[BUG] Opus 4.5 usage limits significantly reduced since January 2026 - most restrictive since Opus 4.5 launch

Can confirm. Max x20 subscriber here, also using Opus 4.5 since November.

My detailed timeline is in #16157 — went from comfortably using my weekly quota over 50-80 hours to burning through it in a fraction of that. Today alone: 8%→19% session usage on trivial work (test runs, commits).

I've submitted debug logs per @whyuan-cc's request (Feedback ID: 7e57e16c-d6a7-45c2-acfe-472b80df376c). Happy to provide more data if needed.

This has been ongoing since Jan 8. At this point we need an official acknowledgment and a timeline for resolution — paying $200/month for Max x20 and getting effectively unusable limits isn't sustainable.

Found 2 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/16270
2. https://github.com/anthropics/claude-code/issues/17083

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

I hope this is simply a mistake; the pre-New Year 2x limits were rolled back (manually or automatically), and something went wrong.

Can confirm as well. Max 5x.

I haven't changed my pattern of usage, and in 1 hour of trivial tasks, I've reached 67% of the session usage.

Even tho I use the Pro plan, it's unacceptable as my current session usage limit increased by 3% by itself. I intentionally monitor the usage limit % via the Claude desktop app, and just by looking at the last answer from Claude code in VScode it increased in a couple of minutes

20x plan, the same

This situation is kind of funny — my limits are getting used up in just a few hours. I have a corporate Premium Seat subscription, but it feels like now the maximum I can work is about two hours, and only around ten hours per week. Something doesn’t seem right.

According to the usage statistics I only used up 11% of my daily limit but 30% of my weekly limit within the last 24 hours. This makes little sense. Is this an actual usage limit change or a reporting bug?

Is this an actual usage limit change or a reporting bug?

We don't know, hope for a bug.

As a 20x Max plan subscriber, I confirm this.

Affects regular pro user on Sonnet as well. (me). Sonnet could get an easy 2 hour session in most of the time, now, I'm luck to get an hour. Gone through half a week of usage Friday night and Saturday morning - and slept 9 hours.

My suspicion is this:

<img width="1278" height="135" alt="Image" src="https://github.com/user-attachments/assets/130b22f3-4981-4615-8efd-99f9423b46a2" />

https://github.com/anthropics/claude-code/blame/0359f24538eae61e0ef6ac6220e8a861811fe1bb/CHANGELOG.md#L110

Background agents (of which there are 3 for every session) that used to start with haiku, now start with opus (one still starts as Haiku but not the other two).

This is one conversation startup (first of the session) in a completely empty directory. I ran /context after the 3% usage was already showing, but for the screenshot to show what was loaded.

<img width="1906" height="787" alt="Image" src="https://github.com/user-attachments/assets/b8b9ea6e-0de7-44ab-abb1-0c06c0fac925" />

Here are the details of three background agents that are auto spun up every time you simply open Claude Code. This is likely done this in preparation for a conversation, but I don't think they should be Opus. Also sometimes I just wish to open Claude to do something simple not start a whole conversation and so these background agents are just a waste.

<img width="948" height="115" alt="Image" src="https://github.com/user-attachments/assets/af5fd5b6-9e82-4f9c-b97c-828986b833a8" />

1.
<img width="1169" height="1588" alt="Image" src="https://github.com/user-attachments/assets/716e85ce-f503-4066-a5bd-07f45e6de5ed" />
2.
<img width="1169" height="1720" alt="Image" src="https://github.com/user-attachments/assets/5946c650-d6bb-4f8f-9a00-57c81f5e7fd8" />
3.
<img width="1169" height="1720" alt="Image" src="https://github.com/user-attachments/assets/6cd7b3fc-7693-483b-9f2f-6bbbeec30d4b" />

Feedback ID: a8262735-8707-4859-9330-5458ca7a01e5

Regarding the subagents issue in changelog:

Context: https://x.com/rezadev1/status/2004241484640371178?s=20
Fix: https://x.com/bcherny/status/2004246795333914874?s=20

Background agents (of which there are 3 for every session) that used to start with haiku, now start with opus (one still starts as Haiku but not the other two).

That is definitely not the reason. I use a flow with an orchestrator, where all subagents do the main work (at least 3 different subagents) and all use Opus. This pattern has not changed for me since November. It could not have affected the limits.

All changes to the limits occurred on Anthropic's server side. This has happened before, and this time the situation may be the same.

P.S. I'm still using v2.0.76 (not "affected" by model changes). Now I used 87% of the limits:

<img width="1483" height="893" alt="Image" src="https://github.com/user-attachments/assets/4cd2f13b-89fe-463a-83b3-e7a124e7d101" />

I think you misunderstood what I was saying. I'm talking about the default agents that are launched at startup. Your ticket said Jan 2026 and that is when the 2.1 versions where released.

