Codex hits 20M users as OpenAI credits every account a reset after limit complaints

OpenAI's Codex lead says the agent passed 20 million active users this week and is handing every account a banked usage reset today, hours after tracing a wave of "my limits drain faster" complaints to users piping subscriptions through a shared API proxy. Anthropic's own status page logged its 12th incident in nine days, two of them Thursday, less than an hour apart. Patronus AI open-sourced 200 hours of expert Figma work as an agent training set and says a 27B model trained on it beats Opus 5 on GUI tasks. An anonymous coding model called Ox Alpha landed free on OpenRouter with no builder named. And a catch-up this newsletter owed you: Cognition's CEO denied Bloomberg's report that SpaceX tried to buy it, five days after SpaceX closed on Cursor.

Platform moves

Codex hits 20M users as OpenAI credits every account a reset after limit complaints

OpenAI's Codex lead Thibault Sottiaux said Friday morning that Codex passed 20 million active users this week, and that every Codex and ChatGPT Work account gets a "banked reset" credit today, usable whenever you like. The celebration sits on top of a complaint thread. Users reported limits draining faster; his overnight post, read 1.8 million times, says the team found many affected users running sub2api, a tool that turns one subscription into API traffic shared across many users. That pattern gets flagged by fraud prevention. Signing in with ChatGPT from the official clients, or from open-source ones like Pi and OpenCode, is fine, he says. On the wider drain reports: "we're not seeing anything abnormal," but an investigation is open. If your limits look short, check how you're connected before you blame the meter.

Anthropic logs a dozen Claude outages in nine days, two more Thursday

Anthropic's status page recorded two more incidents Thursday, 44 minutes apart: 29 minutes of degraded Google Sheets, Docs and Slides connectors, then 26 minutes of "elevated errors on requests to multiple models." Pulled from Anthropic's own incident API, that's 12 incidents since August 12: one critical, four major, six minor, one unrated, including Tuesday's nearly three-hour multi-model outage already covered here. Most close without saying how many requests failed. If you ship on Claude Code, budget for a short interruption most weeks.

Claude Status

Welcome to Claude’s home for real-time and historical data on system performance.

Claude logo

What shipped, and what the data shows

Patronus releases 200 hours of Figma design work as an agent training set

Patronus AI open-sourced FigmaTrace Thursday: 3,469 agent trajectories cut from 200-plus hours of recorded expert Figma work, CC-BY-4.0 on Hugging Face, 2,883 train, 586 eval. Ten skills, from accessibility fixes to prototype wiring. Patronus says a Qwen3.8-27B fine-tuned on it beats Claude Opus 5 on two out-of-domain GUI benchmarks by 6.4 and 11.8 points. That's the vendor's number, nobody else's yet. If you're building agents that drive Figma, this is the first real training set.

Ox Alpha lands on OpenRouter, a free, anonymous coding specialist

An unnamed developer put a new model on OpenRouter Thursday, the gateway many coding agents call to reach hundreds of models. It's free, with a 1-million-token context window, described only as built for "efficient coding, sustained agentic work, and real-world production use." No builder is named, no benchmark published, just a request for feedback. OpenRouter has hosted stealth previews before that turned out to be labs testing ahead of a named launch. Treat this one as a preview, not a result.

Money moving through coding tools

Cognition's CEO denies a SpaceX buyout bid, days after Cursor's $60B sale (catch-up)

Bloomberg reported Wednesday evening that SpaceX approached Cognition, maker of the Devin coding agent, about an acquisition, five days after closing its $60 billion purchase of Cursor. Cognition CEO Scott Wu denied it within half an hour: "This is not true. Huge respect for the SpaceX team but Cognition is not for sale and we haven't been talking." Bloomberg says the talks are now inactive. Two coding-agent makers approached by the same buyer in one week says more about SpaceX's ambitions than either deal's outcome.

Cognition CEO denies report that SpaceX tried to acquire the startup | TechCrunch

SpaceX was reportedly in talks to buy AI coding startup Cognition. SpaceX has already acquired Cursor as it races to catch up to rivals like OpenAI and Anthropic in enterprise AI.

TechCrunchRebecca Bellan

Also worth your time

• Codex on AWS Bedrock can't set prompt-cache controls, and one team's bill shows it — the native Bedrock provider sends no cache options for GPT-5.6 Sol, so an agentic workload logged 171.9 million cache-write tokens, about $1,182 of a $1,386 four-day estimate. Open issue, filed August 9, no maintainer reply yet; it reached 141 points on HN overnight.

• DeepSeek ships V4 Flash vision, its first V4 model that reads images — deepseek-v4-flash-vision-exp takes JPEG, PNG, GIF and WebP through the OpenAI-compatible Chat Completions and Responses APIs. Experimental, per the name; relevant if your agent reads screenshots on a budget.

• Vomit pipes Claude Code's terse output through a local model to make it readable — a Go tool that hooks Claude Code and rewrites its output via Ollama or Llama.app, fully local. The author calls it vibe-coded, slow, and Mac-only, and warns the translation can miss the point. 273 points and 268 comments on HN.

• A hobby editor swaps chat prompts for a persistent pseudocode file — Huzzah saves your intent as a .hz file and regenerates code from the diff when you edit it. One person's experiment, no independent hands-on yet, and the week's most-upvoted Show HN at 338 points and 192 comments.

• GitHub's postmortem on the August 17 outage names a Copilot retry loop — errors elsewhere triggered client retries that added load while GitHub was recovering. The root cause was capacity; the detail to know is that Copilot's own error handling made a bad day slightly worse.

Know someone who'd want this in their inbox? Forward it — that's how this grows. And if we got something wrong, or you think we buried the real story today, hit reply. A person reads every one.

The New Way is human-curated — a person picks every story. The summaries are written with AI (Claude) and reviewed before we hit send.