# Did Codex Overtake Claude Code? The 7M-User Question > Source: > Published: 2026-07-15 04:44:12+00:00 Codex hit 7 million active users on July 13, 2026 — up from 600,000 at the start of the year. That is a 10x surge in six months, with the last million arriving in a single day following [GPT-5.6's launch on July 9](https://thenewstack.io/gpt-5-6-codex-user-surge/). Latent Space ran the headline everyone was thinking: ["did Codex overtake Claude Code?"](https://www.latent.space/p/ainews-codex-usage-up-10x-in-6-months) 📖 [Read the full version with charts and embedded sources on AgentConn →] Meanwhile, Polymarket bettors are pricing Anthropic at [94.8% odds of having the best AI model](https://fourweekmba.com/polymarket-anthropic-95-percent-best-ai-model/) — a market with $22 million in volume where real money says Claude's model family holds both the #1 and #2 slots globally. Claude Code itself hit $2.5 billion in annualized revenue by February, with Dario Amodei telling the May 2026 Code with Claude conference that they had planned for 10x annual growth and saw 80x annualized in Q1. So which is it? Did Codex overtake Claude Code, or did Claude Code overtake everything? The answer depends entirely on what you measure — and the divergence between those measurements is the most interesting thing happening in coding agents right now. [View original post on Latent Space →](https://www.latent.space/p/ainews-codex-usage-up-10x-in-6-months) The Codex growth trajectory has been vertical. Here is the timeline, drawn from [confirmed milestones](https://www.getpanto.ai/blog/codex-ai-statistics) and [financial analysis](https://finance.biggo.com/news/5e9d35fa-7bbd-4be0-8067-6595b21bbdc4): That is a [730% surge in five months](https://finance.biggo.com/news/5e9d35fa-7bbd-4be0-8067-6595b21bbdc4). Two key inflection points drove it: the February desktop app launch that made Codex accessible outside the CLI, and the April "Codex for (almost) Everything" update that added general-purpose agent capabilities — computer use, browser, Excel, documents. The July 9 GPT-5.6 launch then poured rocket fuel on the curve. Claude Code's numbers tell a different story — not slower growth, but a different kind of growth. In Q1 2026, Claude Code had [4.2 million weekly active users](https://finance.biggo.com/news/5e9d35fa-7bbd-4be0-8067-6595b21bbdc4) while Codex was below 1 million. By Anthropic's last public disclosure in February, Claude Code was generating [$2.5 billion in annualized recurring revenue](https://serpsculpt.com/claude-code-usage-statistics/) — the fastest any developer tool has ever hit that mark. Anthropic has been notably silent on updated user counts since then, even as it raised Claude Code's weekly limits [three times in five weeks](https://medium.com/@AdithyaGiridharan/anthropic-just-raised-claude-code-limits-three-times-in-five-weeks-36a321082e4a) in what one analyst called "a lab being repriced in real time by a competitor it did not expect to lose ground to." The number to watch:Codex leads on WAU (weekly active users). Claude Code leads on revenue per user. At $2.5B ARR with an estimated 4-5M users, Claude Code generates roughly $500/user/year. Codex's revenue-per-user, boosted by 20% non-developer users on free/lower tiers, is almost certainly a fraction of that. User count and business health are not the same metric. There is a critical detail buried in the Codex growth data that changes the narrative: the GPT-5.6 launch did not just add coding users. On July 9, [OpenAI merged the standalone Codex app into the ChatGPT desktop app](https://thenewstack.io/gpt-5-6-codex-user-surge/), launched ChatGPT Work as a new agentic mode for knowledge workers, and began sunsetting the Atlas browser — all in a single day. That means the 7M number is not 7 million developers choosing Codex over Claude Code. It is 7 million ChatGPT power users who now have Codex capabilities embedded in their existing workflow. [OpenAI itself reported](https://techjacksolutions.com/ai-brief/openai-codex-passes-5-million-weekly-users-and-1-in-5-arent/) that 20% of Codex users are non-developer knowledge workers — project managers, marketers, analysts running automations. By the time you subtract the ChatGPT-integrated casual users and the non-developers, the "pure