TL;DR
André Lindenberg's post points at a Strands Agents case study (2026-11-09): two Amazon engineers replaced Claude Sonnet 4.6 with Strands Decider, an open decision model, for labeling tickets in a large internal ticketing system into one of 20 categories. By their list-price math, cost fell from about $22,200 to about $500 per million tickets (98%), average call time from 4,215 ms to 897 ms, and total tokens per ticket from 6,724 to 2,010.
The point is a model class, not a single model. A decision model cannot write text. It takes some state and a question with a fixed set of options, and returns a probability per option plus a confidence score, in one forward pass. For closed choices that is all an agent needed, and it is far cheaper than generating JSON with an LLM.
How a decision model works
From the Strands Decider 2B launch post (2026-10-01, Marc Brooker, Mike Chambers, Fabio Nonato de Paula):
Architecture. A pre-trained LLM torso (Qwen3.5-2B) with the LM head removed. A small pointer head (just over a million parameters) scores the hidden state at each option position against the hidden state at an
<answer>position. The torso is fine-tuned with a rank-16 LoRA. The released checkpoint is v19, and the repo documents each iteration, including an earlier slot head that did worse.Outputs. Choices and scores (yes/no, one-of-N, a 0 to 1 score), always from the offered options, with a calibrated confidence per decision. Several questions can share one state in a single request.
Speed. About 115 ms median per decision on an RTX 3090 and about 153 ms on an M3 MacBook for small tasks, growing roughly linearly with input size. The context window is 4,096 tokens.
Evaluation. Accuracy and Brier-score calibration on JevBench's public set; the team reports 3rd of 33 in the 2B class. The class was popularized by TypeSafe AI's Jev.
Openness. Weights, training data and scripts are released (Apache-2.0);
pip install strands-decider.Not for coding, chat, summarization or anything needing multi-step reasoning; it is worse than reasoning models on hard problems by design.
Suggested uses from the vendor: model routing, tool selection, evals, guardrails, memory admission, context management and policy classification, and hybrid agents where an LLM makes the hard calls and a decision model makes the rote ones.
The ticket-labeling swap
Model. Despite the headline, production runs the E4B Gemma 4 build of Strands Decider, which beat the original 2B on their test set, served on SageMaker
ml.g5.xlarge(one A10G, about 19 GiB used, about 640 ms per request).Pipeline. Ticket events go to SQS, a Lambda reads the ticket and calls the SageMaker endpoint, AppConfig holds a switch, CloudWatch carries logs and alarms, CDK defines Beta, Gamma and Prod.
Prompt rewrite. The production label descriptions alone exceeded the 4,096-token window, so each of the 20 labels was rewritten as a one- or two-sentence criterion and sent as the options of one choice question, plus a "none" option. Long tickets drop their oldest comments to fit.
Thresholds. They tuned how confident the model must be before a label is applied.
Three-way rollout switch.
BEDROCK(Sonnet only),DUAL(both in parallel, Sonnet applied, results compared) andDECIDER(decision model only). Rollback was a 45-second config change. The authors call this the part they would copy regardless of model: judging every disagreement on real traffic told them more than offline evals.Gotcha. JSON key order in the state changed output probabilities by up to 0.19. They now sort keys on every request.
| Claude Sonnet 4.6 | Strands Decider E4B | |
|---|---|---|
| Average time per call | 4,215 ms | 897 ms |
| Input tokens per ticket | 6,554 | 2,009 |
| Output tokens per ticket | 170 | 1 |
| Cost per million tickets (list price) | about $22,200 | about $500 (about $250 at full instance use) |
Tradeoffs and caveats
No accuracy numbers. The case study says the decision model "performed just as well" and beat production Sonnet on a separate ticket set, but publishes no accuracy, agreement rate or per-label breakdown. The cost and latency figures are measured; the quality claim is asserted.
Vendor source. Both posts are on the Strands Agents (AWS) blog, about AWS's own model, priced against Bedrock list prices.
The comparison is generous to the swap. Sonnet was running a long system prompt and emitting JSON with reasoning; a smaller or cheaper generative model with a short prompt and constrained output might close much of the gap. No such baseline is reported.
Fit is narrow, and they say so. It works for a closed choice from a known catalog at high volume where latency matters. It does not fit labels that need explanatory text, catalogs that change weekly, or inputs that cannot fit in a few thousand tokens.
Sensitivity to formatting. The key-order effect suggests decisions can move on irrelevant input changes; canonicalize state before calling and before comparing runs.
Self-hosting. Cost depends on instance utilization and burst headroom, and you own the endpoint, scaling and model updates.
Implications
Many "agent" steps are classification in disguise: which tool, which route, is this argument grounded, should this fact be stored, does this violate a policy. Those steps can move from a generative call to a calibrated decision call that is roughly two orders of magnitude cheaper and fast enough to sit in every request path, including before each tool call (the launch post's example asks whether tool arguments are grounded in what the user said and whether calling now is premature, then returns Proceed, Deny, Confirm or Guide).
List the closed-choice steps in your agent (routing, tool gating, memory admission, labels) and their call volume
Move one high-volume closed choice to a decision model behind a three-way switch and judge every disagreement
Set confidence thresholds per decision, and send low-confidence cases to the LLM or a person
Canonicalize inputs (sorted keys, stable ordering) before every call
Report accuracy and calibration alongside cost when you make the swap
Related entries
R026 HybridAgent: another way to stop paying for text the agent does not need, there by reasoning in latent space.
R027 The Tool-call index: hidden token and latency cost of tool-call failures; gating tool calls with a cheap decision model attacks the same cost.
R020 The model proposes, the harness enforces: a decision model is a cheap judge the harness can call on every proposal.
R015 Psychology of programming and deterministic workflows: steady steps become scripts; a decision model is the middle ground for steps that need judgment but not text.
Sources
Amazon saves millions by replacing Sonnet 4.6 with Strands Decider, Carr Onstott, Mehul Jaiswal and the Strands Agents Team, 2026-11-09
Introducing Strands Decider 2B, 2026-10-01
André Lindenberg's LinkedIn post, 2026-11-10