The coding tool wars are getting a scoreboard, and OpenAI is winning on raw numbers. Meanwhile, Anthropic's silence on its own metrics is starting to look less like strategic restraint and more like a communications problem. Both stories are below, and they connect.
01
OpenAI's Codex just hit 7 million users. Where's Claude Code's number?
Latent Space flagged something worth sitting with: OpenAI's Codex coding tool has grown more than 10x in six months, from roughly 550,000-700,000 users at the start of the year to 7 million as of this week. The last public Claude Code figure was around 2 million users back in February. Anthropic hasn't updated that number since, which is notable when your main competitor is dropping milestone tweets every 48 hours. Latent Space offers a charitable explanation: Anthropic may have shifted the bulk of its coding activity to Claude Tag, a Slack-based tool with different usage patterns that's hard to compare directly. That might be true. But "hard to compare" is also a convenient reason not to publish a number that's losing ground. ---
Why it matters: If you're a developer choosing where to build your workflow, user growth is a real signal about which tools are getting better faster. A product with 7 million active users generates more feedback, more edge cases found, and more pressure on the team to ship fixes. Codex's lead, if it holds, compounds.
Anthropic's communications problem, named directly
Product executive Peter Yang posted a pointed critique: Anthropic builds great models but communicates poorly, and should take notes from how OpenAI engages its community. The post got 142 likes and 26 replies, which isn't viral, but the replies were substantive enough that it's worth flagging. This connects directly to the Codex story above. When you don't publish user numbers, don't do developer livestreams, and don't engage the discourse on X, you cede the narrative by default. OpenAI has figured out that community management is a product decision. ---
Why it matters: Anthropic's reticence costs real users. Developers who are on the fence between Claude Code and Codex are choosing based on what they hear in the community, and right now they're hearing a lot more from one side.
Microsoft is using AI agents to verify the code that secures Windows
Microsoft Research published a detailed post on how they're using AI agents, alongside formal mathematical proof tools, to verify the cryptographic code inside SymCrypt, the security library that underpins Windows and Azure. The project has released verified implementations of SHA-3 and ML-KEM, a post-quantum cryptography standard. The practical upshot: instead of trusting that security-critical code works correctly, Microsoft can now machine-prove it does. ---
Why it matters: Cryptographic bugs are catastrophic and notoriously hard to catch in code review. If AI agents can help scale formal verification from a research specialty into a standard engineering step, the security implications for cloud infrastructure are significant. Your bank's encryption keys pass through code like this.
Nikunj Kothari, an investor in the AI space, posted an observation that landed with the SF crowd: lots of founders claim to be running elaborate AI agent setups, but when asked what they're actually building and for whom, they go quiet. His point is that spinning up subagents feels productive without requiring the harder work of knowing what you want to accomplish. ---
Why it matters: This is the 2026 version of a founder who has a beautiful pitch deck and no customers. If your AI workflow can't be explained in one sentence to a potential user, more compute won't fix it.
**The Claude Fable subscription extension, and what it might mean** — Thibault Sottiaux noted that Claude Fable's subscription status was quietly extended to cover Team, Edu, and other paid tiers, with the comment "quite simple really." The timing, during the week Codex hit 7 million users, has people speculating. Anthropic hasn't said anything officially, which is on-brand.