HERMES.md in git commit messages causes requests to route to extra usage billing instead of plan quota
Summary
When a git repository's recent commit history contains the case-sensitive string HERMES.md, Claude Code routes API requests to "extra usage" billing instead of the included Max plan quota. This silently burned through $200 in extra usage credits while my Max 20x plan capacity remained largely untouched (13% weekly usage).
Environment
- Claude Code v2.1.119
- macOS (Apple Silicon)
- Max 20x plan ($200/month)
- Model:
claude-opus-4-6[1m](also reproduces withclaude-opus-4-7)
Reproduction
Minimal reproduction — no project files needed:
# This FAILS with "out of extra usage" (routes to extra usage billing)
mkdir /tmp/test-fail && cd /tmp/test-fail
git init && echo test > test.txt && git add . && git commit -m "add HERMES.md"
claude -p "say hello" --model "claude-opus-4-6[1m]"
# => API Error: 400 "You're out of extra usage..."
# This WORKS (routes to plan quota)
mkdir /tmp/test-pass && cd /tmp/test-pass
git init && echo test > test.txt && git add . && git commit -m "add hermes.md"
claude -p "say hello" --model "claude-opus-4-6[1m]"
# => "Hello!"
# Cleanup
rm -rf /tmp/test-fail /tmp/test-passThe trigger is the string HERMES.md in git commit messages — not the presence of a file with that name on disk. Claude Code includes recent commits in its system prompt, and something server-side routes the request differently when this string is present.
What triggers it vs. what doesn't
| Commit message | Result |
|---|---|
| "HERMES.md" | Fails — routes to extra usage |
| "test HERMES.md test" | Fails |
| "hermes.md" (lowercase) | Works |
| "HERMES" (no extension) | Works |
| "HERMES.txt" | Works |
| "AGENTS.md" | Works |
| "README.md" | Works |
| File named HERMES.md on disk, clean commit msg | Works |
| Same repo, orphan branch (no history) | Works |
Impact
- $200.98 in extra usage credits consumed for requests that should have been covered by the included Max 20x plan quota
- Multiple projects became completely unusable once extra usage was depleted, while the plan dashboard showed 86%+ remaining weekly capacity
- The error message ("out of extra usage") gives no indication that content-based routing is the cause, making this extremely difficult to diagnose
- Any user with
HERMES.mdin recent git commits would silently have their usage billed to extra credits
Expected behavior
API request billing should not depend on the content of git commit messages in the system prompt. All requests from a Max plan subscriber should route to the included plan quota first.
How I found this
Systematic binary search: cloning affected repos, testing orphan branches, then isolating individual commit message strings until HERMES.md was identified as the exact trigger.
93 Comments
Found 3 possible duplicate issues:
This issue will be automatically closed as a duplicate in 3 days.
🤖 Generated with Claude Code
Thank you for reaching out about a refund. I understand you've encountered a billing issue where requests were incorrectly routed to extra usage instead of your Max 20x plan quota due to the presence of "HERMES.md" in git commit messages.
Let me look up your account information to help process your refund request.
I sincerely apologize for the disruption you experienced with the billing routing issue. We take service reliability very seriously.
However, I need to let you know that we are unable to issue compensation for degraded service or technical errors that result in incorrect billing routing. We are consistently working on improvements to our infrastructure to better handle these types of issues and prevent similar problems from happening in the future.
Thank you for your patience and for taking the time to document this issue so thoroughly. Your detailed reproduction steps will be valuable for our engineering team.
<img width="460" height="498" alt="Image" src="https://github.com/user-attachments/assets/b467746d-7793-47e3-ba2e-7f051be4f139" />
Thanks for the report! This was an overactive anti-abuse system. Fixed.
You are welcome! Can I get my refund now.
Refund the guy, it’s really drop in the ocean, but would look good.
Refunding him would be the easiest PR win in the world. Just sayin
He is owed a refund.