developer coding agent" comparison shrinks considerably. Claude Code, by contrast, is terminal-first. You install it. You configure CLAUDE.md. You learn the hooks and skills system. Every Claude Code user chose to be there. That self-selection produces a different kind of usage — deeper sessions, higher token consumption per user, and the $2.5B ARR that follows. The Hacker News threads on this topic are revealing. In the ["Is Codex really on Par with Claude Code?"](https://news.ycombinator.com/item?id=47750069) thread, user **d-lo** wrote: "I've mostly switched to Codex (GPT-5.4 high) over Claude Code (Opus 4.6)" — but the reason was workflow, not quality. Codex's sandbox model lets you fire off tasks and come back to review, while Claude Code requires interactive presence. User **palguna26** made the same point: they switched due to token constraints, but acknowledged "Claude Code generates superior code quality initially." In the ["Do you use Claude Code, Codex, or something else?"](https://news.ycombinator.com/item?id=48612758) thread, **magicalhippo** described what is emerging as the consensus pattern: "I like to do brainstorming and solution design with Codex GPT 5.5 High, and then have Claude Opus do the implementation." This is not a winner-take-all market. It is a task-routing market. Developer Mckay Wrigley captured the shift more bluntly on X: ["for code, I went from 80/20 claude/gpt to 80/20 gpt/claude in <3 months. surprised by this tbh."](https://x.com/mckaywrigley/status/2047806060061507971) But he immediately added the qualifier: "claude still mogs gpt for non-coding agent stuff." An [analysis of 500+ Reddit developer comments](https://dev.to/_46ea277e677b888e0cd13/claude-code-vs-codex-2026-what-500-reddit-developers-really-think-31pb) found that Claude Code won 67% of 36 blind code quality tests, but Codex dominated on speed and cost efficiency. On SWE-bench Verified, Claude Code leads 59% vs Codex 56.8%. But on Terminal-Bench — which tests real CLI workflows — Codex crushes it at [77.3% vs Claude Code's 65.4%](https://www.morphllm.com/comparisons/codex-vs-claude-code). The benchmarks themselves don't agree on who is winning. Anthropic's response to Codex's growth has been capacity, not marketing. Three limit increases in five weeks tell the story: On the same day as that third increase, [OpenAI offered two months of free Codex access to enterprise customers willing to switch from Claude Code within thirty days](https://pasqualepillitteri.it/en/news/2494/claude-code-weekly-limits-50-percent-anti-codex-anthropic-2026). Both companies staked out territory on the same day, and neither blinked. But OpenAI's most surprising move was not competitive — it was collaborative. In early July, [Romain Huet from OpenAI announced](https://x.com/romainhuet/status/2038677236304245087) the open-source Codex plugin for Claude Code: "We've seen Claude Code users bring in Codex for code review and use GPT-5.4 for more complex tasks, so we thought: why not make that easier? We love an open ecosystem!" The [codex-plugin-cc repo](https://github.com/openai/codex-plugin-cc) lets you run `/codex:review` for automated code reviews and `/codex:rescue` for task delegation — directly from your Claude Code session. That is OpenAI shipping a plugin that runs inside Anthropic's product. The message is clear: we do not need to kill Claude Code to win. We just need to be present wherever developers are working. Contrarian take:The "did Codex overtake Claude Code" framing is already outdated. OpenAI shipping a pluginforClaude Code is the tell. The companies themselves have moved past the winner-take-all narrative that the tech press is still running. The real competition is not Codex vs Claude Code — it is which model and harness combination earns the highest-value tasks in a multi-agent workflow. Here is the divergence that should make everyone pause. On Polymarket, traders betting real money — [$22 million in volume](https://polymarket.com/event/which-company-has-best-ai-model-end-of-june) — give Anthropic 94.8% odds of having the best AI model. Not just the best coding model. The best model, period. The "second best AI model" market has Anthropic at 90.5% as well — meaning the market consensus is that Claude holds both the #1 and #2 global positions. How do you reconcile "Codex has 3x the users" with "Anthropic has a 95% probability of having the best model"? Two possible explanations: **Explanation 1: Distribution beats quality.