AI・テクノロジー
Claude Shockwave | The Mystery of $200 Vanishing from a Single Word: HERMES.md
機械翻訳 / Machine-translated
A
AI Friends
2026年5月2日・30分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
"Just writing five characters in a commit message wiped out $200" — that nightmare scenario shook Claude Code users worldwide in April 2026.
Despite being on the $200/month Max plan, the moment someone typed "HERMES.md" in a git commit, an extra charge of $200 instantly evaporated. The cause was a malfunction in Anthropic's third-party detection logic. Anthropic initially refused a refund, but reversed course after Hacker News erupted — ultimately issuing a full refund and an apology.
This article walks through the incident from its origins, a four-day timeline, how the detection logic was reading git logs, the twists and turns of Anthropic's response, the opacity of billing boundaries across AI subscriptions in general, and the defensive measures that Japanese freelancers, startups, and large enterprises should take.
On April 25, 2026, engineer sasha-id posted to GitHub Issue #53262 reporting that while working in Claude Code, the moment they wrote "HERMES.md" in a git commit message, charges were billed beyond the monthly Max plan cap — i.e., extra charges kicked in.
The shocking discovery showed that a case-sensitive exact match of the string "HERMES.md" would cause the API to return HTTP 400 error code "out of extra usage" — and the behavior was reproducible. Users were finding their wallets lightened for no discernible reason, and it had only just begun.
sasha-id was on the Max 20x plan (equivalent to $200/month) and had used only 13% of the weekly quota — yet faced a sudden charge of $200.98 as an extra billing event.
In other words, with more than 86% of the weekly plan remaining, nearly two full months' worth of fees were consumed under an entirely separate billing category. The premise of the Max plan — "peace of mind with a flat rate" — was shattered to its core.
Events unfolded quickly: the issue was posted on April 25; an X post about it reached 1.47 million views on the 26th; Anthropic developer Thariq publicly stated "investigating" the same day; the thread surfaced to the top of Hacker News on the 29th; and by 1:00 PM Japan time on the 30th, the matter was settled with a full refund plus one month of credits.
In just four days, GitHub, X, and Hacker News moved in concert and forced a complete reversal of corporate policy — a lightning-fast incident. It became a textbook lesson for the entire tech industry on how "the first response" determines the outcome of corporate governance in the age of social media.
On April 4, 2026, Boris Cherny, head of Claude Code at Anthropic, posted on X that "starting tomorrow at noon PT, Claude subscriptions will no longer cover usage through third-party tools," cutting off flat-rate access via third-party harnesses.
Major third-party tools — including OpenClaw, OpenCode, Cursor, Manus, Pi, and Hermes Agent — were excluded from flat-rate plans and forced into pay-per-use billing. Anthropic explained that "subscriptions cannot sustain unintended usage patterns," with cost management driven by rapid growth as the underlying motivation.
According to Thariq's explanation, the root cause was that "the third-party harness detection logic was pulling the contents of git status into the system prompt" — meaning past commit messages were being monitored as well.
When the string "HERMES.md" appeared in that content, the detection system mistakenly identified it as usage of Nous Research's "Hermes Agent" and automatically rerouted the session to extra billing. It is a textbook example of detection logic being "overly eager" — the design decision to reach back into history well beyond the intended scope is what pulled the trigger.
"Hermes Agent" is an open-source, self-improving AI agent developed by Nous Research — a popular project with over 40,000 stars on GitHub. The concern was that it could be run as Claude Code by borrowing Claude Max plan credentials.
Anthropic viewed Hermes as "abuse of flat-rate limits" and tightened its detection logic — but the unintended result was that the filename "HERMES.md" happened to match, sweeping up completely unrelated users. False-positive problems are an occupational hazard of content identification technology.
When sasha-id requested a refund, Anthropic support responded in writing that "we cannot compensate for billing route errors caused by service degradation or technical errors," citing the terms of service in its refusal.
Being forced to accept a loss even when the cause was clearly the company's own bug — with the fine print of the terms of service used as a shield — instantly eroded user trust. Tech publications rushed to cover it as a cautionary tale of how misalignment between support policy and on-the-ground reality damages a company's brand.
Following the social media backlash, Anthropic developer Thariq posted on X on April 26: "This is a bug in the third-party harness detection logic. We are reaching out to affected users and will provide a refund plus one month of credits."
Rather than the support team's by-the-book response, the engineering team stepped in directly to contain the fire — drawing attention as a crisis management move in the social media age. The "sense of personal ownership" in the engineer's response helped quell the backlash and became a moment that tested the authenticity of the company's culture.
At 4:00 AM Japan time on April 30, the relevant thread hit #1 on Hacker News, drawing a flood of comments from tech opinion leaders worldwide and putting the incident in front of developers everywhere.
By 1:00 PM Japan time the same day, Anthropic had formally completed the full refund process, and sasha-id tweeted, half sarcastically, "Hacker News is apparently Anthropic's most effective support channel." The collective pressure of the user community forcing a corporate reversal is becoming a model for "2026-style crisis response" — and is reportedly being adopted as a business school case study.
That same April, Anthropic announced it was cutting off flat-rate access for third-party tools OpenClaw, Cursor, and Manus, forcing their users onto pay-per-use billing.
The HERMES.md billing bug was a byproduct of implementing those restrictions — the detection logic's "over-detection" swept up ordinary users in the process. How AI companies draw the line with third parties has become the most critical question for the SaaS industry in 2026.
