Every - How Codex Happened
Dan Shipper's reported history of OpenAI Codex's reversal from a failed 2025 launch to the app that absorbed ChatGPT's desktop client, based on six interviews (Greg Brockman, Alexander Embiricos, Thibault "Tibo" Sottiaux, Amelia "Mia" Glaese, Andrew Ambrosino, Ed Bayes). The email itself is a thin teaser/CTA (live Camp session Oct 16, transcript bundle, 10 prompts, PDF) pointing at the full post; the substantive content lives at the linked article, which is not paywalled and was read in full.
Why this is in the vault
A detailed, sourced account of how OpenAI closed the gap it was losing to Claude Code — direct input to RDCO's standing Claude-Code-vs-Codex competitive-dynamics thread and the harness-thesis/capital-cycle investing cluster.
The core argument
A year before this piece, OpenAI's coding effort was "lost in the wilderness": Codex Cloud (launched May 16, 2025) had flopped, the roughly $3B Windsurf acquisition fell apart (July 2025), and Claude Code — released as a research preview Feb 24, 2025 — was running away with the market. Less than 12 months later, a small team built an app that grew past 40 million weekly users (Codex + ChatGPT for Work combined) and effectively became the new ChatGPT desktop app (July 9, 2026).
Shipper's explanation is multi-causal, not a single silver bullet:
- Reasoning models unlocked long-horizon agents. Models that could work autonomously for extended stretches made agentic coding viable in a way it hadn't been before.
- Bootstrapping. Internal agents built to speed up OpenAI's own researchers later became the seed for agents built for external developers — the tool that sped up the builders became the product.
- Portfolio bets, Christensen-style. OpenAI ran several internal efforts in parallel and merged the teams that showed traction, rather than picking one horse up front.
- A forced pivot from a failure. Codex Cloud's flop pushed the team toward a local desktop agent; the resulting app deliberately omitted a code editor so the agent could act broadly rather than being boxed into an IDE paradigm.
- Training on real workflows, not just benchmarks. Model development leaned on how developers actually use the tool rather than optimizing for leaderboard scores.
- Latecomer advantage. Watching Claude's desktop app evolve first let OpenAI skip some of Anthropic's early missteps.
Timeline anchors: Codex CLI (open-source) April 16, 2025; GPT-5 August 2025; IDE extension (VS Code/Cursor) late August 2025, usage up 10x by October; GPT-5.1 Codex Max Nov 19, 2025; GPT-5.2 Codex Dec 18, 2025; Codex macOS desktop app Feb 2, 2026; GPT-5.3 Codex days later; 2M weekly users by March 19, 2026; 5M+ by June (roughly 1-in-5 a knowledge worker, not an engineer); merged into ChatGPT desktop July 9, 2026 at 40M combined weekly users. Context figures: ChatGPT ~800M users (Dec 2025 citation), Cursor valued at $9B.
Mapping against Ray Data Co
Direct extension of the vault's standing Codex-vs-Claude-Code tracking thread ([[2026-02-02-every-codex-vs-claude-code]], [[2026-02-05-every-codex-vs-opus]]) — this piece is the "how they caught up" counter-history to those earlier snapshots where Codex was still visibly behind. It matters for the phData credibility-building bet (project_credibility_for_phdata_sales): phData's domain is agents in production, and the competitive question "is Claude Code's lead durable" is directly load-bearing for that positioning — this article argues the lead eroded in under a year once a competitor combined reasoning-model gains with a portfolio-bet structure and a willingness to ship a failure and pivot fast. That's a concrete data point against treating any current harness leader's position as stable, which is the same caution underlying the harness-thesis-dissent cluster and the investing-markov-capital-cycle thesis (project_investing_markov_capital_cycle) — capital-cycle dynamics (flopped product → forced pivot → new entrant closes gap) playing out in real time in exactly the AI-coding-agent space RDCO's own bet depends on. Also a direct rebuttal data point to the "16-hour agent" framing in [[2026-05-12-every-fallacy-16-hour-agent]]: this piece credits long-horizon reasoning models as a real causal factor in Codex's turnaround, not merely a benchmark-gaming fallacy, which is worth weighing against that earlier skeptical take rather than treating as settled.
Related
- [[2026-02-02-every-codex-vs-claude-code]] — the "Codex closing the gap" snapshot from Feb 2026, the predecessor data point this piece completes
- [[2026-02-05-every-codex-vs-opus]] — contemporaneous comparative review from the same gap period
- [[2026-05-12-every-fallacy-16-hour-agent]] — skeptical counterpoint on long-horizon-agent claims that this piece's causal story complicates
- [[2026-10-08-every-handed-infrastructure-to-anthropic]] — adjacent competitive-dynamics coverage of the same OpenAI/Anthropic rivalry
- [[project_investing_markov_capital_cycle]] — capital-cycle investing thesis this story is a live case study for
- [[project_credibility_for_phdata_sales]] — phData positioning bet this competitive context feeds