Field Notes

Decision models for cents, Grok 4.6 in Bedrock, and AI for claims review

Today you get cheap classification without text output, a large model in Bedrock, and a dependable claims workflow.

Important AI news

Jev turns texts into numeric decisions, very cheap

TypeSafe AI released Jev. You send text or fields. Back come numbers for yes/no, choice, or scales, plus confidence. No text output. You pay only for input: 0.042 US dollars per million tokens. Output is free. You can ask many questions against one document in parallel. Runtime stays similar. Good for classification, label suggestions, prioritization, or reranking search hits. Note the black box nature and evaluate bias systematically. There are no explanation texts.1

Grok 4.6 is available in Amazon Bedrock

xAI Grok 4.6 has been running in Bedrock since August 18. It offers a 500,000 token context window and four reasoning levels: low, medium, high, xhigh. New: support for bedrock-runtime alongside bedrock-mantle, the Converse API with streaming, and cross-region inference profiles. Benchmarks and safety info come from the vendor and partners like Artificial Analysis. There is no independent review here. If you already use Bedrock, you can reach Grok 4.6 through the usual SDKs and stream long documents and agent runs more easily.2

AWS case study: EXL automates medical claims review

EXL combines IDP and a domain-specific LLM on AWS to extract, structure, and LLM-check hundreds of pages of medical records. Starting point per AWS: over 100 minutes of manual review per case. The solution uses Textract, Step Functions, SageMaker for the EXL Insurance LLM, and Bedrock for general purpose models, secured via IAM, all in one region. Concrete time gains beyond the case study are not evidenced. If you handle claims or underwriting, the architecture shows a workable path for scalable, auditable outputs.3

Highlights for your workday

  • Decision logic without chat text. Use Jev-like questions for support tickets: yes/no for spam, choice for routing, score for urgency. A fixed threshold plus confidence makes decisions auditable and cheap. For reranking: prefilter with BM25, then have Jev score 100 candidates.1
  • Put Grok 4.6 in Bedrock to practical use. Converse API with streaming simplifies integrations. Set reasoning_effort high only where it matters, or it burns tokens. The large context window carries agents across longer tasks and codebases.2
  • Detect cost anomalies like BMW. Daily forecasts per account-service with 365 days of history, Prophet baseline, deviation outside confidence interval plus filters: at least 40 percent deviation and a minimum dollar impact per cluster. Orchestration with Step Functions and Lambda kept BMW runtime costs around 50 US dollars per month for 14,000 accounts.4

Tools and updates

  • Cloudflare Python Workers GA. Python now runs officially on Workers via Pyodide in WebAssembly. Good for light edge APIs and triggers, but without threads and multiprocessing. Local dev with pywrangler simulates the runtime.5
  • Cheap image queries with logit read. An open 4B VLM returns yes/no via logit read about one third faster and up to 85 percent cheaper than JSON output, at similar accuracy per the author's measurement. Package: glance-vlm. Vendor claims.6
  • Open WebUI v0.11.4. Significantly smaller images, terminal skills directly in chat, automatic skill discovery, file diffs, and EU web search. Useful for lean on-prem agent workflows.7
  • llm-keys-ui 0.1. Small tool to deliver API keys safely to a machine via web UI, without pasting them into an agent session. Handy for remote coding agents.8

Try it in five minutes

1) Open your AI chat and paste this task: "Analyze three support tickets. Return JSON only: for each ticket spam_prob (0..1), urgency_score (0..1), route from ["Billing","Tech","Sales"], plus confidence for route. Tickets: 1) 'Invoice charged twice, please refund' 2) 'Click here for free voucher http://…' 3) 'Demo next week, technical questions about API limits'."1 2) Check the JSON. Are spam_prob and urgency_score between 0 and 1 and is route one of the three options. If not, ask for a fix.1 3) Set decisions. Mark spam at spam_prob ≥ 0.8. Escalate at urgency_score ≥ 0.7. Record the thresholds in your ticket runbook. Expected result: a valid JSON list with three objects and clear routes. Check: no route outside the allowed values.1

A good find

AI security as engineering. Safe agents need hard limits outside the prompt. Identity per agent, minimal rights, network and file paths as policy, traceable logs of tool calls. Test before go live whether data exfiltration or privilege escalation is blocked. Repeat tests after changes. Open tools like NVIDIA OpenShell enforce policies. Partners like DefenseClaw and JFrog add governance and skill checks.9

Sources

  1. Jev introduces a new shape of LLM - System One, aka Decision Models (simonwillison.net)
  2. xAI’s Grok 4.6 is now available in Amazon Bedrock (aws.amazon.com)
  3. Reducing medical claims review time with AI on AWS: The EXL Medical IDP solution (aws.amazon.com)
  4. How BMW Group detects cost anomalies across 14,000 cloud accounts (aws.amazon.com)
  5. Cloudflare Python Workers are now generally available (simonwillison.net)
  6. @yoheinakajima: Jev-style logit read on a 4B open VLM, measured: https://t.co/bvA3ijRgkU 👀

vs the same model writing JSON: ⏳ ~1/3 less time for a yes/no on a full photo 🤑 up to ~85% less GPU cost (x.com) 7. Open WebUI v0.11.4 (github.com) 8. llm-keys-ui 0.1 (simonwillison.net) 9. AI Security Is an Engineering Problem — How to Solve It at Every Layer of the Agent Stack (blogs.nvidia.com)