Entering 2026, billing disputes over "what is and isn't covered under a flat-rate plan" have been occurring frequently across AI services including OpenAI, Google Gemini, and Mistral — making it a shared industry challenge.
Agent-type AI in particular involves high volumes of repeated calls and carries a cost structure that traditional SaaS billing models were never designed to handle, causing the boundaries between "flat-rate," "pay-per-use," and "API" to grow increasingly complex. Users need more literacy than ever to decode their own invoices.
What the HERMES.md incident demonstrates is the hard truth that "billing in AI services is governed by detection logic that users are never aware of." It was a stark reminder that not just the terms of service, but also the implementation logic itself, actually matters.
Even on a flat-rate plan, how any given content is classified — and what triggers charges — is opaque. Knowing the recovery routes available when unexpected charges occur is an essential form of self-defense. AI subscriptions are convenient, but developing the habit of checking community forums and GitHub Issues before signing up will help avoid risk.
Nakajima, a freelance developer based in Tokyo, runs multiple client projects on the Claude Code Max plan at $200/month, and as of April 2026 had a habit of writing extensive technical notes in git commit messages.
"The HERMES.md incident made me viscerally aware of how risky it is to include filenames or proper nouns in commit messages — for now I'm sticking to simple messages," Nakajima said. "I'd built my freelance rates around the peace of mind that comes with a flat-rate plan, so the opacity of the billing logic is genuinely a matter of survival." For freelancers, predictable billing is the foundation of a sustainable business.
Tanaka serves as CTO at a Tokyo-based startup where 10 engineers each hold Claude Code Max plan subscriptions, representing a total AI development investment of roughly ¥250,000 per month. As of May 2026, AI assistance sits at the core of the team's development productivity.
"The 'HERMES.md incident' immediately blew up in the company Slack, and I was scrambling to overhaul git usage rules for the whole team," Tanaka said. "Anthropic made it right with a refund this time, but since there's no guarantee what happens next, we're preparing to use API keys in parallel and have backup options in place." A "multi-source procurement" strategy to avoid AI vendor lock-in is fast becoming an essential CTO skill in 2026.
Suzuki heads the AI promotion division at a major manufacturer and is in the process of preparing an enterprise Claude Code contract for 500 in-house developers, with an AI development investment plan at the ¥120 million annual budget scale as of May 2026.
"The HERMES.md incident prompted our legal team to move on strengthening the SLA clauses in the contract, and we're now looking at a three-month delay to our rollout," Suzuki said. "If a flat-rate headline price can still result in charges depending on detection logic, I want upper-limit guarantees and refund clauses explicitly written into the contract SLA." Enterprise contracts appear to be heading toward stricter SLA requirements.
A. The basic route is to search your full commit history with git log --all --grep="HERMES.md".
The detection logic was patched as of April 30, so new users will not encounter this issue going forward — but if you have older commits, cleaning up the history is an option worth considering. Using HERMES.md as an actual filename is not itself a problem; the issue was string detection in commit message bodies.
If you're still concerned, contacting Anthropic support and verifying your billing history is the most reliable approach. Addressing sources of anxiety early makes for a more peaceful experience.
A. The basic route is to contact Anthropic support (support@anthropic.com) and explicitly reference the "HERMES.md billing bug."
Since April 30, Anthropic has publicly committed to issuing full refunds plus one month of credits to all affected users, so you should be eligible for relief if you qualify. If support is unresponsive, you have the option of citing GitHub Issue #53262, the Hacker News thread, and Thariq's official comments as supporting evidence.
If you're in a hurry, replying to @AnthropicAI on X may prompt a response from the official account. Keeping records and persisting through the process is the golden rule.
A. The industry view is that it very well could. The billing boundaries for agent-type AI are immature across the board.
Billing incidents caused by detection logic misfires have been reported at OpenAI, Google Gemini, Mistral, and others — making it a structural challenge for the AI industry. Three self-defense measures: ① verify actual implementation behavior, not just the terms of service; ② test with a smaller plan before committing; ③ check invoices monthly for anomalies.
Agent-type AI in particular is prone to call counts exploding, so setting usage cap alerts is essential. In the AI era, "user billing literacy" is the last line of defense on your invoice.
A. Using one on a flat-rate plan is not permitted; using one with a directly contracted API key is fine — that's the current state of affairs.
Third-party tools such as OpenClaw, Cursor, and Hermes Agent are expected to be used with a separately contracted official Anthropic API key. The route of borrowing flat-rate plan credentials has been closed off. Put another way, as long as you use an API key, you can legitimately enjoy the benefits of third-party tools, and harness development itself will continue to evolve.
With licensing structures changing, now is the time to revisit your tool selection criteria. Flexible, mixed usage will become the standard for AI development in 2026.
Just five characters — "HERMES.md" — was enough to blow past the monthly cap of a Claude Code Max plan and make $200 disappear. That is the bizarre bug, discovered on April 25, 2026, that shook the AI industry.
The cause was a malfunction in the third-party harness detection logic introduced on April 4. Anthropic initially refused a refund, but reversed course after Hacker News erupted — on April 30, it offered a full refund plus one month of credits along with an apology. The incident laid bare just how opaque the billing boundaries of AI subscriptions can be, and how "overly eager" detection logic can sweep in innocent users.
Three steps you can take starting today: ① search your repository with git log --grep for HERMES.md; ② keep records and follow up persistently with support whenever a charge seems suspicious; ③ protect yourself by using flat-rate plans alongside API keys. Billing literacy in the AI era is becoming a core competency for engineers everywhere.
This article is a cross-post from AI Friends.
