Daily
Wednesday, 23 September 2026
Top 50 · open 01–03
01GPT-6 Sol and Luna OpenAI shipped GPT-6 Sol and Luna, the cheaper Astra siblings: Sol is $2 in / $10 out, Luna $0.10 / $0.50 — half the GPT-5.6 promo list — plus caching they say already cut GitHub's uncached tokens in half. The chart is agent cost, not a new brain: Sol at xhigh scores 33.2% on AutomationBench versus Opus 5 at eleven times the spend, and DeepSWE 68.8% sitting next to Fable 5. They're in Work, Codex, and the API today, still missing from Chat, which is why this is a thousand-point spreadsheet fight.
02Claude Opus 5.5 Anthropic's Opus 5.5 is the first 5.5 drop: they put it next to Fable 5.1 on most work, 40% cheaper than Opus 5 ($4/$20, cache reads down to $0.20), and 66.4% on Terminal-Bench 4.0 against Astra's 57.9%. METR and Frontier Design got a look; cyber and biology get Fable-class gates that bounce a lot of cyber tasks to Opus 4.8. I opened it because this is the new default coding agent, not because I needed another Elo table.
03'We hacked the FBI:' Hackers say they have data on all FBI employees ShinyHunters told 404 they hit FBI-related PeopleSoft with a zero-day, moved into GovCloud, and walked out with 2–3 TB they say is every employee and applicant — names, home addresses, phones, spouses. 404 saw a 5,000-row sample; fbijobs.gov got a seizure-notice defacement and then went unavailable; the Bureau is only confirming it's looking at the jobs site. I looked because home addresses of agents is the kind of leak you assume already happened, and this one is claiming it on the record.
04OpenAI GPT–6 Astra breaks Enigma message that has resisted solution since 2005must 05ReBarUEFI: Resizable BAR for almost any UEFI system 06Microsoft killed FoxPro in 2007. Anyway, here's FoxPro revivedmust 07What California is learning from solar panels built over irrigation canals 08Claude Opus 5.5 Intelligence, Performance and Price Analysis (Max) 09SAML: A fractal of bad design 10Pentagon says overreliance on AI contributed to missile strike on Iran school 11WordPress: Unauthenticated path traversal leading to conditional RCE 12Unreal Agent 13How did AMD Ryzen get 50% faster in two years?must 14The current balance of power in open models 15Native apps written in TypeScript and CSS 16MUNI Heritage Weekend in San Francisco 17Show HN: JevBench, a reproducible benchmark for typed decision models 18Markdown in /src 19OpenAI is well positioned to fast-follow Jev 20Obscura: VPN that can't log your activity 21Make Math Automatic with Mathy 22The UV index is not the warm sensation of sunlight on bare skin 23George Lucas Returns to Earth, Bearing Gifts 24The new CC, an AI agent built for families 25Show HN: Training a model to identify AI web content from structure alone 26People hooked on vapes try a new way to quit: cigarettes 2716-bit Intel 8088 chip (c. 1985) 28Launch HN: Coverage Cat (YC S22) – Umbrella insurance via your personal agent 29Apple has added persistent 'ads' to iOS, and it's driving users crazy 30The JavaScript Midlife Crisis 31The trouble with 'ntile()' 32There's a high chance of devices being sold with GrapheneOS preinstalled in 2027 33Did OpenAI solve the wrong Navier-Stokes problem?must 34Can gzip be a language model? 35Porsche puts wireless EV charging into production 36Rabbit Hole: Minimum L-seams 37Writing Rust code that's fast by asking agents to make the code faster 38Solitaire Alone Together 39Show HN: Drop – A rootless Linux sandbox with gVisor support 40AMD's random number generator can't generate a 0?must 41ArchiveBox v0.9 released with new native apps and 50 new plugins 42One Minute Park 43An update on how we confirm your age group on Discord 44Show HN: AI·rete·RAG – a Rete rule engine decides, RAG explains why 45Truman World 46GPT-6 Sol (Max) Intelligence, Performance and Price Analysis 47The Economics of Open-Weight Inference 48Relativistic raytracing 49MiMo v2.6 50Show HN: InstinctFlash – High-Performance Serving Runtime for Robotics Models