Codex CLI vs Claude Code Observations

Claude code max is burning your token limits 40% faster, and anthropic never told you why A developer set up an HTTP proxy to capture full API requests across 4 different claude code versions and what he found should be making way more noise than it is Claude code v2.1.100 silently adds ~20,000 invisible tokens to EVERY SINGLE REQUEST server-side. you can't see them. they don't show up in /context. they just disappear from your balance The proof: v2.1.98 → 49,726 billed tokens v2.1.100 → 69,922 billed tokens same project. same prompt. same account. v2.1.100 actually sends FEWER bytes but gets billed 20K MORE tokens The inflation is 100% server-side and completely invisible to you And billing is actually the smaller problem Those 20K hidden tokens enter the model's ACTUAL CONTEXT WINDOW which means: Your CLAUDE.md instructions get diluted by content you can't see Quality degrades faster in long sessions When Claude ignores your rules you have no way to know if invisible context is the reason You're flying blind inside your own context window npx claude-code@2.1.98 Share this. most people using claude code max have no idea this is happening.

Any agent(including coding agents like claude code, codex, etc) has 4 parts. Here they are in order of decreasing impact on quality of output: Model - most impactful Tools Control loop - the flow of reasoning and action an agent has been configured to follow (for eg claude code has perception-action-verification loop) System prompt A great place to start for building your own agents is by checking out LangGraph: https://lnkd.in/g-xsRPGb or sdk provided by cli coding agents like opencode: https://lnkd.in/gWCCgxH2

Claude Code's source code just leaked. Not a marketing stunt — Anthropic accidentally shipped a source map file in their npm package. Someone extracted 512K lines of TypeScript. Findings: • Voice mode, 24/7 background agents, multi-agent coordination are all built — just feature-flagged off • Internal employees get better prompts and anti-laziness patches • There's an "undercover mode" that hides AI attribution when employees contribute to open source Most practical discovery: the community found two cache bugs causing 10-20x API cost inflation. Patches are now available. From Openclaw to OpenClaude!

🪨 why use many token when few do trick? This open-source plugin blew my mind: caveman turns Claude Code into a caveman - and saves an average of 65% output tokens without losing any technical accuracy. React bug explained? From 1,180 to 159 tokens. Auth middleware fix? From 704 to 121 tokens. No small talk. No filler phrases. No pleasantries. Just the answer. A March 2026 paper even backed this up scientifically: constraining LLMs to brief responses improved accuracy by up to 26 percentage points on certain benchmarks. Caveman not make brain smaller - caveman make mouth smaller. 👉 One-line install: npx skills add JuliusBrussee/caveman Built by Julius Brussee - free, open source, MIT license. https://lnkd.in/dyr3ygkK

The Claude Code Leak: 512,000 Lines of TypeScript On March 31, 2026, a routine npm update accidentally exposed nearly 1,900 TypeScript files over 500,000 lines from Anthropic’s Claude CLI in 2.1.88 version. A large source map file was published, linking compressed code to the full source. Although Anthropic quickly fixed the issue, the code was already mirrored on public sites. No model weights, training data, or user info were leakedjust the client-side orchestration layer. The root cause was a default source map from Bun and a missing .npmignore rule. The leaked code shows features like "undercover mode" for protecting internal info and advanced agent coordination for complex tasks and some other features for background operations and proactive monitoring.

I tested Qwen2.5-Coder (7B and 14B) locally to see how close we are to replacing tools like Claude Code with local models. I ran both on an M4 Pro with 48GB of RAM. Even with a small 8K token context window, each call took 8 to 25 seconds. When you're used to Claude Code scaffolding entire features in a few dozen of seconds, waiting 20 seconds for a single function suggestion is not bearable anymore. The progress is real, but so is the gap. Claude Code and Codex have pushed the bar so far forward that open source now has more catching up to do than it did just a couple of months ago. We live in such strange times.

The cheapest way to run Claude is to make it talk like Kevin from The Office. Someone built a Claude Code skill that makes Claude talk like a caveman. It cuts token usage by 75%. A React re-rendering explanation that normally takes 69 tokens drops to 19. A web search task goes from 180 to 45. Bug explanations save up to 87%. Thinking tokens stay untouched, so reasoning quality is preserved. There was even a paper that found that forcing brevity actually improved accuracy by 26pp. Why waste many token when few do trick.

Every API call returns a status code… But most developers only know 200 & 404. Here’s a simple breakdown of HTTP status codes you should actually understand.

The backend isn't ready. Develop, test, and deliver anyway. That's the whole idea behind this project – and Part 2 walks through exactly how to do it. Clone the repo, run one command, and your client is connecting to a live local mock server in under a minute. From there: push any message, craft any response interactively, simulate timeouts, test malformed payloads, verify reconnection logic. No backend changes. No code deploys. Just you and the dashboard. Full guide + technical deep-dive live now, link to Part 1 included 👇 https://lnkd.in/dNecjTVq

Code mode or programmatic tool calling enables your coding agent to write code that calls your tools programmatically within a code execution container, rather than requiring round trips through the model for each tool invocation. [1][2] [1] Carey, M. (2026, February 20). Code Mode: Give agents an entire API in 1,000 tokens. The Cloudflare Blog. https://lnkd.in/ezE4SsRP [2] Anthropic. (n.d.). Programmatic tool calling. Claude API Docs. https://lnkd.in/eW75u3eB