[Bug] Anthropic API: Credit consumption rate increased 3-5x after January 1st

Exact same problem with my Pro plan. Same issue, same dates, same usage pattern, same behavior.

Same. It's clear that it has 5x'd since Oct/Nov the last time I used it. I didnt bother using it AT ALL over the entire month of Dec and into Mid Jan. I come back from holidays and get greeted with a Claude Code session that lasted about 13 minutes before it rudely said I had reached the usage max and all I had it doing was summarizing the NFL Playoffs outlook with some Football Fantasy odds and what not. I didnt even get a CHANCE to do any coding that day with claude because I couldnt even use it :/ Once it reset, I tried to use it again, this time for coding, and it failed MISERBLY. It was completely unusable because of the limits and I pay 20 bux for this. I had to use Codex to finish it because Codex doesnt have limit problems with the $20 account.

It's basically not possible to use Claude Code Pro anymore. 1 prompt is consuming 5 hour window quota and always letting unfinished job

I totally agree, the limit is reached with very little actions performed

This is ridiculous. Pro plan. Fix this and comp everyone a free month or two.

I'm really pissed off. I thought this was a "me" problem because I was scanning some large files, it reached my monthly limit (what never happened before doing a lot longer usages), and I added "extra credit" to start with another task. Using small files, small prompts and a new session, in 20 minutes it has eaten nearly 10€ of credit.

Ok, if this serves of any help. I'm seeing something. First of all, web client does not have this problem, I have uploaded the large files and the consumption is a tenth of the one in Code.

Next, I have requested Claude to instead of performing himself the tasks, to use an agent to do the tasks, and the consumption is minimal. I know is not the same because of thinking and token price, but is not possible that the same task takes 6-8x more.

And the final thing that I think is the key, the context size skyrockets, even a simple answer with a couple of lines increases the context size in a 1/2%, supposedly the context is 85k, so according to that a couple of sentences is using 800 to 1600 tokens!!!

And yes, I have taken into account thinking, I have it always expanded as I review the reasoning. As an example, this has consumed a 3% of context:

The user is asking if the fix I made is generic enough to handle INT_TYPES|BOOL correctly, or if it was just a specific fix for NUMERIC|STRING.

Let me look at the code I wrote. The expand_type function now:

Parses the input by splitting on |
For each token, calls add_group_types()
add_group_types() checks if the token is a known group (NUMERIC, INT_TYPES, ALL_TYPES) and expands it, otherwise adds it as a single type
So for INT_TYPES|BOOL:

Split by | → ["INT_TYPES", "BOOL"]
Call add_group_types(out, "INT_TYPES") → adds BYTE, SHORT, INTEGER
Call add_group_types(out, "BOOL") → adds BOOL (as single type)
Result: BYTE, SHORT, INTEGER, BOOL - which is correct!

So yes, it's generic. Any combination of groups and single types will work.

Sí, es genérico. El nuevo código:

Divide por | → obtiene tokens individuales
Para cada token llama a add_group_types()
add_group_types() expande si es grupo conocido (NUMERIC, INT_TYPES, ALL_TYPES), sino añade como tipo simple
Así que INT_TYPES|BOOL → BYTE, SHORT, INTEGER, BOOL ✓

También funcionaría NUMERIC|STRING|BOOL → BYTE, SHORT, INTEGER, REAL, STRING, BOOL

Another example, this has eaten a 10% of the context. The file was already in context (you can see that it does an edit directly, not a read) and only a line was removed:

The user wants me to remove result_varies from the generator tool. I need to find and remove the line that emits it.

Edit [G:\source\pico\BasicComputer\BasicEngine\LanguageSpecs\SpecsBuilder\main.c](<removed for privacy>)
Removed 1 line
Done. I removed the line that emits result_varies. The user will remove it from the struct definition.

Listo, eliminado del generador.

How is possible that this uses 8500 tokens???

I still DO NOT believe that this is a claude code issue. I just had a 7 minute "deep research" task (opus 4.5) on the website take the ENTIRE 5 HR usage for the "pro" account type in about 10 minutes and it didnt even finish. This appears to transcend the cli and is also affecting the website models as well. This is completely abnormal as deep research use to take MAYBE 1-2%, sometimes 3-5 if it took a long time (45+mins). So not only are my claude code sessions only able to last a few minutes, website usage is now in the same boat. This seems to have started sometime around Jan 1st. Someone F'd up the holiday rollback that is 100% clear now. I have reports from Aug 2025 that ran FAR longer and didnt take up nearly as much usage.

Ok, the "mistery" is solved, Anthropic is screwing us unintentionally, but still screwing us.

Today I started a new chat, and without telling Claude anything he already knew the content of some files in a project that I was working on. But not the resume that is supposedly to have in memory about what we have worked on, no, the damn full document!

It seems that something is injecting documents we have used in the past, even with new chats and even different clients (this happened in the Web client).

So, our contexts are skyrocketing because files we do NOT provide to claude in a specific session are being injected to the context even when these are not needed or explicitly named.

Anthropic should be more transparent on what is doing and set an option to enable/disable it.

EDIT: Just to be clear, is not that Claude checked actively the memory, it was injected automatically, here is the proof, Claude himself acknowledging it:

<img width="899" height="661" alt="Image" src="https://github.com/user-attachments/assets/480fc420-b5fa-4b41-a205-852b395a9137" />

BTW, this also explains why is even worse in Code. When you work in a project with big files, even when you start a new conversation, as the info about the project and the current open file is sent, the automatic system detects it and injects again and again the big files.