I have data. It went from about 1,000 input tokens to 23,000 many of which are cached input tokens which cost more. I reverted back to 2.0.72 and checked. It uses less data on startup.

I'm not saying that is the full story but it is certainly happening. If you upgrade you will likely see even worse usage, but fair enough if that is not the problem being discussed by your issue.

I can confirm this with concrete data from my usage dashboard (CodexBar). I'm also on the Max plan.

The drop in limits is massive, approximately an 8x reduction in effective allowance between last week and this week.

1. Last Week (Jan 5)
I spent roughly $99 over two days, and it only consumed 11% of my weekly limit.

• Calculated Weekly Limit: ~$900/week.
• Projected Monthly Limit: ~$4,000/month.

<img width="512" height="512" alt="Image" src="https://github.com/user-attachments/assets/85e89808-842f-49b9-839d-b29984ba3bf2" />

2. This Week (Jan 12)
I spent just $13.52 today (start of the week), and it has already consumed 12% of my weekly limit.

• Calculated Weekly Limit: ~$112/week.
• Projected Monthly Limit: ~$500/month.

<img width="512" height="512" alt="Image" src="https://github.com/user-attachments/assets/2e32cf0f-2f93-4321-b90d-4e203c895fcf" />

My effective usage limit has dropped from ~$900/week down to ~$112/week.

Yesterday evening, I hit 100% usage:

<img width="1468" height="882" alt="Image" src="https://github.com/user-attachments/assets/aacdfd42-22df-4d6a-a9ed-e3ed609ec779" />

This has never happened before, and yet here we are.

I have the same issue, today, I hit 100% usage

Are they even aware of this issue? Will this even be addressed. This has been going on for over a week now

+1 experiencing the same issue.

My numbers (January 17, 2026):

• Weekly "All models": 38% consumed with identical usage patterns to December
• Extra usage: €20.34 (102% of €20 limit) — exceeded mid-month
• Sonnet-only: 7% in 20 hours

Same codebase, same workflows, same everything. The only thing that changed is Anthropic's server-side limits.

This is NOT about the holiday bonus expiration — current limits are more restrictive than November 2025 baseline.

As of 17/01/26 the bug has been acknowledged by anthropic. The current version (i.e v2.1.9) has solved the previous issue but there exists a new token burning issue that is happening because of compact (either auto or manual)

<img width="1451" height="472" alt="Image" src="https://github.com/user-attachments/assets/0f9e2daf-b954-4723-ae3b-3f38a0e2d064" /> in 12 minutes

I have the same issue. I reverted my VS Code extension back to version 2.0.75

It hasn't been fixed, and there is no "software" solution — you can't upgrade/downgrade your Claude Code, or any IDE extension — it won't help, since the issue is on the Anthropic server side.

It hasn't been fixed, and there is no "software" solution — you can't upgrade/downgrade your Claude Code, or any IDE extension — it won't help, since the issue is on the Anthropic server side.

As it been confirmed by Anthropic?

As it been confirmed by Anthropic?

I have not seen any public confirmation from Anthropic that limits were reduced.

Based on the reports in this thread (same workflows, same codebase, same patterns, but materially higher % consumption), this looks less like a Claude Code client bug and more like a server-side quota or accounting change. In other words, even if something is “wrong”, it is happening on the service side, not because we suddenly started using the tool differently.

Also, there is precedent that Anthropic does not publicly confirm or detail quota changes in GitHub issues. For example, in a very similar case (#8449) there was no explicit “yes, we reduced your limits” statement, so I do not know why we should expect a clearer confirmation now.

Finally, I would not expect an official representative to write something blunt like “yes, we cut your limits and we do not care”. If anything, the best we can reasonably ask for is clarity, either:

• confirm this is an intentional policy / allocation change (and update expectations in docs), or
• confirm it is a metering / accounting regression on the server side and give a timeline for a fix.

Until then, the most accurate statement is: no confirmation, but the evidence strongly points to a server-side change.

[1]: https://github.com/anthropics/claude-code/issues/8449 "[BUG] Opus usage limit reached unusually fast with Max 20x plan since Claude Code v2 · Issue #8449 · anthropics/claude-code · GitHub"
[2]: https://github.com/anthropics/claude-code/issues/17084 "[BUG] Opus 4.5 usage limits significantly reduced since January 2026 - most restrictive since Opus 4.5 launch · Issue #17084 · anthropics/claude-code · GitHub"

Max x20 user and literally can't use Opus 4.5 since last Friday 16/01/2026 anymore due to the limits or bug whatever it is. For value for money, definitely there is no value and if it's not fixed I will cancel my entire subscription. Not going to pay $200 per month for a premium subscription only to be capped.

