Top Stories

GPT 6.1 Sol: Near-Astra Intelligence for a Fifth of the Price

966 points · openai.com

OpenAI dropped GPT-6.1 Sol just a week after GPT-6 Sol, and the headline is price: near-frontier “Astra-class” intelligence at roughly a fifth of the cost, landing around $2 per million input tokens and $10 out. Independent trackers already put it near the top of the intelligence charts with a 1M-token context window. HN’s interest here is less about the benchmark bragging and more about the economics — when a model this capable gets this cheap, the calculus for building agents, batch pipelines, and always-on tooling shifts overnight. The thread is full of people re-running their own evals and recalculating unit costs for products they’d shelved as too expensive.

Livenerf: Has Opus 5.5 Been Nerfed Yet?

657 points · github.com

A tongue-in-cheek but genuinely useful project: a live tracker that continuously benchmarks Anthropic’s Opus 5.5 to answer the perennial community question, “has the model been quietly nerfed?” It taps into one of the deepest recurring anxieties among heavy LLM users — the suspicion that a model silently gets worse after launch, whether from quantization, routing changes, or cost-cutting. Whether or not the effect is real, the fact that someone built automated, public monitoring for it says a lot about how much trust (and paranoia) now surrounds the frontier labs. The comments are a mix of “finally, receipts” and methodological nitpicking about how you’d even measure this reliably.

Dots: Always-On Agents

638 points · openai.com

OpenAI’s second big launch of the day introduces Dots, a framework for always-on agents that persist and act continuously rather than waking up only when you prompt them. This is the natural next step in the agent story: instead of a chat you poke at, you get background workers that watch, react, and take initiative. HN is chewing on the obvious tensions — reliability, cost when something runs 24/7, and the governance question of what an autonomous agent should be allowed to do without a human in the loop. Paired with the cheap GPT-6.1 Sol pricing above, it’s easy to see why the two announcements landed together.

How Delhi Cut Electricity Loss from 50% to 5%

532 points · spectrum.ieee.org

IEEE Spectrum tells the story of how Delhi slashed its aggregate technical and commercial power losses from a staggering 50% down to about 5% — a mix of metering, anti-theft engineering, privatization incentives, and dogged operational work. HN readers love a concrete infrastructure turnaround, especially one with hard numbers and transferable lessons for other grids in the developing world. The thread digs into what actually drove the change, how much was technology versus policy, and whether the model generalizes beyond Delhi’s specific circumstances.

PS5 Relapse Exploit

316 points · github.com

A fresh PlayStation 5 exploit chain published on GitHub, complete with the technical write-up console hackers live for. This is prime HN territory: reverse-engineering locked-down consumer hardware, documenting the vulnerability, and showing exactly how the pieces fit together. Beyond the homebrew and preservation angle, readers debate the cat-and-mouse dynamics with Sony’s firmware team and what the exploit reveals about the PS5’s security architecture.

A Staff Engineer’s Guide to Inventing Work

279 points · sujithjay.com

A sharp essay on a failure mode senior engineers know too well: manufacturing work that looks impressive but doesn’t actually move the needle — gratuitous rewrites, premature abstractions, and projects that exist mostly to justify a promotion packet. It resonated because it names something uncomfortable about incentives at the staff-plus level, where visibility and real impact can quietly diverge. The comments are a candid mix of confessions, disagreements about what counts as “real” work, and war stories about org structures that reward the wrong things.

NASA Asked Former SR-71A Staffers to Help with a Secret Restart

185 points · aviationweek.com

Aviation Week reports that NASA quietly reached out to former SR-71 Blackbird personnel to assist with a restart effort — reviving institutional knowledge from one of the most storied aircraft programs ever built. HN has a deep soft spot for the Blackbird’s engineering lore, so the story doubles as both news and nostalgia. Readers are speculating about what NASA actually needs the expertise for, and marveling at how much hard-won knowledge lives only in the heads of the people who were there.

A Virtual Power Plant Hidden in Vermont Homes

218 points · bbc.com

The BBC profiles a Vermont utility that’s effectively replacing conventional peaker plants with a fleet of home batteries, coordinated as a single “virtual power plant” that keeps the lights on during storms. It’s a tangible, working example of distributed energy resources doing real grid work rather than living in a slide deck. HN’s engineers are into the control and coordination problem — how you orchestrate thousands of household batteries reliably — and the economics of paying homeowners to lend their storage back to the grid.

Language Models for Text Classification: From Bag-of-Words to Jev

134 points · magazine.sebastianraschka.com

Sebastian Raschka traces the full arc of text classification, from classic bag-of-words models through transformers and up to the newest Jev-style approaches. It’s the kind of well-illustrated, historically grounded explainer HN reliably rewards, giving practitioners a mental map of how the field got here and when a simpler method still beats a giant LLM. The thread turns into a practical discussion of when classic techniques remain the right tool versus reaching for a modern model.