Opus 4.6 on Max plan: usage seems disproportionately high per prompt

Found 3 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/38345
2. https://github.com/anthropics/claude-code/issues/40881
3. https://github.com/anthropics/claude-code/issues/38350

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

Yeh, im hitting abnormal usage too...
Normally 1x 5 hour limit is 8% for my total week.. During promo it was 5-6%...
Today? Its 14 fking percent... Thats HALF my usage per week...
What the actual fk is going on here???????
Max20 here...
Now im also getting this...
Ive always, for over a year worked in 10+ terminals without a single issue...
Its also not consistent, sometimes its on 7 terminals sometimes on 10... the only thing that is consistent is that it kills ALL MY terminals AND ALL my work.... blanket on every single client...
"API Error: Rate limit reached".... Cant even find info on it...

Found a contributing factor on my end that others should check: my ~/.claude/settings.json was bloated with redundant permission entries. This caused ~100k tokens of overhead on every single prompt just from the config being sent as context.

How to check: Open your settings.json and look at the permissions section. If you've been using Claude Code for a while, it accumulates allow/deny rules over time — many of which are duplicates or overly broad. Cleaning up the redundant entries significantly reduced my per-prompt token usage.

This isn't documented anywhere and likely affects anyone who's been hitting usage limits unexpectedly fast.

Follow-up — still no official response

It's been 2 days since this was filed. No acknowledgment from the Anthropic team — not even a "we're looking into it."

This is a wider issue

Other Max plan users are reporting the same problem:

• #37944 — "Max plan usage limits reported as much lower than they should be"
• #28848 — "Max plan usage limits silently reduced since Claude 4.6 release"
• #3650 — "Unexpected Reduction in Claude Code MAX Plan Usage Limits" (older, closed)

The comment from @TheAuditorTool above confirms this isn't isolated — Max 20x users are seeing the same disproportionate consumption.

What we've done on our end

Following the thread, we've already:

• Audited and trimmed MCP servers (removed unused ones)
• Disabled unnecessary plugins (reduced from 12 to 8 active)
• Cleaned up settings.json overhead

Usage improved marginally, but the core issue remains — Max plan limits feel closer to Pro-level limits, especially with Opus 4.6.

What we're asking for

1. Acknowledgment that this is being investigated
2. Transparency on whether Max plan limits were intentionally adjusted post-Opus 4.6 release
3. Guidance on expected token budget per 5-hour window for Max plan users

Paying $170/month and running out of capacity after a handful of prompts is not sustainable. Even a quick "we're aware and investigating" would help.

Yeah, this usually happens when the visible prompt looks small, but the actual payload being sent each round is much bigger than expected.

Things like CLAUDE.md, MEMORY.md, MCP config, tool manifests, and prior session state can quietly stack up and get resent over and over.

So the prompt itself can look lightweight while the real token load isn’t.

Have you checked whether most of the cost is coming from the actual message vs. everything being attached around it?

Confirming this. Max 20 ($200/mo), Opus 4.6, v2.1.89. April 1: 100% in ~70 minutes after reset.

Usage per prompt has clearly increased dramatically — same workflow that lasted full sessions for months now burns through the entire quota almost immediately.

Full report: #41788
Related: #38335, #38239, #40790, #6457, #41055, #38345, #41174, #41550, #41617, #41663, #41779, #41802

We need a fix for that, Claude Code is virtually unusable for days!! We rely on that and cannot afford to loose even more time!

I've been experiencing the same issue on Max 20 ($200/mo) — rate limit 100% exhausted in ~70 minutes.

After setting up a monitoring proxy using the official ANTHROPIC_BASE_URL env var, I identified two cache bugs as the root cause (#40524, #34629) and measured the impact: cache read ratio dropped to 4.3%, meaning ~20x token inflation per turn. After applying workarounds it stabilized at 89-99%.

Full analysis with per-request measured data, safe workarounds, and community references (including cc-cache-fix): https://github.com/ArkNill/claude-code-cache-analysis

<img width="583" height="360" alt="Image" src="https://github.com/user-attachments/assets/2cea7f11-1d83-4a8b-8301-7f9456da7905" />

Done in ~3 days and i slowed down the last day and i was on .88. and .89...
Zero difference for me... Thats about HALF my usage this week...

