[ AI Models & Platforms
](https://www.unite.ai/series/artificial-intelligence/)
[Add Unite.AI to your preferred sources on Google](https://www.google.com/preferences/source?q=unite.ai)
Amazon’s (AMZN ) engineering teams have found cases where moving work from hand-written code to AI models blew through project budgets, including $1.8 million spent running Anthropic’s Claude Sonnet on a job that never shipped. Senior engineers laid the cases out for staff at an internal meeting on July 28, 2026 and described the results as “catastrophically expensive,” the Financial Times reported, citing multiple people familiar with the presentation. They told colleagues they are now building automated guardrails to cap what future projects can spend.
The largest example matched author records against product listings on Amazon’s retail site. Spending came in 860% above what that project had budgeted, took five months to surface, and the deployment failed anyway. Two smaller cases were shown alongside it: about $541,000 in unplanned costs on a set of financial auditing tools, and $134,000 on work to improve delivery speeds across Amazon’s logistics network, caught after more than two weeks.
Amazon said it is “experimenting, learning and improving” how it uses the technology, including how it drives cost efficiencies. The company also said that presenting a handful of isolated examples as business as usual misrepresents how its teams work with AI, and that the cases covered a small number of groups inside a corporate workforce of roughly 300,000.
What token billing does to a coding mistake #
Staff were told that mistakes costing almost nothing in conventional systems turn expensive once a model does the work. The reason sits on the price list. Anthropic charges $3 per million input tokens and $15 per million output tokens for Claude Sonnet 4.5 and 4.6, with Sonnet 5 at an introductory $2 and $10 until August 31, 2026 before it moves to the same rates. At Sonnet’s standard input price, $1.8 million buys on the order of 600 billion input tokens.
A retry loop that re-sends the same context, or a batch job pointed at an entire catalogue instead of a sample, throws no exception and crashes nothing. It produces an invoice, and the invoice arrives on a monthly billing cycle rather than in a build log. That distance between the mistake and the number is what turns a bad configuration into a five-month problem.
The unit of billing has been moving too. Anthropic’s documentation notes that Claude 4.7 and later models use a newer tokenizer producing roughly 30% more tokens for the same text, so identical work bills higher on a newer model at the same headline rate. The wider shift from flat seats to metered tokens is the backdrop: Unite.AI covered it when Claude Fable 5 arrived priced as a metered utility and again as the cheap era of always-on Claude agents ended.
The controls AWS already sells #
The levers for this exact problem are published on Amazon’s own Bedrock pricing page. Batch inference runs at half the on-demand rate. A Flex service tier carries a 50% discount to standard pricing, with a Priority tier at a 75% premium for work that needs the speed. Intelligent Prompt Routing costs $1 per 1,000 requests and pushes simpler prompts to a cheaper model in the same family, which AWS says can reduce costs by up to 30% without hurting accuracy.
Anthropic’s side adds two more. A cache read is priced at a tenth of the standard input rate, so re-sending a fixed system prompt or document is the cheap part once caching is switched on. And Claude Haiku 4.5 lists at $1 and $5 per million tokens, a third of Sonnet’s rates, which is the largest single saving available to any team that picked the frontier model by default.
Every one of those is a per-workload decision: which model, whether context is cached, whether the job runs interactively or in a batch window, and what ceiling the account carries. Vendors have begun selling the enforcement layer itself, including Spectro Cloud’s PaletteAI inference launchpad, pitched at enterprises trying to hold token costs down.
What Amazon is changing #
Amazon has already removed one incentive pointing the wrong way. Earlier in 2026 it shut an internal leaderboard ranking employees’ use of its Kiro developer platform after staff inflated their token consumption to climb it, a habit that picked up the name “tokenmaxxing”. The automated guardrails engineers described are the next move, putting budget enforcement into the deployment path instead of a monthly review.
The scale explains why $1.8 million reads as a governance story rather than a financial one. Amazon is expected to spend about $200 billion in capital expenditure across 2026, most of it on AI and data-center infrastructure, against quarterly revenue near $180 billion. The company reports second-quarter results after the close on July 30, 2026, and the capital-spending line will take most of the attention. The measure that shows whether the guardrails work is smaller and internal: how fast a runaway job gets flagged. The author-matching project set that bar at five months, and the automated checks are meant to move it into the build.