Top Stories

Claude Opus 4.7

1528 points · anthropic.com

The runaway #1 story on HN today. Anthropic released Claude Opus 4.7 to a flood of interest — over 1,000 comments in the first twelve hours. The thread is equal parts benchmark analysis, day-one user reports, and speculation about where the frontier is headed. Much of the discussion centers on whether the gains over 4.6 are meaningful for real coding workflows, how it compares to OpenAI’s Codex and the new open-weight Qwen release, and pricing vs. the existing Sonnet tier. A reminder that every Anthropic model drop remains the biggest moment of the week on HN.

Qwen3.6-35B-A3B: Agentic coding power, now open to all

942 points · qwen.ai

Alibaba’s Qwen team released Qwen3.6-35B-A3B, a mixture-of-experts model with 35B total parameters and only 3B active — small enough to run on a single high-end consumer GPU while claiming agentic coding performance competitive with frontier closed models. The HN thread is buzzing: developers are posting benchmarks, laptop runs, and side-by-side comparisons against Opus and Codex. The open-weights angle is the real story — if the results hold up outside cherry-picked benchmarks, this is a serious shot across the bow of the closed-model incumbents.

Codex for almost everything

717 points · openai.com

OpenAI pushed Codex deeper into the stack today, expanding its remit from a coding agent to something closer to a general-purpose shell-and-browser agent. The rollout lands on the same day as Claude Opus 4.7 and Qwen3.6, turning HN into a three-way agent bake-off. Comments range from excitement about the expanded tool surface to frustration with quota limits and reliability gaps on longer autonomous tasks. If you were waiting for the “ambient agents” moment, today was it.

The future of everything is lies, I guess: Where do we go from here?

523 points · aphyr.com

Kyle Kingsbury (aphyr) — best known for Jepsen — turns his attention to the AI-generated content epistemics crisis in a long, characteristically sharp essay. The 572-comment thread is one of the most substantive HN debates in weeks, wrestling with whether provenance, cryptographic signing, or social trust networks can do anything meaningful when the marginal cost of plausible text approaches zero. Essential reading if you care about what the internet becomes in 2027.

Qwen3.6-35B-A3B on my laptop drew me a better pelican than Claude Opus 4.7

322 points · simonwillison.net

Simon Willison runs his now-famous “draw a pelican riding a bicycle” SVG benchmark against today’s two big releases — and the open-weight Qwen model wins, on his laptop, against the newly-minted Claude Opus 4.7. The post is part benchmark commentary, part live-blog of the new agentic coding workflow, and it hit the HN front page within an hour. If you want the quickest gut-check on whether Qwen3.6 lives up to its claims, start here.

Cloudflare’s AI Platform: an inference layer designed for agents

249 points · cloudflare.com

Cloudflare announced its AI Platform — an inference layer explicitly positioned for agentic workloads rather than traditional chat. The pitch: edge-local inference, per-request sandboxing, and tight integration with their Workers and Durable Objects primitives. HN is parsing the announcement carefully, with plenty of commentary on pricing, model selection, and whether Cloudflare can credibly compete with hyperscaler inference APIs. A companion post on “Artifacts” (versioned Git-speaking storage for agents) also cleared 169 points today.

Codex Hacked a Samsung TV

218 points · calif.io

A developer turned OpenAI’s Codex loose on a Samsung TV and documented the results — the agent successfully chained vulnerabilities to gain code execution in a way that looks a lot like a junior pentester at work. The 121-comment thread is a fascinating mix of “this is awesome” and “this is why we should all be terrified,” plus pointed discussion of what happens when the marginal cost of vulnerability research trends toward zero. Pairs well with today’s cybersecurity debate.

AI cybersecurity is not proof of work

204 points · antirez.com

Salvatore Sanfilippo (antirez, creator of Redis) writes a direct rebuttal to yesterday’s viral “cybersecurity looks like proof of work” post. His argument: AI is already upending the economics of both offense and defense, and the “compliance theater” framing misses the actual inflection point happening right now. A thoughtful, experience-heavy counterpoint from one of the most respected engineers on HN — and a rare instance of a top HN post directly responding to another.

The “Passive Income” trap ate a generation of entrepreneurs

183 points · joanwestenberg.com

A pointed essay arguing that the creator-economy/passive-income dream of the 2010s produced a cohort of founders who optimized for distribution grind rather than building anything durable. 136 comments of founders, indie hackers, and former dropshippers working through their feelings on Shopify arbitrage, newsletter mills, and YouTube theater-of-productivity. A useful read if you’re re-evaluating what you’re actually building.

Artifacts: Versioned storage that speaks Git

169 points · cloudflare.com

Cloudflare’s second big ship of the day: Artifacts, a versioned object store that exposes a Git-compatible interface — specifically designed for agents that need to read, write, and roll back their own outputs. John Graham-Cumming posted it himself, and the HN comments are digging into the API design, pricing, and whether this becomes the default scratchpad for long-running agents. An under-the-radar piece of infrastructure that might end up mattering more than today’s flashier announcements.