Two harnesses shipped a build that would not work, and fixed it the same night.
Claude Code 2.1.275 went out at 22:33 UTC. Every request failed with 400 … Input tag 'advisor_20260301' when ANTHROPIC_BASE_URL pointed at a proxy or gateway.
No feed. One issue a day. Then you are done.
Claude Code 2.1.275 went out at 22:33 UTC. Every request failed with 400 … Input tag 'advisor_20260301' when ANTHROPIC_BASE_URL pointed at a proxy or gateway.
One issue every weekday. The whole thing, not a teaser.
Three harnesses, seven models, the same thirty tasks. The success rates are indistinguishable. The bills differ by up to 5.1x. We framed this wrongly on Wednesday and the correction is the piece.
One Claude Code release broke two indirect paths at once. The gateway path was repaired in 3.3 hours. The third-party endpoint path took 47.9 — and 43 of those ran after the failure had already been reported three times.
Fable 5.1 cut cache reads to $0.25 per million tokens and promised agentic work up to about 45 percent cheaper. For three days after Claude Code made it the default Fable model, the client re-sent tool-turn context uncached — so the product that was supposed to deliver the discount billed the full rate instead.
All week the desk printed Claude Code's channel lag as policy. The money and credential fixes landed on latest first — and stable only moved after the billable bugs had already shipped past it.
In five weekdays the agent stack moved work into vendor-held sessions, files and repos. The gate that used to sit in a local click has to be declared before the session starts, or it is not a gate.
Claude Code made a classifier the default gate on Friday. The study that sold the change measured what human approval actually does under load — and the default that landed still lets the agent push main.
What it actually takes to point Codex at DeepSeek in production, and where the economics stop working.