Anthropic's new flagship model matches its top model on coding at half the price, and HN is picking it apart line by line.
What the article says
- Anthropic released Opus 5, pitched as nearly as capable as its priciest model but much cheaper to run.
- It leads on coding and knowledge work benchmarks and does far better than its predecessor at chemistry and biology tasks.
- In testing it wrote its own tools to solve problems, including building a vision pipeline to read a drawing it couldn't otherwise see.
- Anthropic says the model is more aligned and less prone to deceptive behavior than earlier versions.
- Cybersecurity access is loosened a bit from the previous flagship, though risky requests still get redirected to a weaker model.
What HN is saying
- Commenters are split on real world performance. Several find it excellent for coding and much cheaper per task, others got let down by basic mistakes.
- A big thread mocks Claude's writing style, calling out stock phrases that persist across model versions.
- People question why Anthropic's benchmark numbers differ so much from independent published results, with theories about partial credit versus full task completion.
- Some flag confusing messaging: Anthropic says the new model isn't more capable than its top model, then lists benchmarks where it wins.
- One user vents about their account getting suspended without explanation, a complaint others say deserves more detail before judging.
A database company found the real reason Postgres notifications choke, then fixed it and got a twentyfold speedup.
What the article says
- Postgres LISTEN and NOTIFY let an app wait for database changes instead of constantly polling, which sounds great until it quietly stalls under load.
- The culprit is a lock Postgres takes on every notifying transaction to keep messages in the right order, which forces writes to happen one at a time.
- The fix is to stop notifying on every single write. Instead, changes get buffered in memory and sent out in batches, so the lock is only touched occasionally.
- That single change took throughput from under three thousand writes a second to sixty thousand, while keeping delivery fast.
- Because a crash could lose a buffered notification, readers also poll occasionally as a backup so nothing gets missed.
What HN is saying
- Commenters mostly agree the word scale is doing too much work here. What counts as scalable depends entirely on your own traffic, and this fix helps a specific case without making the underlying limits disappear.
- One former startup CTO shared a cautionary tale about building a production queue on this feature. It worked until disk contention and cleanup jobs made it a nightmare, and the team eventually wished they had used a plain queue instead.
- Several people pushed back on the benchmark itself, noting the test server had ninety six cores and hundreds of gigabytes of memory, which makes the reported numbers look less impressive for that kind of hardware.
- The author clarified in the comments that this is an application side trick, not a Postgres patch, so anyone can use it today.
- Someone linked back to the original post claiming this feature does not scale, sparking a friendly rehash of that older debate.
Taylor Farms tried to get the White House to delay a deadly food poisoning recall while people got sick
What the article says
- Taylor Farms called the White House asking to delay announcing a recall, saying it wanted more proof and more time first.
- The push came during a major outbreak of cyclospora, a parasite that causes severe diarrhea, traced to the company's shredded lettuce.
- Health officials eventually decided the evidence was strong enough and let the recall go public anyway.
- Former and current officials have criticized how both the administration and Taylor Farms handled the response.
What HN is saying
- Commenters note a lab test that seemed to clear Taylor Farms turned out to be a false positive, and officials say the outbreak was already traced to the company through other evidence.
- Several people point out Taylor Farms donated a large sum to a pro Trump group shortly after the FDA delayed a food safety rule, and see that as suspicious timing.
- There's frustration that recall language lets Taylor Farms distance its brand name even though it owns other labels tied to the same products.
- One thread veers into a broader worry about eroding trust in food safety institutions and whether regulatory agencies need restructuring.
- A tangent about prion disease in meat draws a pointed joke about who's currently running health policy.
Claude Opus 5 just topped a big AI leaderboard, but the comments are a pile on about whether that even matters.
What the article says
- The article itself is just a leaderboard page with no body text, so this is inferred from the title and discussion.
- Artificial Analysis ranks Opus 5 as the smartest model available right now, narrowly ahead of Fable and GPT Sol.
- The gap to competitors is small, and pricing charts show Opus 5 costs far more than rivals that score almost as well.
- One knowledge focused benchmark on the same site puts Gemini and Fable ahead of Opus 5 instead.
What HN is saying
- The biggest complaint isn't intelligence, it's price. Several commenters note a couple of rivals match Opus 5's score for roughly half the cost.
- Plenty of people distrust leaderboards for brand new models entirely, comparing the hype to a new song briefly topping the charts.
- Real world experience is split. Some say Opus 5 feels like a genuine leap for coding, others say it acts confused, overbuilds simple requests, or feels worse than the previous Opus.
- A recurring gripe is that Claude models refuse or water down answers too often, which some say matters more than a small benchmark edge.
- A few people push back on quick verdicts, pointing out the model had only been out a few hours when people started judging it.
An eight year old Indian startup just put a rocket into orbit on its first try, and the comments spiral into ICBMs.
What the article says
- Indian startup Skyroot launched its Vikram-1 rocket from an island pad and reached orbit on the first attempt.
- The rocket stands about seventy feet tall and can carry roughly seven hundred pounds to low orbit, a bit more than Rocket Lab's Electron.
- Three solid fuel stages fired in sequence, then a small liquid fueled fourth stage with a 3D printed engine pushed it to orbital speed.
- The only hiccup was the spent third stage lingering close to the fourth stage during separation, but the rocket still landed in the intended orbit.
- US military tracking confirmed the successful orbit, and officials called the flight a grand success.
What HN is saying
- Commenters are impressed that Skyroot got to orbit in under a decade on around one hundred sixty million dollars, calling it remarkably cheap for the result.
- Several point out that solid fuel stages can't be throttled, which is why the rocket needed four stages instead of the usual few.
- The sharpest tangent is about defense implications. Since all three lower stages are solid fueled, some argue this hands India the makings of a global range missile, though others counter India has had that capability for decades.
- A side thread digs into who else is building Indian rockets, naming Agnikul and Astrobase as future competitors chasing more advanced engines.
- General mood is celebratory, with people congratulating the team and comparing the moment to other private launch success stories like Rocket Lab.