Published on July 25, 2026 at 06:00 CEST (UTC+2)
Claude Opus 5 (1381 points by alvis)
Anthropic has released Claude Opus 5, a new large language model that delivers near-frontier intelligence at half the cost of its higher-end sibling, Claude Fable 5. Opus 5 sets new benchmarks in coding and knowledge work tasks, significantly outperforming its predecessor Opus 4.8 while reducing cost per task. It excels on evaluations like Frontier-Bench and CursorBench, achieving performance within 0.5% of Fable 5 at half the price. The model features adjustable "effort settings" allowing users to trade off between speed, cost, and output quality, and it is now the default on Claude Max and the strongest option on Claude Pro.
Postgres LISTEN/NOTIFY actually scales (236 points by KraftyOne)
This article challenges the common belief that PostgreSQL’s LISTEN/NOTIFY feature doesn’t scale, demonstrating that with proper optimization it can support high-throughput, low-latency streaming. The authors achieved 60,000 writes per second with millisecond latency on a single Postgres server by using LISTEN/NOTIFY instead of inefficient polling. The mechanism enables durable, pub/sub-style communication directly within the database, making it suitable for real-time applications like streaming LLM tokens. While NOTIFY uses a global lock that causes unintuitive performance behavior, the post argues this doesn’t equate to non-scalability when engineered correctly.
Taylor Farms Called White House to Try to Delay Cyclospora Recall (51 points by JumpCrisscross)
The article reports that Taylor Farms contacted the White House in an attempt to delay a recall of its products linked to a Cyclospora outbreak. Cyclospora is a parasite that causes foodborne illness, and public health agencies typically mandate swift recalls to prevent further infections. The piece highlights tensions between food safety regulators and agricultural companies over timing, transparency, and public risk during contamination events. However, due to unavailable content, specific details about the outcome or justification for the delay request remain unclear.
Sperm Whales blow bubbles to achieve restful, vertical sleep (51 points by hhs)
Researchers from the University of St Andrews and Université de Neuchâtel discovered that sperm whales release gas bubbles while sleeping vertically just below the ocean surface to maintain neutral buoyancy. This behavior counteracts their natural positive buoyancy—caused by spermaceti oil in their heads—and prevents slow drifting upward during breath-hold rest. Data collected via suction-cup tags off Norway’s coast captured clear acoustic evidence of bubble release synchronized with resting postures. The finding explains how these deep-diving mammals achieve restful, stable sleep without expending energy to maintain depth.
India's first privately-developed rocket reaches orbit on debut launch (533 points by sohkamyung)
India’s first privately developed orbital rocket, Skyroot Aerospace’s Vikram-1, successfully reached low-Earth orbit on its debut launch from Sriharikota Island. The 22-meter-tall, four-stage rocket—three solid-fueled and one liquid-fueled—delivered its payload to a 280-mile orbit after a brief delay. Capable of carrying up to 350 kg, Vikram-1 is slightly larger than Rocket Lab’s Electron and marks a milestone for India’s commercial space sector. The successful launch positions Skyroot as a new player in the global small satellite launch market.
My security camera shipped a GitHub admin token in its login page (534 points by hhh)
A security researcher discovered that Hanwha Vision security cameras shipped with a hardcoded GitHub admin token embedded in their web login page. While analyzing firmware extracted from publicly available blobs, the researcher decrypted nested archives using model-specific passphrases and found sensitive credentials left in plaintext. This severe security oversight could allow attackers to access private repositories, compromise development pipelines, or exfiltrate source code. The finding underscores the risks of poor credential management in IoT and enterprise hardware, especially as cameras become more programmable and network-connected.
Opus 5 is currently #1 on Artificial Analysis Intelligence Leaderboard (184 points by aarondong)
According to the Artificial Analysis Intelligence Leaderboard, Claude Opus 5 (at max and xhigh settings) currently ranks as the highest-intelligence AI model available. The leaderboard evaluates models across multiple dimensions including intelligence, output speed, latency, cost, and context window. While Opus 5 leads in intelligence and cost-effectiveness for complex tasks, other models like Mercury 2 and HyperNova lead in raw speed, and Llama 4 Scout offers the largest context window. The data reinforces trends showing specialization—different models excel in different operational trade-offs, making model selection context-dependent.
Show HN: I simulated closing the Strait of Hormuz on real oil trade data (125 points by eliotho)
A researcher built a simulation tool using real 2025 oil trade data to model the economic and logistical impact of closing the Strait of Hormuz—the world’s most critical oil chokepoint. The interactive platform allows users to adjust disruption severity, duration, and elasticity parameters to observe effects on oil prices, trade flows, and emergency stockpile depletion. Based on fluid-stochastic network modeling and published in an arXiv paper, the simulator shows how bypass pipelines and strategic reserves mitigate—but don’t eliminate—systemic risk. The project demonstrates how open data and supply chain modeling can inform energy security policy.