According to this link https://www.reddit.com/r/ClaudeAI/s/NIt60lb8am

!Image

What a bunch of crap, we pay a premium fee and get garbage. This is really pathetic from Anthropic.

this looks less like a Claude Code client bug and more like a server-side quota or accounting change

-- @alex-feel

At the very least it is both. I saw token use on the background agents increase by 11 times! From about 900 input tokens to about 10K input cache tokens (the more expensive kind). Both Opus and that is just by opening Claude Code and not typing a single character.

I was asked by an Anthropic team member to send details, but when I did I got a baffling response that they couldn't see anything being used on their side. However, I sent the same screenshots as above with friggin arrows pointing to the 10K number. I don't know how much more clear I can make it.

I haven't rechecked recent versions because people didn't even acknowledge the original numbers so why bother.

That's right, I can't even get through a single complete task with opus

According to this link https://www.reddit.com/r/ClaudeAI/s/NIt60lb8am

↓

P.S. I'm still using v2.0.76 (not "affected" by model changes). Now I used 87% of the limits

Downgrading won't help; I never upgraded and consistently hit the limits (already twice).

> 根据此链接 https://www.reddit.com/r/ClaudeAI/s/NIt60lb8am ↓ > P.S. 我仍在使用 v2.0.76（未“受”模型变更影响）。现在我已使用了 87% 的额度 降级也无济于事 ；我从未升级过，却始终触及限额（已经两次了）。

I've got it now: Max users have been reduced to one-fifth of their original number, and Pro users have been reduced to half. This is what Anthropic is doing in the first month of 2026

Sharing some data from tracking Claude Code transcripts. Max $100 plan, ~18K messages over 18 days (Jan 9-27).

Cache Behavior - The 5-Minute Cliff

| Gap Between Messages | Messages | Avg Cache Write |
|---------------------|----------|-----------------|
| < 1 min | 16,322 | 4,589 |
| 1-2 min | 632 | 3,591 |
| 2-5 min | 602 | 3,731 |
| 5-10 min | 197 | 70,919 |
| 10-30 min | 106 | 76,275 |
| > 30 min | 104 | 73,868 |

Cache expires at 5 minutes. After that, you're rebuilding context - 16x more tokens per message.

Cold Start Impact

• 407 cold starts detected (2.3% of messages)
• Avg cold start: 73,068 tokens
• Avg warm start: 4,524 tokens
• Extra tokens per cold start: ~68K
• Total wasted on cache rebuilds: ~27M tokens

Session Length Matters

| Session Size | Sessions | Avg Cache Hit % |
|--------------|----------|-----------------|
| 1-10 msgs | 14 | 74% |
| 11-50 msgs | 10 | 88% |
| 51-100 msgs | 9 | 94% |
| 100+ msgs | 32 | 91% |

Longer sessions = warmer cache = better efficiency.

Week-Over-Week: Tokens Per 1%

| Period | Total Tokens | Usage % | Tokens/1% |
|--------|--------------|---------|-----------|
| Last Week (Jan 16-22) | 849M | 56% | 15.2M |
| This Week (Jan 23-27) | 617M | 56% | 11.0M |

Result: -27% reduction in tokens per 1% - same usage %, fewer tokens.

Still early data. Will report back after more billing cycles.

---

Data parsed from Claude Code transcript JSONLs.

Looks like I am getting 1h cache exclusively now... (per Snoop https://github.com/rcrsr/snoop)

Stop says: [how66wnw | 9m 11s | 434 msgs | 582,023 in (43 p / 48,973 cw1h / 533,007 cr / 92% ce) | ~3,290 out | 3 subagents | 76 tools (Bash, Edit, Glob, Read, Task, Write)]

Seems to improve situation considerably (on Max 20x).

Subbed first time, this has been going back and forth since october as I read. Session and weekly limits on top of each other that burns through and barely matters when sub you are on. Not a good look. Def not not renewing this waste.

Can confirm. Max 20x subscriber here ($200/month), renewed on Feb 4, 2026.

Today I hit the usage limit on Claude Code with only 72% session usage shown on the claude.ai usage page(and vs code plugin Claude Code、Claude CLI showed the same). The strange part: Claude Code is completely blocked, but claude.ai web chat continues to work fine at the same time. This contradicts Anthropic's documentation which states all surfaces share the same usage limit.

Additionally, I discovered that Claude Code was previously authenticated via API mode instead of my Max subscription, which may have caused some of my earlier session consumption. After switching to Max Account login (/logout → /login), the plan now correctly shows "Claude Max" — but the limit block persists even after the switch.

Environment:

• Claude Code v2.1.7
• Model: Opus 4.5
• OS: Windows
• Plan: Max 20x ($200/month)

This is unacceptable for a $200/month plan. We need an official response from Anthropic.

Well... it seems those that had their reduced usage tokens might have been an a/b test for Opus 4.6...??

Same issue