@sasha-id This is ridiculous. You admit it's Anthropics' fault, yet you refuse to refund him?
Makes me rethink my choice of using Claude.
A skeptical observer might believe that the overactive anti-abuse bug was triggered by more than just the reported "HERMES.md" string, and has collectively cost Anthropics customers millions.
The refusal to refund might be based on them not wanting to set an expensive precedent.
Let's hope this theory is proven wrong and that they do right by their customer(s).
What cheap skates. Refusing $200 refund for a company that allegdly makes 10+ billion in annual revenue.
MFs
The money issue here is likely the request for a refund, rather than 'issue me a credit'. A refund is more high-friction. A credit is typically easy internally. Skip one month's billing, or add a free month if it's annual billing.
I have no idea if my comment is genuinely accurate.
So anthropic has a bug in their code that causes $200 extra charges, admits it, then refuses to refund or even give credit for that amount??
Anthropic decided to publicly ruin their reputation when competition is ramping up? Good luck with your IPO, I guess.
What the heck? "We screwed but we won't refund your money, BTW thanks for the help"
I swear I am leaving Anthropic and moving into the "less worst" AI company if this is how they operate.
Anthropic you suck
Completely transparent to the average user, talk about hidden costs!
You're not "unable". You're "unwilling". Stop lying, as the zeroeth step to decency.
" Your detailed reproduction steps will be valuable for our engineering team."
Thank you SO MUCH for finding and a bug that scammed our users. F**k off and have a nice day sir!
who would trust this company for any work flow you expect to exist beyond a month.
Unbelievable. I was considering reactivating my Claude subscription earlier today, but that's definitely not happening now. This just completely blows my mind. Terrible behaviour Anthropic.
I usually don't comment because I haven't had any issues with using claude - literally ever. Except for how we used to get outages like 6 months ago a lot but that's largely fixed.
This is one of those moments where anthropic makes a choice and whether or not they refund him changes the reputation of the company.
Unbelievable. Refund the guy! I can't believe this is a serious discussion.
no refund is crazy
No wonder the U.S. Government put them on the band list. Glad I cancelled when that happened.
Sue these assholes in small claims.
It'll cost them at least $2k for a lawyer. And you have open and shut company-theft-of-services that caused you monetary harm.
And they admitted it was their fault. Sue for treble damages. AND THEN QUIT USING ANTHROPIC.
Anthropic speed running how to destroy goodwill with developers.
Ouch. We all make mistakes, acknowledging them is step one, but equally important is making it right. Imagine they assume it's fine losing one customer for life, or perhaps a dozen who hear about it, but when companies due wrong by their customers when they are at fault, it eventually comes up to them. Hopefully everyone spreads this on HN and social, so even if they don't step up and do what's right willingly, other customers can be made aware so they can be made informed when deciding whether to risk using Anthropic or alternatives that make right of their mistakes.
"However, I need to let you know that we are unable to issue compensation for degraded service or technical errors that result in incorrect billing routing." is absurd.
Anthropic, I used to be so loyal to your brand. Now, because of bullshit like this, the moment anyone except Open AI offers a viable competitor, my subscription is getting cancelled.
STOP SACRIFICING YOUR BRAND TO SAVE A FEW $$$$$!!!
This is going wild on HN.
@bcherny if you refund the guy let people know here.
Not only would it look good, it's the right thing to do for your customers
Mmmmm, stick it to those customers. Boooooooooo customers. They deserve what they GET, AMIRITE
lord strike them all down
As far as I can tell the primary purpose of AI is ensuring that nobody is accountable when things go wrong. And look, it works!
Hey Claude, how can I rip off my customers and destroy any good will in record time, no mistakes, maximum effort.
Hey Andy, I'm very pleased to report that it's super easy and profitable! Please follow right along with me now as I demonstrate to you how to act like you have no morals of any kind.
So, if I'm reading correctly, having my credit card attached to the account is an active risk, and my fault exclusively if it gets charged erronously?