If coding has been solved, why does software keep getting worse? (609 points by pchm)
The author argues that despite rapid advances in AI coding tools and promises of the “Agentic Era,” software quality is noticeably declining. Examples include buggy banking apps requiring multiple logins, focus-stealing applications on macOS, and broken warranty forms. The piece suggests that inflated expectations, rushed AI integration, and misaligned incentives are leading to brittle, user-hostile software—even as developer productivity metrics rise. The paradox highlights a disconnect between engineering efficiency and end-user experience in the age of AI-assisted development.
Designing an Ethernet Switch ASIC (134 points by random__duck)
An engineer claims to have designed the world’s first open-source Ethernet switch ASIC, currently taped out on the Tiny Tapeout gf26b shuttle program with silicon expected in November 2026. The project aims to fill a gap in open hardware, as most open-source silicon efforts focus on RISC-V CPUs rather than networking infrastructure. Spanning 711.2 x 325 µm, the ASIC supports basic switching functionality and responds to growing concerns about proprietary networking hardware in light of recent FCC regulations. The post advocates for transparency and community-driven development in critical network infrastructure.
Model Specialization and Tiered Intelligence Are Becoming Standard
The release of Claude Opus 5 alongside higher-end models like Fable 5 shows a clear industry shift toward offering multiple performance tiers within a single model family. This allows users to optimize for cost, speed, or capability depending on task requirements.
Why it matters: It reflects a maturing market where “one-size-fits-all” models are giving way to nuanced, application-aware deployment strategies.
Implication: Developers and businesses should adopt dynamic model routing systems that select the right model (or effort setting) per request to balance budget and quality.
AI Cost-Efficiency Is Now a Primary Competitive Metric
Opus 5’s marketing emphasizes “double the performance at lower cost per task,” and leaderboards now prominently feature cost-per-task alongside raw intelligence scores.
Why it matters: As AI adoption scales, operational expenses (especially token costs) dominate total cost of ownership, making price-performance ratios critical.
Implication: Teams should prioritize benchmarking not just accuracy or speed, but cost-normalized performance—e.g., “intelligence per dollar”—when selecting models.
Agentic and Coding Benchmarks Are Driving Model Development
Evaluations like Frontier-Bench, CursorBench, and the Coding Agent Index are now central to model announcements, indicating a strategic focus on autonomous software engineering.
Why it matters: LLMs are transitioning from assistants to autonomous agents capable of end-to-end development, shifting R&D priorities toward planning, tool use, and error recovery.
Implication: Organizations should prepare infrastructure for agent-based workflows, including sandboxed execution environments and observability for multi-step AI actions.
Real-World Software Quality Is Degrading Despite AI Advances
Article #9 highlights a paradox: AI coding tools boost productivity metrics while user-facing software becomes buggier and less reliable.
Why it matters: Overreliance on AI-generated code without rigorous testing, design oversight, or user-centric validation can amplify technical debt and UX failures.
Implication: Engineering teams must reinforce code review, integration testing, and usability feedback loops—even (or especially) when using AI copilots—to avoid “productivity theater.”
Hardware and Infrastructure Are Critical Enablers (and Risks) for AI Systems
From open-source ASICs (#10) to insecure IoT firmware (#6), the physical and low-level layers of the stack are increasingly relevant to AI deployment.
Why it matters: AI models run on real hardware, and vulnerabilities or bottlenecks at the firmware, network, or silicon level can undermine even the most advanced models.
Implication: Developers should adopt a “full-stack security” mindset and consider hardware transparency (e.g., open networking ASICs) as part of trustworthy AI infrastructure.
Real-Time Data Systems Are Being Reimagined Around Simpler, Durable Primitives
The Redis-style use of Postgres LISTEN/NOTIFY for streaming (#2) shows a trend toward leveraging battle-tested, ACID-compliant databases for real-time AI applications (e.g., token streaming).
Why it matters: Instead of adding complex message queues or stream processors, teams are finding scalability in “boring” database features when properly engineered.
Implication: Prefer extending existing durable systems over introducing new stateful components unless absolutely necessary—simpler stacks reduce failure modes in AI pipelines.
Public AI Leaderboards Are Shaping Market Perception and R&D Priorities
The prominence of third-party benchmarks like Artificial Analysis (#7) influences both user adoption and vendor roadmaps, creating a feedback loop where leaderboard performance drives model design.
Why it matters: Standardized, multi-axis evaluation frameworks are becoming de facto regulators of AI progress, potentially homogenizing innovation toward measurable (but narrow) tasks.
Implication: Practitioners should treat leaderboards as useful but incomplete; supplement with domain-specific evaluations that reflect real-world constraints like latency, safety, or composability.
Analysis generated by qwen/qwen3-max