This is max20... 200 usd per month... Unacceptable....

Everyone should cancel their sub, even if you intend to resume it in x weeks? Cancel it... If 10-100k of us do it? It will be statistically big enough to force their hand... Its the only way at this point...

Follow-up — precautions (April 2, 2026)

Additional drain accelerators beyond the cache bug:

Avoid: --resume (token bomb), /dream & /insights (background drain), v2.1.89 (unresolved bugs)
Careful with: sub-agents (Haiku = 0% cache read, 317K/31 calls), multiple terminals
Recommended: v2.1.81 fixed + fresh sessions + lean CLAUDE.md + proxy monitoring

Ref: https://github.com/ArkNill/claude-code-cache-analysis

Update (April 2): Much of the perceived Opus 4.6 consumption spike was likely compounded by the cache bug (Sentinel regression in standalone binary). v2.1.90 fixes this — benchmark shows 95-99% cache read in stable sessions, dramatically reducing effective token cost.

Update to v2.1.90 to isolate actual model cost from cache bug overhead: https://github.com/ArkNill/claude-code-cache-analysis

April 3 update: v2.1.91 fixes the cache regression that caused the worst drain. If you are still hitting limits after updating, there are additional unfixed mechanisms: a 200K tool result budget cap, a client-side false rate limiter, and silent context stripping — all confirmed via proxy testing. Anthropic acknowledged peak-hour tightening on X (Lydia Hallie) but stated "none were over-charging you." Measured data and analysis: claude-code-cache-analysis

👍 #45333

The problem is back and worse than ever. Token consumption on max $100 at 20x. I was forced to upgrade to max $200. Very unhappy as this appears to be punishment for max users that attempted to manage token use. I will be experimenting with Codex now as a alternative as Anthropic has broken my trust.

Can one of you number crunching guy check a theory i have....
Does any of the numbers correlate to cache read (the 90% discount) being removed.
Thinking since they said to "clear often", "dont run resume idle chats".... maybe they just stopped discounting it, thats why there is so wild jumps seemingly random...
Would tie into what i saw below earlier today....

Ive noticed that pretty much any prompt with meat on it (i.e more than "hello") costs a consistent 10% on the 5 hour limit but working within that problem space has the normal consumption... switching between sonnet/opus is a death sentence as it reloads entire context for entire chat, costing 10-15% of the 5 hour limit....

Its weird because 50% is normal usage and then consitently on new chats/topic eats 10-15% just to "load the convo" up...and then eats 1% slowly and normally and repeat until limit is gone...

I believe there is couple things happening... they lowered limits for everyone, that is obvs but there is also some bug in calcuöation which they are to incompetent to see is happening and think everyone is just complaining over their intended lowered limits and are ignoring it full stop as "noise"

I have not done the deep dive research on claude token weighting yet, but:I will comment based on my experience. Demand for resources requires prepositioning or a bubble of resources on standby. Depending on how quickly your horizontal scaling can respond to demand, determines how much wasted resources you preposition. Anthropic, Google, xAI, OpenAI and others are all balancing how much infrastructure, training, and inference they need and when. It is messy and the financial operations behind it is why companies fail or succeed in a competitive market. Companies holding customer data like Salesforce use it like a moat to protect them. ~ SeanOn Apr 11, 2026, at 03:31, TheAuditor @.***> wrote:TheAuditorTool left a comment (anthropics/claude-code#40895)
Can one of you number crunching guy check a theory i have....
Does any of the numbers correlate to cache read (the 90% discount) being removed.
Thinking since they said to "clear often", "dont run resume idle chats".... maybe they just stopped discounting it, thats why there is so wild jumps seemingly random...
Would tie into what i saw below earlier today....
Ive noticed that pretty much any prompt with meat on it (i.e more than "hello") costs a consistent 10% on the 5 hour limit but working within that problem space has the normal consumption... switching between sonnet/opus is a death sentence as it reloads entire context for entire chat, costing 10-15% of the 5 hour limit....
Its weird because 50% is normal usage and then consitently on new chats/topic eats 10-15% just to "load the convo" up...and then eats 1% slowly and normally and repeat until limit is gone...
I believe there is couple things happening... they lowered limits for everyone, that is obvs but there is also some bug in calcuöation which they are to incompetent to see is happening and think everyone is just complaining over their intended lowered limits and are ignoring it full stop as "noise"

