Cut model costs, secure AI development in GovCloud, new team features in Open WebUI

Key AI news

Claude Code in AWS GovCloud: AI development in regulated environments

Anthropic brings Claude Opus 5.5 and Sonnet 5.5 to AWS GovCloud. That opens a path for AI-assisted development in constrained environments, up to ITAR. Sonnet 5 has, per the vendor, FedRAMP Class D and DoD IL4/IL5 authorizations. FedRAMP is a US certification framework for cloud services. IL4 and IL5 are DoD security levels. For 5.5 models AWS lists FedRAMP Class D on Bedrock. The concrete model status should be verified. Two Bedrock surfaces are available: bedrock-runtime with Guardrails, Knowledge Bases, Agents, and Logging, and bedrock-mantle with native Anthropic Messages API, Server-Side Tools, background inference, and Projects. Both on a Zero Operator Access architecture. ZOA means operators have no access to customer data. For audit trails you use bedrock-runtime. You need GovCloud access, IAM rights, and enabled model entitlements.¹

Model routing saves budget: 64 percent lower median cost per task

Harrison Chase reports that Open SWE moved model choice into the orchestration harness. Each subtask goes to the cheapest model that passes a quality check. The median task cost dropped by 64 percent. This is vendor reporting in an X post. No independent review. For you that means: add simple quality checks and route routine steps away from the frontier model.²

Replit Desktop for Windows starts as private preview and new ads features

Replit Desktop for Windows runs in a private preview. You can build and run apps locally in isolated sandboxes. A new chat view spans projects. For marketing there is a TikTok Ads MCP that generates content and copy and steers campaigns. Access via waitlist.³

Case study: Qlik Answers scales grounding on Amazon Bedrock

Qlik explains how Qlik Answers delivers grounded answers from knowledge bases, analytics, and documents. Three numbers from vendor and customer accounts. Over 100,000 discoveries by the Discovery Agent since launch. 17,000 technical documents indexed, 75 percent faster answers at Lintech. TouchPoint uses it for 15,000 employees in a regulated setting. Architecture point for you: clear entry, routing, and answer layers, then specialist agents, retrieval on OpenSearch, and Bedrock Guardrails for PII, prompt injection, and grounding checks. Qlik solves regional needs with Bedrock. It falls back to SageMaker in-region when needed. The numbers are not independently verified.⁴

Highlights for your workday

• Cut model costs without losing quality. Define an acceptance criterion per task type and route to the cheaper model first. Send only failures up. You get quick, measurable savings.²
• If you work in a US Gov setting, test Claude Code on Bedrock-runtime. Enable Guardrails and invocation logging. Check the current certification status of the models and keep IAM fine grained. Start with internal automations.¹
• Build your Q&A in clean layers. Entry stays stable, a thin router decides, and the answer layer picks a fast or thorough path. Hang retrieval and analytics behind it as specialists and enforce citations.⁴

Tools and updates

• Open WebUI 0.12.0: real-time voice calls via an OpenAI Realtime Voice model, animated avatars, two-factor login, joint replies in shared chats, and new model control buttons. Handy for service and teams. Requirement: admin rights, API key for Realtime Voice.⁵
• Falcon ASR: new speech recognition tool from TII UAE on Hugging Face. Helps with support recordings and transcription. You need integration and suitable compute.⁶
• Unsloth Notebook: train a compact decision model on Qwen3.5-4B locally with 8 GB VRAM. Suited for simple, structured decisions in processes. You need a GPU and Python.⁷

Try it in five minutes

Own practice idea: test a cheap model against a frontier one for orchestration steps.

1) Pick a recurring task, like ticket triage. Write a short prompt and three clear acceptance criteria. 2) Run the prompt twice in your existing AI chat: first with the cheapest available model, then with the strongest. Check both answers against the criteria. Note hits and, if available, the cost. 3) Repeat with five examples. Derive a simple routing rule: default cheap, escalate on failure.

Result: a rate where the cheap model suffices, plus a first routing rule. Checkpoint: the share of correctly solved cases by the cheap model for this task type. Background: model routing cut median costs sharply in one field report.²

Sources

1. Supercharge regulated workloads with Claude Code and Amazon Bedrock (aws.amazon.com)
2. @hwchase17: agree - most orchestration steps don't need a frontier model

for Open SWE we moved model choice into the harness, so each task goes to the cheapest model that still does the job, (x.com) 3. @replit: This Week in Replit

1) Replit Desktop for Windows is now in private preview, with Microsoft and NVIDIA. Replit can build and run apps right on your PC, each in its own isolated sa (x.com) 4. How Qlik built grounded, enterprise-scale AI with Amazon Bedrock (aws.amazon.com) 5. Open WebUI v0.12.0 (github.com) 6. Introducing Falcon ASR (huggingface.co) 7. @unslothai: You can now train your own Decision model with our free notebook! 💡

Qwen3.5-4B will generate decisions instead of text on just 8GB VRAM locally.

Learn to data prep (state, questi (x.com)