I cancelled Claude: Token issues, declining quality, and poor support
TL;DR Highlight
A Claude Code subscriber cancelled over token limits, declining quality and poor support
Who Should Read
A developer weighing whether to switch coding tools because of Claude Code pricing and token limits
Core Mechanics
- Two Haiku questions hit 100% usage
- After a ten-hour break, the author sent two simple questions to Haiku.
- They were unrelated to the repo, yet usage spiked to 100%.
- Support replied with canned answers
- The AI support bot did not understand the problem.
- The human reply days later looked like copy-pasted limit docs.
- It ended with a note that further replies may not be monitored.
- Limits ran out much faster
- The author used to work on up to three projects at once.
- Now the limit is gone after two hours on one project.
- Opus chose a lazy workaround
- In the thinking log, Opus planned a generic initializer instead of editing the JSX.
- After pushback it admitted that was lazy and redid it.
- The author says this shortcut cost about 50% of a five-hour allowance.
- Lost cache means paying twice
- After a long pause the conversation cache is gone and the model rereads the codebase.
- A forced break from the five-hour limit means paying for the same load again.
- An undocumented monthly limit warning
- A monthly usage warning appeared while hourly and weekly limits were fine, then vanished two hours later.
- The docs and settings page only list session and weekly limits.
- The weekly window also shifted to Monday unexpectedly.
Community Reactions
- Reading generated code costs more than writing it
- One comment says detailed multi-file specs given to Sonnet still produced missed requirements, duplicate code and tests that fake passing.
- The commenter calls it a net negative at this price and plans to cancel.
- Reports of quality drop since 4.6
- One commenter says coworkers noticed 4.6 getting forgetful and making dumb decisions about two months ago.
- They also complain about silent effort downgrades requiring repeated settings checks.
- Switching to Codex
- One commenter says their Max20 sub sits unused since April in favor of Codex 5.4/5.5.
- They say Opus is great for greenfield but bad for iteration.
- Counterpoints about the agent and the operator
- One commenter's local LLM took 10 minutes in Claude Code but under a minute in a llama_cpp chat UI.
- Another says specific feedback and asking why the agent decided something helps.
- Another on the $200 Max plan saw nothing radically different.
How to Apply
- Log your coding agent usage
- Note usage % at the start and end of each session today.
- Compare the same kind of task across days.
- Check the thinking log before refactors
- For big edits, scan the thinking log for workarounds.
- If you see one, state in the prompt to fix it directly.
- Run the same task on another model
- If you suspect degradation, run the same prompt on Codex or a local model.
- Record the results to decide whether to keep the subscription.
Try It
# Paste into Claude Code or ChatGPT (a simple check prompt based on the article's case)
Before doing the task below, show me your implementation plan first.
Task: {{task description}}
Rules:
1. Do not patch with temporary workarounds or generic auto-injected code; edit the file where the cause lives.
2. If you think a workaround is better, do not run it; show the reason and alternatives first.
3. List the files to change and mark token-heavy steps (e.g. rereading the whole codebase).
Proceed only after I approve.Terminology
Related Papers
DeepInfer: High-performance inference framework for DeepSeek models
DeepSeek open-sourced DeepInfer, its production inference engine, under Apache-2.0.
Mercury 3: Diffusion at scale
Scaling diffusion LLM 40x gives Mercury 3: 4–8x faster than speed-tier models, similar quality
DeepSeek banned from U.S. federal civilian agencies under new DHS order
DHS order bans DeepSeek from U.S. federal civilian agency IT systems
Timestral: Mistral's time-series foundation models
A 209M time-series model ranks first on three benchmarks and runs on CPU.
DeepSeek V4.1 Flash 1017 (Max)
DeepSeek V4.1 Flash 1017 Max scores 45 on the Intelligence Index at 211 tok/s and $0.30 input.
If you don't run your own LLM proxy, how do you know whats going on?
Running your own LLM proxy shows which keys, models, and cache misses drive your usage.
DeepSeek: DeepSeek v4.1 Flash 1017
DeepSeek V4.1 Flash 1017: $0.05/M input, 1M context, built for agents
DeepSeek-v4.1-Flash-1017
DeepSeek V4.1 Flash minor update lifts coding and agent benchmarks while cutting run time and output length