The whole AI industry in 5 minutes a day — no noise

Weekly digest

Codex Rewrites Its Own Prompt Every Turn, Claude Code's Boost Runs to August, NVIDIA's Benchmark Swings Wildly

Aug 16, 2026 — Aug 22, 2026

One GitHub issue this week put a number on a problem a lot of teams have felt but not measured: a Codex CLI session on AWS Bedrock logged 6.7 million cache-write tokens and zero cache hits, because the CLI rewrites its stable instruction prefix from scratch on nearly every turn. That's not a capability story — it's an invoice story, and it sets the tone for this edition. Anthropic, meanwhile, is still calling its 50% weekly usage boost for Claude Code 'limited-time,' even as it extends the deadline again, this time to August 31, 2026 — a reminder that weekly limits, not model quality, are what actually gate agentic coding work. And NVIDIA open-sourced SkillEvaluator to bring rigor to agent-skill claims, only to find its own verified skills ranging from a 77% token savings to a 120% token bloat. Three different stories, one shared lesson: the tooling around agentic coding is outrunning the accounting for it.

• •Codex CLI on Bedrock rewrites its full instruction prefix on nearly every turn, so a fixed cost becomes a recurring one — one session logged 6.7 million cache-write tokens and zero cache hits, according to a GitHub issue report.
• •Anthropic has extended Claude Code's 50% weekly usage boost for at least the second time, now running through August 31, 2026 per its own support documentation, with no stated plan to make it permanent.
• •NVIDIA open-sourced SkillEvaluator to score agent skills objectively, but its own catalog shows results ranging from a 77% token-use reduction to a 120% token-use increase — 'verified' isn't yet the same as 'performs well.'
• •Qwen 3.8 27B matched GPT-5.6 Luna on Artificial Analysis's Intelligence Index this week, narrowing the gap between a small open model and the largest frontier systems to a single point.