Claude Code bug can silently 10-20x API costs

66 points
1/21/1970
2 days ago
by wg0

Comments


esperent

The first couple of days of this week (Claude session week I mean, Max 5x plan) I was doing some heavy data analysis and I hit 30% on day 1, then finished up the job at 50% on day 2.

But this was a lot of work. Analyzing thousands of files and extracting data from them using headless Claude sessions with a concurrency of 20 sessions at once. I have no complaints from those two days.

Then yesterday (day 3 of the week) I did some minor refactoring and bug fixes for 2 or 3 hours. Somehow that took me from 50% to 90%. This morning I finished off some of the bug fixes, maybe 45 minutes of work across ~4 twenty minutes sessions, each of involved identifying the test gap, then using a single subagent to fix it with TDD.

That took me to 100%.

Also, I'm on Asia time so I'm working entirely during off peak hours.

I now have two days to wait before the session rollover and I'm feeling quite frustrated. Whether it's a bug or they just silently reduced session limits, it's not acceptable for $200 a month.

2 days ago

chatmasta

What’s their official policy on purchasing two subscriptions?

2 days ago

pmx

I've done this with Cursor because I have similar issues with inconsistent allowance consumption there. I mostly use Claude models but I've had to disable Opus 4.6 because it just EATS tokens in it's thinking steps.

a day ago

ncr100

Q: Claude bills you on unverifiable work? Is that how these things go?

((Is there any way to audit this aside from reverse engineering? You get less than you pay for?))

a day ago

hulitu

This is a feature.

2 days ago

barazany

[dead]

9 hours ago

lossoth

[dead]

12 hours ago

BernardinoGM

[dead]

a day ago

fxtentacle

Use AI! It'll make you a 10x engineer! (cost-wise) /s

I've recently had the displeasure of Opus 4.6 hallucinating an API. It would have been great if that API had existed, but it did not. Still, it then looped until I manually terminated it while trying to make tests pass. In my case, I used up about $12 of usage in 30 minutes. My guess would be mostly through the (pretty verbose) thinking tokens.

But it's not just Anthropic. I had the same issue with Gemini 3.1 Pro.

a day ago

colwont

[dead]

a day ago

hu3

Some manager just hit their KPI for an entire decade with 1000% revenue increase. /s

a day ago