** Codex is bundled with ChatGPT, which has 400M+ monthly users. Claude Code requires a separate installation. OpenAI has an enormous distribution advantage that is independent of model quality. This is the Microsoft Office playbook: being default is worth more than being best. **Explanation 2: Different products for different jobs.** Codex has become a general-purpose agent platform (computer use, browser, documents, automation), while Claude Code has stayed focused on deep coding. They are not competing for the same job. A 20% non-developer user base on Codex is not eating into Claude Code's developer base — it is expanding the total market. Both explanations are probably partially true. And both point to the same conclusion: raw user count is a poor proxy for competitive position in this market. The data points toward a specific pattern: developers are not choosing between Codex and Claude Code. They are routing tasks to each based on the work type. **Codex wins on:** **Claude Code wins on:** The [Codex plugin for Claude Code](https://github.com/openai/codex-plugin-cc) codifies this pattern. You stay in your Claude Code session — with your CLAUDE.md, your hooks, your skills — and delegate specific subtasks to Codex. The harness stays Claude. The execution can be either model. This is exactly what our previous coverage predicted. [The agent harness, not the model, is the moat](https://agentconn.com/blog/agent-harness-not-model-guardrail-stack-2026). And the harness ecosystem — Anthropic's CLAUDE.md config, community skills directories, and the [multi-agent CLI switching](https://agentconn.com/blog/cc-switch-cli-claude-code-openclaw-codex-gemini) layer — is still decisively in Claude's orbit. If you are a developer or engineering lead trying to make a decision right now, here is the actionable read: **1. Stop treating this as a binary choice.** The developers who are shipping fastest in mid-2026 are running both. Codex for the ticket backlog, Claude Code for the architecture decisions. The [multi-agent CLI layer](https://agentconn.com/blog/cc-switch-cli-claude-code-openclaw-codex-gemini) makes this trivially switchable. **2. Watch revenue-per-user, not WAU.** Codex's user count includes ChatGPT power users who tried the coding features once. Claude Code's user count is developers who configured a CLAUDE.md file and run multi-hour sessions. The engagement quality is categorically different, and that is what determines which tool will get the most investment and fastest iteration. **3. Build your harness to be model-agnostic.** Your CLAUDE.md, your agent skills, your [guardrail stack](https://agentconn.com/blog/agent-harness-not-model-guardrail-stack-2026) — these should survive a model swap. If your entire workflow breaks when you switch from Opus to GPT-5.6, your architecture is fragile. The Codex plugin pattern shows where this is going. **4. The real question is not "who won."** It is whether your team has a coherent task-routing strategy. Which tasks go to the background sandbox (Codex)? Which tasks need interactive presence (Claude Code)? Which tasks need both in sequence (brainstorm on one, implement on the other)? That routing decision, not the tool choice, is what separates high-output teams from the rest. Did Codex overtake Claude Code? In raw user count, yes — decisively. In revenue, no. In model quality according to prediction markets, not even close. In developer preference for deep coding work, it depends on the task. The better question is whether "overtake" is even the right frame. OpenAI is not trying to kill Claude Code — they shipped a plugin for it. Anthropic is not panicking — they are raising limits and printing $2.5B in ARR. The market is bifurcating, not consolidating. The 7 million number is real and impressive. But the 94.8% Polymarket number is also real. The $2.5B ARR number is also real. And the developer who brainstorms on Codex and implements on Claude — that person is also real, and they are the future of this market. Stop picking a winner. Start building a routing strategy. *Originally published at AgentConn*