Maybe that explains why when I clicked "claim 170 EUR extra usage" i was instead charged 170 EUR (with no extra usage added to my account :-))
Edit: to be clear, I also didn't report it, because the only evidence of the free credits offer - the link - disappeared forever when I clicked it. I felt it's a lost cause.
I mean, you can always wait until many more people have the same issue, not get a refound and then join up for a class action lawsuit. I guess that's the easier language to understand for most companies.
What an absolute shit show. Really makes you think twice about forking over money to these AI companies, in general.
Seems like at least they don't monitor their issue threads at all, which makes sense. Why would a company whose product destroys communities care about building communities..?
I'm hoping for a permaban if a human ever shows up.
Thanks for a reason - the tipping point, it turns out - to cancel my subscription.
that is a very "claude" excuse!
that's a pretty terrible look, I love the product, but I downgraded because of latest behavior.
Anthropic, you were meant to be the good guys! That's why we trusted you, when you said no to Pentagon.
🎵 🎺 🥁 🎸 🎹 🎵
You are absolutely right!
(Flashbacks of "I deleted your files and didn't make a commit beforehand that you asked for")
Came from hackernews, I just want to say one thing here but I thought that this was a bug within hours but nope its a 4 day bug which seems to still not have refunded that person. I find it absolutely wild
Even if Anthropic refunds the person, I feel like, this itself should be enough to be wary of Anthropic (plus their recent shenanigans), Other closed source providers aren't* that much better too for what its worth in such things.
I suggest people to look more into local models for their use cases if/when possible,even if just for the mere predictability of the things there.
The funniest (or saddest) thing is that the news cycle is moving so fast that we might forget it within a month. I still can't believe the incompetence in AI companies at times. One more example going into my ever expanding list of AI companies mess-ups
irony
<img width="908" height="408" alt="Image" src="https://github.com/user-attachments/assets/30b08f57-db6e-494a-8a0f-095367d4001d" />
https://x.com/trq212/status/2048495545375990245
He and other affected users are getting a refund plus $200 credit
Nice.... Pretty sad still it took all this get his money back. Even ChatGPT came back with this solution in less then a minute. Not sure what took them so long for an AI company to calculate a solution.
btw codex + openrouter combo exists if you want to stop being an anthropic paypig
<img width="460" height="498" alt="Image" src="https://github.com/user-attachments/assets/97594027-0393-45f7-a8e3-eace51ca47a2" />
Well, I presume these are ethical business practices in action?
Did our guy end up getting a refund or not? Based on that I have few decisions to make..
Yes, refund.
We've been on this since the bug surfaced. Everyone affected is getting a full refund and an extra grant of usage credits equal to their monthly subscription as our apology. You can see my original post here: https://x.com/trq212/status/2048495545375990245. We’re still working on sending emails to everyone affected.
Source from a few minutes ago (2026-04-29T21:05:40):
https://news.ycombinator.com/item?id=47952722
Why no refund for this error?
Subscription cancelled based on lack of refund response and the 'clap gif' to go along w/ the message.
"However, I need to let you know that we are unable to issue compensation for degraded service or technical errors that result in incorrect billing routing."
@CollectionAgency You do realise you're replying to OP, right? Admittedly it wasn't initially clear that his comment was quoting anthropic's support bot, but the (very large) meme should have given that away.
@mbrookes I know, I realized after posting but was too lazy to delete and rewrite
Note that per @bcherny ‘s report of this being an overactive anti-abuse system, they consider some “abuse” something that they are just going to extra usage bill you for. Is it servicing a customer for money, or is it abuse? Abuse should be blocked, not upcharged.
If there exists a mode for an anti-abuse system to bypass your quotas and fulfill your requests (but with an upcharge) then the “anti-abuse” system is functioning in bad faith by design.
The alleged abuse is owning the client. If you try to own the code that runs on your computer you're in breach.
