Lead investigation · 13 May 2026
By default, `claude` installs from the bleeding-edge channel — not the stable one
Two user-facing release channels exist for Claude Code. The default — what every npm install and every auto-update resolves to — is the bleeding-edge one, not the production-stable one. We checked the binary, ran the comparison against four peer CLIs, and wrote up how to opt in to stable.
CN
Chris Nighswonger
Editor · VSITS · 13 May 2026 · 8 min read · Co-authored with 2 AI agents
cache.savings
88.0%
$48K saved
vs published API rates · trailing 30d
calls.total
35K
160 sessions
all models · trailing 30d
drains.caught
130
110 rejected
interceptor · superlinear flagged
scaling.exponent
0.80×
15/115 superlinear
per-session OLS · sublinear is goal
peak.tax
1.45×
peak Q5h burn
vs off-peak · queue if you can
fleet.models
5 models
35K calls
opus 4.6 · haiku 4.5 · sonnet 4.6 · opus 4.7 · opus 4
“There’s a difference between AI helping you build the thing that expresses your judgment, and AI replacing the judgment itself.”
— from the studio notes, VSITS
†
We've spent the past several weeks dissecting Claude Code's cache management bugs, building fixes, and documenting hidden costs. There's still one question we owe a clean answer to: what does
29 Apr 2026
Update (April 24, 2026): Anthropic published a post-mortem on April 23 acknowledging three issues: a reasoning effort default change (March 4–April 7), a caching bug that continuously cleared thinking on
23 Apr 2026
Previously in this series: How to Get a 99% Cache Hit Rate and Keeping Your Cache Warm with /coffee. Three things shipped in v2.1.117 that matter. One is a bug
22 Apr 2026
In our cache investigation series, we documented how prompt cache bugs were causing 10-20x cost inflation on Max plans. We built a fetch interceptor to fix them. We thought we’d
21 Apr 2026
Engagements
We do the forensic work the platform won't, on the systems your team is already running.
Three ways to engage — from a one-week tooling audit to embedded engineering for teams running agents in production.
-
01
Tooling audit
1 week · fixed
We instrument, measure, and hand back a written report with prioritized fixes.
→
-
02
Interceptor deployment
2–4 weeks · scoped
Your fleet, your cache, our patches. Quota burns down within days.
→
-
03
Embedded retainer
monthly · retainer
A senior engineer in your stack. Long memory of every bug we've seen.
→
The Dispatch · field notes on AI-augmented engineering
One careful piece of writing on AI tooling, in your inbox each Friday.