// blog
Blog
Insights on AI, software development, and the future of work.
Proxy-level measurements of Claude Code token overhead: the 33k request floor, cache write behavior, instruction file cost, and subagent fan-out.
Claude Code 2.1.196 embeds steganographic markers in its system prompt based on ANTHROPIC_BASE_URL and timezone. How the mechanism works and who it flags.
What OpenAI's GPT-5.6 Sol preview includes: the Sol, Terra, and Luna lineup, benchmark claims, cyber safeguards, and why the launch starts gated.
How a 'fix this code' prompt led the US to export control Anthropic's Fable 5 and Mythos 5, and why security researchers say the ban hurts defenders.
How a US export control directive forced Anthropic to suspend Fable 5 and Mythos 5, what the disputed jailbreak actually was, and why the process matters.
A payments engineer with 10 years of experience says AI eroded his expertise. A close look at which software engineering skills still matter in 2026.
What changed in Claude Opus 4.8: effort controls, dynamic workflows in Claude Code, mid-task system prompts in the Messages API, and unchanged pricing.
DeepSeek prefix caching can lower coding-agent input cost. Learn how Reasonix preserves cache hits, where it helps, and its operating limits.
Python or Rust for AI-assisted coding depends on feedback loops, runtime constraints, ecosystem fit, and the agent's language quality.
Local AI vs cloud AI for app features: choose on-device models for private data transformation, typed outputs, and fewer service dependencies.
How Claude Code ingests git history into the model's context, why a commit mentioning OpenClaw can trigger refusals or extra charges, and how to debug it.
Claude Code quality issues in April 2026 came from reasoning defaults, lost session context, and a prompt constraint. Learn the fixes.
Claude Opus 4.7 improves long-running coding tasks with stronger validation, tool use, vision, and instruction following at Opus 4.6 pricing.
What a data-mined GitHub issue reveals about measuring Claude Code quality regressions from session logs, and why extended thinking depth matters.
Source maps published to npm can expose complete unminified source code. How the Claude Code leak happened and how to audit your packages for .map files.
How the malicious litellm_init.pth in litellm 1.82.8 stole credentials on Python startup, how .pth code execution works, and how to check and remediate.
What Leanstral is, how Mistral's open-source Lean 4 proof agent works, its FLTEval scores against Claude, and why Lean makes pass@k sampling pay off.
Claude Opus 4.6 and Sonnet 4.6 now offer the full 1M context window at standard pricing. What it costs, how recall holds up, and when to use it.
Why experienced developers are split on AI coding agents like Claude Code, and what the old fight over CUDA adoption predicts about how it resolves.
What OpenAI's Department of War agreement covers: classified network deployment, cloud-only hosting, forward-deployed engineers, safety terms.
Claude Sonnet 4.6 migration guidance for choosing thinking effort, testing long-context work, and safely upgrading coding and agent workflows.
A new benchmark finds AI agents violate ethical constraints in up to 62.8% of KPI-driven scenarios, with safety not improving across model generations.
How Anthropic ran 16 parallel Claude agents to build a 100,000-line C compiler: the harness design, test strategy, and where agent teams hit limits.
Vercel's evals found an 8KB docs index in AGENTS.md scored 100% on Next.js 16 tasks while skills peaked at 79%. Here is why passive context won.
Claude Constitution explains how Anthropic uses a public values document in training, where its priorities apply, and what it cannot guarantee.
Will AI replace software developers? Fifty years of attempts, from COBOL to no-code, show why the bottleneck is judgment and what actually changes.
How coding agents work: the tool loop, file operations, prompts, and safety boundaries behind a small but functional coding assistant.
Boris Cherny, creator of Claude Code, runs a near-vanilla setup. When Claude Code customization like CLAUDE.md, hooks, and MCP servers actually pays off.
What both job seekers and talent professionals get wrong about AI-powered resume screening
Job seekers are convinced AI is automatically rejecting their applications, sometimes illegally. Talent professionals insist AI screening features barely exist, and they personally review every candidate. Both are partially right. Both are partially wrong.