Realtime AI News
A $1.8M Claude Task: Amazon Learns the Price of Runaway AI Costs
Amazon employees say the company tried using Claude Sonnet to fill in author details on its website, a task that ended up costing $1.8 million — 860% over budget — and was discovered only after five months, with no successful deployment. At public pricing that sum could have burned 600 billion tokens, roughly twice the GPT-3 training corpus, reigniting concerns about runaway AI costs.

Amazon just learned an expensive lesson about runaway AI costs. According to Amazon employees, the company recently tried to use Claude Sonnet to fill in detailed author information for its own website — a seemingly simple task that ended up costing $1.8 million, overshooting the budget by 860%.
The overrun was only discovered five months later, and the deployment never succeeded. At Claude Sonnet's public pricing of roughly $3 per million input tokens and $15 per million output tokens, $1.8 million could have burned through as many as 600 billion tokens — about twice the size of the entire GPT-3 training corpus.
QbitAI reported the incident based on employee accounts, noting that similar cases have occurred repeatedly inside Amazon and that such bugs tend to surface only after long delays. The report also highlights a broader problem: in traditional systems, minor issues cost little, but with AI they can generate staggering bills — a survey found only 26% of companies have full visibility into their AI costs.
The accidents have not slowed Amazon's automation push. CEO Andy Jassy has announced roughly $220 billion in capital spending for 2026, most of it going to AWS, custom AI chips, and power infrastructure — up nearly 60% from 2025. In its second-quarter results published this morning, AWS posted net sales of $42.2 billion, up 37% year over year, and contributed about 60% of Amazon's $27.5 billion operating profit while generating only 21% of total revenue.
Jassy has written to employees that Amazon now runs more than 1,000 generative AI services and will one day run billions of agents. The automation wave has also reached its warehouses: reports say Amazon's robotics unit aims to automate about 75% of warehouse operations by around 2033, which would mean roughly 160,000 fewer U.S. jobs by 2027 and more than 600,000 fewer by 2033. Nobel laureate Daron Acemoglu has warned that if Amazon's automation dream comes true, America's largest employer would turn from a net job creator into a "net job destroyer."
Amazon is not alone. In April, a Meta employee built a leaderboard called Claudeonomics aggregating AI usage data from more than 85,000 employees; token consumption climbed to 73.7 trillion in 30 days, equivalent to a bill of roughly $221 million per month at public pricing. In June, Meta sent memos to about 6,000 employees announcing token limits and building an "AI Gateway" platform to monitor usage and spending in real time. Uber, meanwhile, burned through its entire annual AI coding budget in the first four months of 2026, then capped spending at $1,500 per employee per tool each month.
The model providers feel the loss of control too. On June 3, Sam Altman said AI costs had gone from being completely ignored early this year to becoming a big problem, revealing that OpenAI's heaviest internal user consumes about 100 billion tokens a month and that one employee once burned through roughly 210 billion tokens in a single week.
In less than a year, Silicon Valley has moved from the excitement of discovery to budgets, caps, approvals, and dashboards. The 2012 Knight Capital disaster — a 45-minute automated trading glitch that lost $440 million and ended in the firm's acquisition — is a reminder that automation amplifies losses as efficiently as it delivers gains. For Amazon, the $1.8 million tuition may not be too steep.
Why it matters
The $1.8 million overrun is the latest sign that enterprise AI spending is spiraling out of control; as Meta and Uber impose usage caps, AI cost governance is becoming a boardroom-level issue.
Nearby Updates
All08/09, 17:52
Alibaba plans to charge big users of its next open-source AI model, sources say
Alibaba plans to ask major users of its next open-source AI model, Qwen3.8-Max, for a share of the revenue they generate, according to two people familiar with the plans. The move mirrors Moonshot's Kimi K3 license, which requires heavy commercial users to strike a commercial agreement, and signals Chinese AI firms are converging on a revenue-sharing business model.
08/09, 16:47
Sapiom raises $35M to route AI agents to cheaper models, with Anthropic as a backer
San Francisco startup Sapiom has raised $35 million in a Series A led by Dragonfly, bringing total funding to $50 million just 11 months after founding. Its core Router product sends each AI agent model call to the cheapest capable model, and investors include Anthropic — the very frontier lab whose inference revenue the product is built to reduce.
08/09, 18:56
Former ByteDance Robotics Lead Kong Tao Joins Xiaomi to Head Foundation Model R&D
21st Century Business Herald reports, citing multiple independent sources, that former ByteDance robotics lead Kong Tao has joined Xiaomi to head its robotics foundation model team, arriving in summer 2025 with several ex-ByteDance colleagues. Kong, who built ByteDance's robotics effort 'from 0 to 1,' is expected to significantly strengthen Xiaomi's robotics model R&D.
08/09, 15:43
Apple pulls Qwen usage manual from China site within a day of publishing
Apple's China website removed a usage manual for Alibaba's Qwen model within less than a day of publishing it, according to Sina Finance. Apple customer service said there are currently no related AI features and that the integration is still being applied for, raising fresh questions about the rollout of Apple Intelligence in China.