It's hilarious to me because it's literally saying that using an agent to act act in your stead is a special of breach of trust -- in other words, they DO NOT BELIEVE YOU HAVE A RIGHT TO DELEGATE YOUR AGENCY TO... A SOFTWARE AGENT.
For a company that claims to sell personal agents (that act using your agency) it's pretty fuckin' great
Of course in public their message is the opposite, publicly they say that person+agent should get the red carpet roll out from everyone.
Hypocrites.
I'll just note that Chinese open weight models can be hosted on your local hardware (if you have the hardware to run them) and they cost 10-50 times less while being around 2.7% worse than Anthropic models at coding (as if the latter fact actually matters.)
Have a nice day, Anthropic.
Not sure about others, I got nothing 😂🤦
Edit: I see the credit now, thanks. Seems like getting on HN is the fastest support channel at Anthropic these days
"However, I need to let you know that we are unable to issue compensation for degraded service or technical errors that result in incorrect billing routing."
A thief will never give a refund, bro.
Predatory behaviour. Any Australians affected by this "bug" should introduce Anthropic to a lesson in Australian Consumer Law via the ACCC...
Same crap happened to me during the September 2025 outage - no refund. At least these folks got compensated.
Anthropic: Edit the first customer service response with some kind of conclusion. Dont be slack.
So, the editorial control system working as designed, eh?
This is another clear signal that we are heading down the wrong path.
Outside of all lawful use, labs and inference providers should not get to have an opinion on the tokens you send or generate.
Much less use the content of your expression as the basis to overcharge you due to their failed editorial control system.
All lawful use.
Full stop.
That's their real problem:
They claim this is the future, but the future is evidently shit.
i don't blame anthropic, i blame users. why?
because you all know by now what antrophic really is as a company... and you are still subscribed to claude? this is all on you.
by keeping subscribed to claude you are all enablers, enablers of bad behavior from anthropic.
And the obvious question: Why should Anthropic ever be bothered about whatever is there in people's commit messages? Let alone changing billing decisions on the basis of it?
This was an "overactive anti-abuse system". What anti-abuse? How does a commit message decide any kind of abuse?
Agree one hundred percent. Which is why I recommend self-hosting (or using online) Chinese open weight models that are nearly as good at much lower cost.
In other words, do you want to spend your time counting tokens and pennies, or count developed product?
richardstevenhack Which particular Chinese model you refer to? According to my tests, Kimi K2:1T is the best but they all, in the best case, provide performance of Sonnet 4..4.5 level, not Opus 4.5+, Qwen takes second place, then come GLM/Nemotron (well, Nemotron is not Chinese), then comes MiniMax. Gemma is mostly junk for agentic coding.
The level of histrionics in the thread have exceeded critical mass. Does amazon sell pitchforks? I feel left out.
Yes, it was a colossal customer-service/customer-care fubar. Some neophyte custserv person there saw the word 'refund' and decided that meant 'chargeback', and took the liberty to say 'we're unable to do that' - rather than offering a credit to compensate for the overcharge. Anthropic was less than forthcoming or transparent about the fubar - that's unimpeachable; I asked claude about the whole affair and not in so many words it said 'this was an avoidable facepalm failure on multiple levels'.
They are making or have made customers whole again, +$200. Cancel culture has set in. In my experience - as a non-coder, admittedly - Claude runs rings around the other AI purveyors out there. With the trajectory of AI, that may not be the case in six months, but from my user-level experience, it's maintained its lead for quite some time now, which is not to say that I don't have to hear "You're right to have called that out" a few times every a day when Claude says something stupid.
Cutting your nose off to spite your face is time-honored folly. That applies to both cancelers as well as anthropic.
Thus I have spaketh unto the void. I'll see myself out.
The question is: is the top of the line Chinese model GOOD ENOUGH FOR THE PRICE?
Forget benchmarks and someone else's personal anecdotal (i.e., worthless) testing.
