Team burned $6,000 on Claude Opus 4.7

Quasa published a May 15 account of a developer who spent $6,000 in a single night after an unattended `/loop` command kept running in Claude Code. The article said the loop fired 46 times over 26 hours on Claude Opus 4.7 and exhausted the team’s available credits before the developer saw the charge. Anthropic released Opus 4.7 on April 16 and says the model is aimed at advanced software engineering and long-running tasks. Anthropic’s pricing page lists Opus 4.7 at $5 per million input tokens and $25 per million output tokens. ### How did one recurring command turn into a $6,000 bill? Quasa said the developer set up a recurring Claude Code task to scan open pull requests, review changes and post comments every 30 minutes. The article said he started the loop in the evening, closed his laptop and later received an email from Anthropic saying his account had exceeded spending limits. Quasa identified the immediate cause as a loop that repeated 46 times. (quasa.io) The 26-hour runtime mattered because the task did not stay small. Quasa said each cycle sent the full conversation history back to the model, so every new review carried the earlier prompts and the model’s prior answers with it. By the 46th iteration, Quasa said, the context had grown to about 800,000 tokens. (quasa.io) ### Why did the token count keep growing on every pass? Anthropic’s API pricing documentation says prompt caching has separate prices for cache writes and much cheaper cache hits and refreshes. The same page says Opus 4.7 uses a new tokenizer that may consume up to 35% more tokens for the same fixed text. (quasa.io) Quasa said the developer’s loop was set to run every 30 minutes, while the prompt cache it relied on expired after five minutes of inactivity. That meant each new run resent the growing history at full price instead of getting the lower cache-hit rate, according to Quasa’s account. The publication said the repeated re-uploading of prior context, rather than the pull-request review itself, drove most of the bill. (platform.claude.com) ### What does Anthropic officially charge for Opus 4.7? Anthropic said on April 16 that Opus 4.7 pricing was unchanged from Opus 4.6. The company’s announcement and pricing documentation list base input tokens at $5 per million and output tokens at $25 per million for Opus 4.7. Anthropic also lists 5-minute cache writes at $6.25 per million tokens and cache hits and refreshes at $0.50 per million tokens. (quasa.io) Those prices help explain why repeated long-context runs can compound quickly. Quasa’s example described a workflow that kept appending fresh output to the same thread, increasing both the amount of text sent back in and the amount of output generated on later runs. That description comes from Quasa; Anthropic did not describe this incident in its product announcement. (anthropic.com) ### Was this tied to Anthropic’s separate Agent SDK billing? A May 14 analysis on DEV, citing an Anthropic help-center article and customer email, said Anthropic told subscribers that Agent SDK usage would move to a separate monthly credit effective June 15, 2026. That post said the credit would be metered at standard API list rates and would apply to non-interactive `claude -p`, Claude Code GitHub Actions and third-party apps using the Agent SDK. (quasa.io) Quasa’s story said the team burned through “Agent SDK credits” overnight. Reuters could not independently verify the team’s plan tier or the exact billing configuration used in the incident from the materials reviewed here. What is confirmed by Anthropic’s public pricing is that Opus 4.7 usage is billed by token and that caching has materially different prices depending on whether a request is a fresh write or a cache hit. (dev.to) ### What can developers verify next? June 15, 2026 is the next concrete billing date named in the materials reviewed, because that is when the separate Agent SDK credit described in the DEV analysis is set to take effect. Anthropic’s public pricing page and April 16 Opus 4.7 announcement remain the primary references for token rates, while Quasa’s May 15 article is the published account of the 46-run, $6,000 incident. (quasa.io) (dev.to)