Karpathy builds HackerTube in 2h with Claude Code, and it tops the weekend's Hacker News

Andrej Karpathy spent Saturday morning telling Claude Code to build a YouTube out of the videos Hacker News commenters recommend, and HackerTube is the weekend's top Hacker News story. Opus 5.5 pulled every HN comment, had Sonnet label them in a Claude Code Workflow (a batch-run feature) until 40K comments recommending a video remained, and wrapped the 6,454 videos they named across 150 topics in a YouTube-style site; "This takes ~2 hours now," he writes. His own manual work: a YouTube API key, a VPS, a domain. On the Hacker News thread he tallied the bill at "504M input tokens (90% cached), 7.6M output tokens," almost all Sonnet labeling, with "~1M tokens of Opus 5.5" for the app and deploy, and added "Only with Opus 5.5+. The thing is beast mode." Copies followed within a day: HackerBooks ("I had Claude build one (Opus 5.5)"), Wikihack, HackerHack. The site's data export is downloadable; the next one is yours to prompt.

OpenAI merges ChatGPT and Codex into one shared weekly limit

OpenAI is merging ChatGPT's Chat and Work tabs and, with it, pooling ChatGPT and Codex usage into one allowance; on the Hacker News thread, Codex subscribers say chat was unmetered on their plans, and that's ending. Codex "will remain a separate experience in the ChatGPT desktop app," OpenAI's email to users says, per TechCrunch; most won't notice, but "some folks might need to wait for resets more frequently." Select users get the merged tabs first. A September 29 plan to cut the $200 Pro tier's Codex allowance is "canceled."

Source: help.openai.com

Anthropic drops the 5-hour cap on its renamed Ultra plan

Anthropic renamed its $200 "Max 20x" plan "Ultra" and dropped its rolling five-hour usage cap for verified accounts, unchanged in price; "Max 5x" becomes plain "Max." Support docs list Ultra's weekly allowance as "10 times the Pro plan's usage every week," the same figure Anthropic and several Hacker News commenters say "20x" always meant per five hours, not per week. Skeptics remain: one called it "a masterful sleight of hand," another hit a weekly limit in two days right after. The cap removal helps only if you had weekly usage left to spend.

Postgres 19 slips to November under LLM-found bugs, committer says

Robert Haas, a PostgreSQL committer, says Postgres 19, due by mid-October most years, "most likely won't be released until November," and "The short answer is that LLMs happened." Against the same April-to-September span last year, commits back-patched to stable branches jumped "by more than 60%" and fixes reaching older releases "by more than 74%," because people are "reporting bugs in PostgreSQL that they found using LLMs." Three large features were reverted late in beta. He names no tool. The Register reported the slip on October 16; these are a committer's own numbers.

Postgres 19 is late, and it's AI's fault

Over the past decade, the timing of PostgreSQL major releases has become very predictable: no earlier than the middle of September, and no l…

Robert Haas

Lawson's benchmark: Opus 5.5 builds the same app for $2.67 in htmx, $4.97 in Angular

Nolan Lawson, a web developer, had Claude Code, on Opus 5.5 at high effort, build one small app in 13 stacks: Hono+htmx cheapest at a median $2.67, Angular priciest at $4.97. The dollars are Claude Code's API-price readout, not his bill (he's on a Max plan); Angular failed to produce a measurable app in two of five runs. Hacker News's top objection: one app is too simple to separate the stacks. Run it on your own app first.

SemiAnalysis's GDPx benchmark holds every model to 12%

SemiAnalysis launched GDPx, a tougher agentic benchmark across 11 high-value domains, and found "even the best model scores just 12%" overall; the article doesn't name which model. Claude Opus 5.5 is "the best all-around model" outside cyber, biology and alignment, winning quant trading, hardware design and chip design outright. GPT-6 Astra takes recursive self-improvement and materials science. The piece is paywalled past its summary, so read the domain wins as reported, not the 12% ceiling as any single model's score.

Factory ships Droid 1.0, defaulting agents to scripted tool calls

Factory's Droid 1.0 makes "programmatic tool calling" (agents script several tool calls in one step, rather than calling them one at a time) the harness's default, to keep context clean and cut model requests. On Factory's own Terminal-Bench 4 runs, Droid scores "3.5 points higher than Codex at 12% lower cost" on GPT-6.1 Sol, and up to 7.6 points higher than Cursor, OpenCode, Devin and Claude Code on Opus 5.5. Nobody outside Factory has re-run the comparison.

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 written with AI. It gathers the day's stories, checks them against their sources and drafts every summary. A person decides what runs and reviews every issue before we hit send.

Karpathy builds HackerTube in 2h with Claude Code, and it tops the weekend's Hacker News

What changed this week