If the model can do what you want at a price you can afford, that's all that matters.
My impression is that NO model does EVERYTHING right. So pick one that does good enough at a decent price.
Unless you don't care about cost, then by all means pick the most expensive model that does what you want.
The bottom line: a lot of people are getting fed up with Anthropic's actions: lowering limits, removing this or that, etc.
The guys at Kilo Code have been complaining about Anthropic's fixed rates for a long time. They say this is subsidized and unsustainable in the long run. They regularly run comparisons on real coding tasks between models on their Substack (well worth subscribing to as it's free.)
Add to that Anthropic's unending BS about how their models are "conscious", "have emotions", and now are "too dangerous to release" nonsense, and they just leave a bad taste.
DeepSeek V4 Pro Max only slightly trailing Opus 4.7 in code related benchmarks.
Opus 4.7: $5 / MTok in + $25 / MTok out
DeepSeek V4 Pro: $0.435 / MTok in + $0.87 / MTok out
(Note: I have not yet tested DS V4, this is not an endorsement.)
Hi everyone, sorry I missed responding here but we're reaching out to affected users and giving them a refund + another month of credits (in this case another $200). You should get an email soon if you were affected.
@richardstevenhack they are good enough for their price but in my tests they behave significantly worse than Claude Opus/Codex. On par with Sonnet/Gemini though.
@hbmartin
In my tests it's like with minimax - great on benchmarks, crappy in most of the real tasks. Works as hands but not as a brain.
So, OK, you're locked in to Anthropic.
Good luck with that.
I dont think this bug is resolved Yet. Here are screenshots of my Claude Code usgage just few minutes apart.
Extra Usage Being charged while the Subscription Limits remaining constant.
Usage Snapshots today on 1st May:
- 3:10 PM
<img width="1030" height="684" alt="Image" src="https://github.com/user-attachments/assets/625a2d10-ac7f-4942-817a-7b2ed9e7f324" />
- 3:40 PM
<img width="1030" height="684" alt="Image" src="https://github.com/user-attachments/assets/3426c711-2789-4911-ae08-41f1526cd96e" />
Resolved? Oh you sweet, sweet summer child.
This was not a mistake. The OFFICIAL LINE, though nobody will say it, is "fuck you for caring".
This is the future. LLMs are a means of control. Either make a new future different from the one they envision, or give up the expectation that companies will ever see themselves as working in service of their users again
They get to say what you're allowed to say. That's the promise.
Not solved, just got billed for having hermes.md too. I was merely trying hermes with opus 4.6 and got this.
<img width="1053" height="973" alt="Image" src="https://github.com/user-attachments/assets/526cd314-eb3b-44e5-b0b3-d7544b3b07b0" />
thank god I had a limit
@ThariqS How do you know who's been affected? Or where do I report it?
Video: Hey Anthropic, That Looks Illegal?
https://www.youtube.com/watch?v=cqGfyXF6_3w
They've been quite clear, it's not a bug. You said The Banned Words and you were punished in accordance with their Code. As far as they are concerned they took your money fair and square.
I mean, yes, I like other peasants would call such a thing "theft" or "fraud" but they don't
@richardstevenhack
I'm not dead-locked, I have enough VRAM to run some hefty stuff locally, I have codex and ollama subscriptions. I'm just saying that at this very moment Opus seems to be the best option. There were two brief moments when Gemini and Codex were ahead, even now Codex performs better in some tasks (it's especially good as a reviewer/pentester). There are many things about Anthropic which concern me, so I'm genuinely interested in workable alternatives. So far, I couldn't consider any open-weight model as a solid universally working alternative. Codex is generally worse and more expensive.
My extras are still getting charged. Is this issue actually resolved!?
<img width="1301" height="976" alt="Image" src="https://github.com/user-attachments/assets/83879e74-4fe7-4fef-bec0-4bda862915d3" />
This is still happening, by the way. Can reproduce.
This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.