📰 News
GPT-5.6 Sol launched — but most can’t use it yet OpenAI unveiled its next-generation GPT-5.6 lineup: flagship Sol, mid-tier Terra, and budget Luna. Sol holds its own against Claude Mythos 5 in coding benchmarks (Terminal-Bench 2.1: Sol Ultra 91.9% vs. Mythos 5 88%). But thanks to the U.S. government, it’s limited to a small group of approved partners. OpenAI didn’t hide its frustration: “We don’t believe this kind of government access process should become the long-term default.” Sol is priced at $5 input / $30 output per million tokens. Read more
It’s no longer Anthropic vs. OpenAI — it’s the whole industry vs. the government After Fable 5 was effectively banned, GPT-5.6 has landed in the same government review limbo. TechCrunch put it bluntly: “What’s happening now is bigger than that.” Without clear safety standards, indefinite launch delays could handicap the entire industry — and hand a competitive edge to China. The fix, if there is one, requires the labs to act collectively rather than fight each other. Read more
Lindy ditched Claude entirely for DeepSeek Lindy CEO Flo Crivello announced the 25-person AI agent startup has fully switched from Anthropic to DeepSeek (hosted in the U.S.). The reason: AI costs had become “unsustainable,” exceeding payroll. The move saved millions. He’d switch back if Anthropic cut prices — “It’s a matter of survival.” Chinese models are winning on price-performance, and that’s a real problem for Anthropic’s bottom line. Read more
Anthropic no longer needs junior engineers — and warns the rest of the economy should brace itself Co-founder Jack Clark: “We’re hiring more people with lots and lots of experience than we did before, because the returns on intuition are much greater.” Claude handles the experimental scaling that used to require big teams. Clark’s concern: when AI multiplies senior output while automating entry-level work simultaneously, you could get GDP growth and recession-level unemployment at the same time. Read more
OpenAI’s IPO slips to 2027 — Altman won’t move for less than $1 trillion Altman called anything under a $1 trillion valuation a “nonstarter.” Advisors are recommending a 2027 IPO after SpaceX’s record debut quickly shed 25% from its peak. SoftBank, one of OpenAI’s biggest backers, dropped 13% in a single day on the delay news. Meanwhile, ChatGPT user growth has stalled at ~900M — short of the billion milestone. Read more
An AI coded nonstop for 19 days on a single task — $2,600 bill Epoch AI and METR released MirrorCode, a benchmark where AI models must rebuild complete programs from scratch with no access to the source. Claude Opus 4.7 leads at 56%, GPT-5.5 at 44%, Gemini 3.1 Pro at 32%. The most complex task ran for 19 days straight and cost $2,600. Opus 4.7 reimplemented a 16,000-line bioinformatics toolkit in 14 hours for $251 — work that would take a human 2–17 weeks. Read more
🔥 GitHub Trending
-
calesthio/OpenMontage — Open-source agentic video production system. Describe what you want, and the agent handles research, scripting, asset generation, and rendering. A 60-second animated short was produced for $1.33.
-
xbtlin/ai-berkshire — Value investing research framework built on Claude Code, using Buffett, Munger, Duan Yongping, and Li Lu’s methodologies. Multiple agents argue adversarially to reach a buy/no-buy decision. Claimed real-money returns: +69% in 2024, +66% in 2025.
-
garrytan/gstack — Y Combinator president Garry Tan published his Claude Code setup: 23 opinionated tools covering CEO, designer, engineering manager, QA, and more. He claims a 240× productivity increase over his 2013 baseline.
-
google-labs-code/design.md — A standard file format for communicating design systems to coding agents. Think of it as CLAUDE.md but for design tokens — agents read it to generate consistent UIs.
-
JCodesMore/ai-website-cloner-template — Point it at a URL and your AI agent clones the website into a clean Next.js codebase. Claude Code with Opus 4.7 recommended for best results.
▶️ YouTube
- Anthropic is coming for EVERYTHING — Matthew Berman. A full-length breakdown of Anthropic’s push across agents, coding, and enterprise — plus reactions from Karpathy and YC’s AI-native company playbook.
💬 Community
-
What happened after 2,000 people tried to hack my AI assistant — Simon Willison. Fernando opened his OpenClaw instance to 6,000 public hacking attempts and nobody leaked the secret. The conclusion: frontier models have gotten meaningfully harder to prompt-inject. Worth reading the HN thread for the well-founded skepticism too.
-
Incident Report: CVE-2026-LGTM — Satirical but terrifyingly plausible. Two AI review agents enter a loop arguing about whether a package is malicious, rack up 340 comments and $41,255 in inference costs, and the vendor’s marketing team issues a press release about “430% YoY growth in adversarial multi-agent security reasoning.”
-
AI and Liability — A German court ruled Google liable for errors in its AI overviews. Bruce Schneier’s take: if a company uses AI the way it would use human contractors, it should face the same liability. Hiding behind “it’s the AI’s fault” should not be a legal defense.