—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you commented.Message ID: @.***>

More info. I was concerned that I was relying on the quality of Claude logic too much last month. So I built a few safeguards and agents to back stop logic checks. This weekend after I switched to Max 20x, the logic quality is basically gone. My assumption, this is another victim to the oversubscription of infrastructure that Anthropic is dealing with.

So Anthropic was smart. Burn a lot of CAPEX to get customers. Get to IPO as fast as possible. Switch to an investor friendly cash burn rate. But, they have lost me as a loyal customer. As soon as I get out of my code spike, subscription gets downgraded and I will start using Codex.

~ Sean

On Apr 11, 2026, at 12:42, Sean Roberts @.> wrote: I have not done the deep dive research on claude token weighting yet, but: I will comment based on my experience. Demand for resources requires prepositioning or a bubble of resources on standby. Depending on how quickly your horizontal scaling can respond to demand, determines how much wasted resources you preposition. Anthropic, Google, xAI, OpenAI and others are all balancing how much infrastructure, training, and inference they need and when. It is messy and the financial operations behind it is why companies fail or succeed in a competitive market. Companies holding customer data like Salesforce use it like a moat to protect them. ~ Sean > On Apr 11, 2026, at 03:31, TheAuditor @.> wrote: > >  > > TheAuditorTool > left a comment > (anthropics/claude-code#40895) > <https://github.com/anthropics/claude-code/issues/40895#issuecomment-4229271917> > Can one of you number crunching guy check a theory i have.... > Does any of the numbers correlate to cache read (the 90% discount) being removed. > Thinking since they said to "clear often", "dont run resume idle chats".... maybe they just stopped discounting it, thats why there is so wild jumps seemingly random... > Would tie into what i saw below earlier today.... > > Ive noticed that pretty much any prompt with meat on it (i.e more than "hello") costs a consistent 10% on the 5 hour limit but working within that problem space has the normal consumption... switching between sonnet/opus is a death sentence as it reloads entire context for entire chat, costing 10-15% of the 5 hour limit.... > > Its weird because 50% is normal usage and then consitently on new chats/topic eats 10-15% just to "load the convo" up...and then eats 1% slowly and normally and repeat until limit is gone... > > I believe there is couple things happening... they lowered limits for everyone, that is obvs but there is also some bug in calcuöation which they are to incompetent to see is happening and think everyone is just complaining over their intended lowered limits and are ignoring it full stop as "noise" > > — > Reply to this email directly, view it on GitHub <https://github.com/anthropics/claude-code/issues/40895#issuecomment-4229271917>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAJDAOS2ZFXDLXC5M3D27KT4VINGVAVCNFSM6AAAAACXF5EZ4CVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DEMRZGI3TCOJRG4>. > You are receiving this because you commented. >

Context bloat compounds silently — each turn re-sends accumulated tool outputs, progress entries, and metadata. Cozempic strips this automatically so you get more turns per rate limit window. Lossless — nothing meaningful is removed.

pip install cozempic && cozempic init

Feedback welcome.

So put your tool in the middle of all transactions. Is it still man in the middle if Its AI traffic? Asking for a friend. ~ SeanOn Apr 13, 2026, at 12:57, Junaid Q @.***> wrote:junaidtitan left a comment (anthropics/claude-code#40895)
Context bloat compounds silently — each turn re-sends accumulated tool outputs, progress entries, and metadata. Cozempic strips this automatically so you get more turns per rate limit window. Lossless — nothing meaningful is removed.
pip install cozempic && cozempic init
Feedback welcome.

—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you commented.Message ID: @.***>

Closing for now — inactive for too long. Please open a new issue if this is still relevant.