 1 million tokens. OpenAI also says GPT-5.5 matches GPT-5.4 on per-token latency in real-world serving while performing better on coding-heavy benchmarks like Terminal-Bench 2.0 and its internal Expert-SWE test. (developers.openai.com) That is the concrete upgrade developers can actually build around. ### Why does bigger context matter so much? Because older coding assistants kept losing the plot on large codebases. A model could be smart, but if it only saw fragments, it still had to guess how the rest of the repo worked. (cryptobriefing.com) A 400,000-token window means Codex can keep much more of the code, tests, docs, and error traces in one working memory. A 1 million-token API window pushes that even further for custom workflows. ### Does that mean the memory problem is solved? Not really. Bigger context helps, but it moves the hard part upstream. Someone still has to decide which files, logs, diffs, docs, and tickets get packed into that window. (openai.com) If you stuff in the wrong context, the model confidently reasons over the wrong picture. OpenAI’s own Codex updates hint at this shift — browser use, approval reviews, analytics, and context-compaction work all point to orchestration becoming the real product. ### What becomes the bottleneck now? Retrieval first. Then latency. Then trust. Retrieval matters because the best model still fails if the repo slice is incomplete. (openai.com) Latency matters because giant contexts are expensive to assemble and serve. Trust matters because developers will not hand over merges just because benchmarks improved — they need approvals, review trails, and a way to verify what the agent changed. ### Why does this matter beyond OpenAI? Because coding assistants are starting to compete less on autocomplete and more on workflow ownership. If Codex can sit in the terminal, IDE, browser, and cloud agent loop at once, the product is no longer “write me a function.” It is “take the ticket, inspect the app, patch the code, run the tests, and show me the diff.” That is a much bigger wedge into software work. (developers.openai.com) ### Bottom line? The safest read is this: the 90 million figure is probably a noisy install spike, not a clean user count. But the underlying shift looks real. Codex got a materially stronger model, much larger working memory, and more agent-like product plumbing — and that changes where the engineering pain now lives. (openai.com) (studioglobal.ai) (developers.openai.com)