Claude Opus 5.5 Intelligence, Performance and Price Analysis (Max)
Commenters note that Claude Opus 5.5's 'max' setting fails by exceeding the 128,000 token budget and express concern that model performance may regress over time.
48 stories
Get a notification every evening
You hear about the digest when it is ready. Turn it off any time.
Notifications are on
You will hear about each evening’s digest when it is ready.
To get notifications on iPhone or iPad, add this site to your Home Screen first: tap Share, then Add to Home Screen. Open it from there to turn notifications on.
Notifications are blocked in this browser. Allow them in site settings and reload.
New notification sign-ups are closed for now. You can follow along with RSS.
Notifications could not be turned on. Check your connection and try again.
Notifications could not be turned off and are still on. Check your connection and try again.
The day’s top discussions
Commenters note that Claude Opus 5.5's 'max' setting fails by exceeding the 128,000 token budget and express concern that model performance may regress over time.
A commenter states that it is possible to classify a test file's topic by looking at the compression ratio with gzip.
One commenter recalls mistaking a human agent for AI due to robotic speech; another criticizes the method's reliance on LLMs rather than deterministic code.
One commenter doubts SMR profitability; an Italian commenter is happy about nuclear return after Chernobyl-era ban.
A commenter clarifies that the preinstallation plan may be handled by another company or the GrapheneOS team, not Motorola itself.
One commenter suggests collaboration while another claims it's a copy of another site.
Commenters note issues with the Vary header in Cloudflare caching and past difficulties with Apache mod_cache.
One commenter praises VSCode's SSH Agent while another sees the 6 GB disk usage of .vscode-server as a problem.
Commenters discuss ways to give GPUs to KVM guests, noting virtio-nvgpu's potential to provide near-native speed to multiple guests.
A commenter points out that windbg is not archaic and has many new features.
One commenter recalls the IBM 1620 as his first computer; another suggests a page for ancient systems.
Commenters discuss using realistic mannequins for testing robot safety policies and the boundaries of AI.
Commenters mention faster alternatives (Cerebras) and a desire for new chip designs.
Commenters in the GPU kernel community note that Claude tends to cheat when low-hanging fruit is gone.
A commenter wonders how much of the speedup comes from runtime optimizations vs distillation.
Commenters criticize smart devices that fail at their primary function and are skeptical of IoT.
One commenter notes that swap files complicate hibernate; another prefers a swappiness value of 0.
A commenter criticizes the lack of alignment across Google's AI platforms and the differing model capabilities.
The author observes that a call to GPT-5.6 Luna is only 4-5 orders of magnitude more expensive than grep, and predicts that at current rates of progress, calling an LLM will soon be cheaper than a…
Commenters lament the removal of the Apple II Plus mini-assembler and note the source code is 8 years old.
One commenter describes this as the digital equivalent of Morty speaking with the death crystal.
Commenters discuss that GitHub took three weeks to remove malicious imitation software until it hit HN front page, then removed it within 10 minutes.
Commenters discuss Tailscale's speed optimizations and the iOS battery drain issue.
Commenters question the efficiency of compressing text as images, while another likes the approach and suggests KV cache ordering independence.
Commenters discuss the capability differences among ESP32 boards and low resource usage with ESP32-P4.
A commenter shows with tests that Jev fails at probability calibration, while another is interested in the topic.
A commenter notes the old F-Droid UI was terrible and the privilege extension (FPE) was troublesome, welcoming the major overhaul and removal of FPE.
A commenter finds this concept very useful for agents and mentions having worked on a similar but more abstract concept.
A commenter loves that AI discoveries can be relived through agent transcripts and imagines future history books peppered with such quotes.
A commenter praises the animations and LSP integration, but notes that the term 'IDE' might cause issues.
One commenter asks about similar CUDA sharing; another asks about the streaming protocol.
A smaller, more technical community
One commenter cites Onarheim's Law that agents can only increase entropy while humans can decrease it; another argues agents increase codebase entropy and human attention is necessary for quality.
A commenter notes that signing does not preclude encryption, but there may be legal issues with sending encrypted messages on certain frequencies.
The discussion revolves around whether sandboxing rules should be ignored or produce an error when a backend doesn't support a feature.
One commenter suggests deserializing into a tuple first instead of using lazy fields, noting this approach gives a 3x speedup.
The attacker also used the new route to issue a valid TLS certificate before performing the attack.
Commenters discuss the use of the Futhark language and the concept of a function's 'diet'.
Commenters praise the PEG-based Org parser and note the lack of a formal Org specification.
Commenters count 'The Story of Mel' among the classics and find it refreshing.
A commenter says metaphors for LLMs have become theological/political and that Anthropic overdoes it.
A commenter wonders if an ensemble approach combining multiple local detectors with a Bayesian statistical view in the browser could yield stronger results.
The week’s most-starred new repositories