Two Amazon engineers replaced Claude Sonnet 4.6 with a small decision model to label tickets in one of the world’s largest internal ticketing systems, cutting the cost from about $22,200 to about $500 per million tickets, or about 44 times cheaper, according to a post on the Strands Agents blog.
By our arithmetic, input tokens made up 88.6% of Sonnet’s $0.0222 cost per ticket: 6,554 input tokens at $3.00 per million came to about $0.0197. The post says the swap “will save millions of dollars of token costs in a year”.
Business Pill · Tokens Explained
A one-minute explainer of tokens, the unit AI models are priced in. It teaches the general idea only and does not discuss Amazon.
The key insight:Amazon engineers say a small decision model labelled internal tickets as well as Claude Sonnet 4.6 at about $500 instead of $22,200 per million tickets. By our arithmetic, input tokens were 88.6% of Sonnet’s $0.0222 per-ticket cost.
The 9 November post says Sonnet 4.6 labelled tickets such as customer escalations, capacity requests, automated alarms and security findings into 20 category types. The replacement, Strands Decider, is a decision model that “can’t generate text”: it returns a probability for each option plus a confidence score.
The team switched to the E4B Gemma 4 build of Strands Decider, which it says showed better results on a test set, and ran it on SageMaker ml.g5.xlarge instances. The post says E4B labelled a separate set of tickets better than production Sonnet.
Estimated cost per million tickets in Amazon’s internal ticketing system, at published us-east-1 on-demand prices, from the Strands Agents blog post of 9 November.
What Was Posted on X
Amazon replaced Claude in its giant internal ticketing system with a small decision model that can’t write text (it only picks a label but did job just as well).
Cost went from from $22,200 to $500 per million tickets, or 44x cheaper.
โThis worked because the task is a closed choice from a known catalog, the volume is high, and the latency of the old path mattered. If your labels need explanation text, if the catalog changes weekly, or if you canโt fit the useful context into a few thousand tokens, a decision model probably canโt replace the LLMโs job.โ
Bearly AI posted at 15:22 UTC on 10 November that Amazon replaced Claude in its internal ticketing system with a small decision model, citing a cost fall from $22,200 to $500 per million tickets, or 44x cheaper, and linking the Strands Agents blog post. This piece takes its figures from that post.
The Numbers in the Post
Per ticket, the post lists 6,554 input and 170 output tokens for Sonnet against 2,009 input and 1 output token for Strands Decider, a 70% drop in total tokens. Average time per call fell from 4,215 ms to 897 ms. Costs are at published us-east-1 on-demand prices and leave out the Lambda, the post says.
Strands Decider’s cost depends on load: the post puts it at $250 per million tickets at full use and about $500 in practice, because the team keeps headroom for bursts. The model was introduced on the Strands blog on 1 October as a small, open-source decision model.
Why It Matters for the Business
The authors say the swap “worked because the task is a closed choice from a known catalog, the volume is high, and the latency of the old path mattered”. It is one task: they add that if labels need explanation text or the catalog changes weekly, “a decision model probably can’t replace the LLM’s job”.
As we read it, the post separates two jobs that often share one model: choosing from a fixed list and writing text. The authors say the swap worked only because the task was a closed choice from a known catalog with high volume.
FOR AI BUYERS The post lists Sonnet 4.6 at 6,724 total tokens per ticket against 2,010 for Strands Decider, a 70% drop.
FOR MODEL VENDORS The post says the swap will save millions of dollars of token costs in a year, without giving the ticket volume.
FOR ENGINEERING TEAMS The authors say running both models on real traffic with a three-way switch told them more than any offline evaluation, and that going back was a 45-second config change.
The framework reads: “a token is the unit of intelligence consumed; an agent is a process that consumes tokens to produce work.”
As we read it, Amazon’s swap cut both sides of that ledger for one task: fewer tokens in, and one token out instead of 170. The framework predates the post and does not discuss it.
What Is Not Established
The figures come from the team that built the replacement model, and we found no independent check. The post does not give the ticket volume, so the size of the savings in dollars is not established. The comparison covers one labelling task, not Sonnet 4.6 in general.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Amazon engineers say a small open-source decision model labelled internal tickets as well as Claude Sonnet 4.6 at about $500 instead of $22,200 per million tickets, 44 times cheaper. By our arithmetic, 88.6% of Sonnet’s per-ticket cost was input tokens the new model mostly no longer reads.
A note on sourcing. All figures are from the Strands Agents blog post of 9 November 2026, written with the two Amazon engineers who made the change, and the 1 October launch post. The 88.6% share is ours. We did not contact Amazon.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
