Control Bedrock costs, automate exceptions, use ChatGPT safely in GovCloud

The top AI news of the week

AWS lays out a concrete operating pattern for Bedrock costs. Five phases in a loop: attribute, observe, enforce, optimize, operate. Core is attribution to identities in the CUR 2.0 cost report (line_item_iam_principal). That fixes the usual gateway catch-all role with a session-per-user assumption plus session tags. Budgets and hard blocks stop outliers before they get expensive. Figures and coverage come from the vendor.¹

Worldwide Express automates invoice discrepancies with three specialized agents on Bedrock. Reported are 75 percent fewer active exceptions after six months, four weeks faster time to cash, 1 million US dollars annual savings, and a one-time 10 million more free cash flow. Those are company claims in the AWS report. If you process lots of LTL exceptions, this shows a solid target for agentic flows with human in the loop and later bulk approval.²

ChatGPT Desktop can now run via Amazon Bedrock in AWS GovCloud. Work and Codex route inference through Mantle in us-gov-west-1, governed by IAM and regional controls. You pay per token, with no seat licenses. Zero Data Retention is supported. The guide uses config.toml and requirements.toml, a named AWS profile or bearer token, and local network restrictions to the Bedrock endpoint. Good for IL4 or IL5 environments.³

A trend from practice: vibe coding for games jumped in October according to Simon Willison. Claude Opus 5.5 shows clear gains in 3D, music, and game mechanics. He reports four playable browser games, some with minimal prompts. Astra in his view is not quite there yet. These are anecdotes without independent tests. Published early in the week.⁴

GitHub Octoverse 2026 shows the agent effect in numbers. Merges rose to 220 million per month, commits to 7.38 billion per month, much of that in the second half of the year. The industry breakdown uses LLM-based tagging. These are GitHub figures, not independently verified. Also early in the week.⁵

OpenAI reports new protections for teens in ChatGPT and introduces a public safety report. This covers moderation, earlier interventions, and violence prevention. Context: useful for schools and parents. Details come from the vendor.⁶

Elorian tests visual reasoning with DecoyCount. Frontier models averaged under 50 percent per the paper, Astra at 43.17 percent. Elorian’s Glimpse reached 50.48 percent internally in a single run with held-out images. The comparator models were averaged across three runs. Takeaway: if your use case needs counting in images, validate on your own data.⁷

The week for your company

1) Make Bedrock costs finally auditable. Enable IAM principal data in CUR 2.0, tag identities with team, application, and cost center, and activate these tags for cost allocation. Fix the shared role gap by having your gateway assume a role per user and set session tags, with caching up to one hour. Set up alerts and enforcement before budgets are exceeded. You then see who spends how much for which workload pattern.¹

2) Automate exception handling in finance. Start with three lean agents: invoice match, price variance, accessorials. Store each carrier’s rules in a knowledge base and wire the same APIs people use today. Begin with human approval, document each decision in plain text, and scale to bulk approvals when accuracy is high. Target outcome: fewer open cases and shorter time to cash.²

3) Secure GPT use in agencies and regulated firms. Configure the ChatGPT Desktop app so all calls go through Bedrock in AWS GovCloud. Use a named AWS profile or a bearer token, place config.toml and requirements.toml at the specified paths, and block network access to anything but the approved endpoint. Enable Zero Data Retention. That lets you process sensitive data more locally without opening external services.³

4) Marketing without manual toil: daily SEO report with n8n. Pick a template, enter domain and keywords, done. The workflow bundle writes rankings daily to Google Sheets and emails a focus report on competitors. You get a prioritized task list in the morning.⁸

5) PHP to Java migration reproducible, not ad hoc. Kiro pairs a scanner, an eight-stage playbook, and a parity harness. First discover, then translate. Finish only when the Java port passes against the original. You can start interactively in Kiro and later scale portfolio-wide via AWS Transform.⁹

Try next week

1) Turn on attribution. In your AWS account, enable CUR 2.0 with IAM principal data. Tag IAM identities with team, application, and cost center and activate these tags for cost allocation. Wait 24 to 48 hours for the data to be groupable in Cost Explorer.¹

2) Fix the shared role. In the LLM gateway use sts:AssumeRole per user, set RoleSessionName to the user, and pass session tags. Cache the temporary credentials for up to one hour. That puts Bedrock costs back per user in billing.¹

3) Add control. For a test team define a budget limit with an alarm and a hard block. After a week check grouping by iamPrincipal or team and adjust limits.¹

A good find

Anthropic separates brain from hands in Managed Agents. The session and logging part stays stable. Code execution runs isolated in a sandbox with no direct credentials. That cuts latency and reduces risk, because credentials are attached via proxy and the sandbox can only go where it should. Apply the idea to your own agents. Separate runtime and execution, keep secrets out of the sandbox, limit network targets.¹⁰

Sources

1. Managing Amazon Bedrock costs end to end: A playbook for FinOps and platform teams (aws.amazon.com)
2. How Worldwide Express automated freight exceptions with Amazon Bedrock (aws.amazon.com)
3. Configure the ChatGPT desktop app with Amazon Bedrock in AWS GovCloud (US) (aws.amazon.com)
4. The October inflection point for vibe game development (simonwillison.net)
5. GitHub Octoverse 2026 (simonwillison.net)
6. Community safety report and new protections for teens (openai.com)
7. Can Frontier Models See Past a Decoy? Elorian Tests Visual Reasoning (latent.space)
8. @n8n_io: Featured template: a daily SEO rank tracker that writes its own report 📈. Add your keywords + domain. Every morning it logs your keyword positions to Google Sheets and emails you a (x.com)
9. Build a repeatable PHP-to-Java migration engine with Kiro (aws.amazon.com)
10. How Anthropic ships Claude with Claude: Separating the Brain from the Hands (latent.space)