Claude Code's Prompt-Cache Costs Are 85% Fixed Overhead, Developer Finds

·1 views

A developer systematically measured Claude Code's prompt-cache token usage across three launch paths — the VSCode extension, a VSCode-internal terminal, and a standalone terminal like iTerm2 — and found significant differences in cache behaviour. Both VSCode-based paths consistently hit Anthropic's shared prompt cache, registering 10,270 cache-read tokens, while the standalone terminal recorded zero cache reads and paid full cache-creation costs every session. After stripping out user-controlled elements such as CLAUDE.md files, skills, and hooks, roughly 85% of the remaining cache-creation cost proved to be Claude Code's own fixed overhead. The analysis used a short Python script that reads per-turn token usage directly from Claude Code's JSONL session logs stored under ~/.claude/projects/. The findings suggest that choosing the right launch path is a more impactful cost lever than trimming user configuration files.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

CTF Challenge Exposes Client-Side Security Flaw: Flag Accessible via Direct API Call

A BroncoCTF web challenge called 'Unblur Me' tasked participants with correctly solving 500 derivative problems to reveal a hidden flag image. However, the entire access gate was implemented in client-side JavaScript, meaning the server applied no actual restriction on the underlying image. The flag image was loaded unconditionally from a public API endpoint at startup, with only a CSS blur filter creating the illusion of it being locked. A direct curl request to the endpoint immediately returned the full, unobscured PNG file containing the flag. The challenge highlights two common security pitfalls: relying on browser-side logic as an access control mechanism, and mistaking CSS visual effects for genuine data protection.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

One Founder Used 9 AI Agents to Run a Real Gym for 120 Days Straight

A solo founder in Dongguan, China, deployed nine autonomous open-source AI agents to operate a physical fitness studio, logging 120 consecutive days of production use. The system ran on just 3.6GB of RAM and autonomously recovered from bugs over 34 days without human intervention. The founder claims no existing label — such as 'AI fitness' or 'agent OS' — accurately describes the architecture, which they now call a 'verification layer for physical business behavior.' Notably, Anthropic's official handbook later described a similar governance model, and Jack Dorsey had previously outlined a comparable verification concept, with none of the three parties in contact. The builder argues this independent convergence across unconnected parties signals the approach is architecturally sound, even before the category has a widely accepted name.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Builds Minimalist Portfolio Using Only HTML and CSS, No Frameworks

A software developer shared how they built a modern portfolio site using only semantic HTML and CSS grid and flexbox layouts, deliberately avoiding heavy frameworks and large dependencies. The design prioritizes simplicity, featuring clean light backgrounds with subtle animated gradients and a pairing of Outfit and Ovo typefaces. A standout element is a glassmorphic, capsule-shaped navigation bar that floats at the top of the page. The portfolio also includes a live pulsing availability badge on the landing view to signal openness to new engineering roles. The project was shared on DEV Community as a case for choosing simplicity over complexity in web development.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Swagger and OpenAPI Are Not the Same Thing, but Here Is Why They Overlap

Swagger was originally both a specification and a set of tools for describing REST APIs, but it became so widely adopted that its specification was donated to a foundation and rebranded as OpenAPI. Today, the term 'Swagger' is used interchangeably to refer to the old spec, the tooling, or OpenAPI itself, which causes frequent confusion. In practice, most Swagger files are OpenAPI-compatible, meaning they can be processed and converted to documentation using the same workflows. Developers are advised to generate API documentation from existing machine-readable files rather than writing it from scratch. Teams are also encouraged to standardize on one term — either Swagger or OpenAPI — to reduce confusion for new members.

0 comments Read more at DEV Community