Claude 4.7 tokenizer cost shock & Floating-point equality and epsilons - Hacker News (Apr 18, 2026)

Let’s start in AI tooling, with a story that’s less about shiny new features and more about the economics of everyday usage. A developer compared token counts between Claude Opus 4.6 and 4.7 and found that the new tokenizer can noticeably increase token usage for English and code-heavy text—sometimes well beyond the “roughly up to 1.35x” expectation. That matters because pricing and quotas didn’t change, so your effective context window shrinks and long sessions can hit limits sooner. The author did see a modest improvement in strict instruction-following on a small benchmark sample, but the broader point is: model updates can change the practical cost profile even when the sticker price stays the same.

Staying in the “things that bite you slowly” category, there’s a strong essay pushing back on the reflexive advice: “never compare floating-point numbers for equality.” The argument isn’t that floating-point is magically safe—it's that the blanket fix people reach for, epsilon comparisons, can be worse. Epsilons are often arbitrary, and more importantly they can break transitivity, which quietly poisons sorting, deduping, geometry code, and anything that assumes equality behaves consistently. The takeaway is to choose comparisons based on the property you actually need—sometimes exact equality is precisely the right guardrail, like detecting true zero cases, and sometimes the real solution is a more robust algorithm instead of a wider tolerance.

Related, but from a more constructive angle, an open-source “Interval Calculator” takes uncertainty seriously by computing with unions of intervals rather than single numbers. The headline benefit is that it stays honest through operations that normally get awkward—like dividing by a range that includes zero—by returning disjoint ranges when that’s the mathematically correct answer. Why it matters: if you’re doing worst-case analysis, validation, or safety bounds, the promise you want is “the true value is definitely inside this result,” not “it’s probably close.” It’s a reminder that for some domains, the best way to handle numerical fragility is to change the model you compute with, not just tweak the rounding at the edges.

On the developer-workflow front, Emacs got a notable security shift in Emacs 30: files are no longer treated as automatically trusted. That’s good for reducing the blast radius of malicious content, especially after real-world concerns like arbitrary code execution vulnerabilities. But the default friction is also real—features can quietly stop working in untrusted buffers, and users get tempted into blunt, risky workarounds. A new package called trust-manager tries to thread that needle by prompting you at the moment it matters, remembering your decision per project, and making trust status visible and reversible. The broader theme here is important: security controls only work long-term if they fit how people actually work.

In open-source creative tools, Kdenlive published a 2026 state-of-the-project update that’s refreshingly focused on stability and polish rather than chasing novelty. The project still shipped meaningful upgrades—like improved masking and object segmentation for background work—while also investing in performance and crash fixes. Interoperability got attention too, with work around OpenTimelineIO to make it easier to move projects between editing ecosystems. Why it matters: for creator tools, reliability is a feature, and incremental workflow improvements often beat flashy additions when you’re editing on deadlines.

A major figure in computer science passed away this week: Michael Oser Rabin, who died on April 14th at age 94. Rabin’s fingerprints are all over modern computing—from foundational theory like nondeterministic automata, to practical algorithms like Rabin–Karp, and especially to cryptography through the Miller–Rabin primality test. If you’ve ever used secure communication that depends on generating and checking large primes efficiently, you’ve benefited from his work. The significance isn’t just historical; it’s a reminder that a lot of today’s “normal” infrastructure rests on ideas that once looked purely theoretical.

For the more mathematically inclined, there’s a new installment of “Category Theory Illustrated” that uses order theory as a bridge into categorical thinking. Instead of treating order as a scoring system for objects, it focuses on relations and the laws they obey—showing how partial orders, preorders, and equivalence classes fit together. The key category-theory punchline is that preorders behave like “thin” categories, where the usual category notions—composition and identity—show up as transitivity and reflexivity. And joins and meets line up with coproducts and products. Why it matters: it’s a clean mental model for moving from everyday “less than” intuition to the diagram-driven way category theory lets you reason about structure.

In retrocomputing and digital preservation, the Amiga Graphics Archive posted an update that sounds small but is actually the heart of archival work: they uncovered very early images by artist Jo-Anne Park, including versions from the Commodore 64 era, helping confirm attribution and show her artistic progression. Preservation isn’t just about storing files—it’s about provenance, credit, and context. When the original source files are missing and work survives only through scattered scans and secondhand copies, every confirmed link between an artifact and its creator becomes part of rebuilding the record of an era.

And finally, from space to the lungs: ESA highlighted a lingering problem from Apollo that’s getting new attention as agencies plan longer stays on the Moon. Astronauts reported irritation and allergy-like symptoms from lunar dust—fine, abrasive particles that stuck to suits and got everywhere. The open question is the long-term risk: could inhaled lunar dust lead to serious respiratory damage over time? The Moon’s environment makes the problem nastier—low gravity keeps dust suspended longer, and electrostatic charging can help it cling and infiltrate habitats. This matters because “dust control” isn’t a housekeeping issue for future bases; it’s a health requirement and a systems reliability issue rolled into one.