Field Notes

Keep AI spend in check, answers right in Chrome, simulations from spoken instructions

Today with a cost model for Bedrock, practical Chrome help, and an open simulation tool for training.

Important AI news

Cost per task instead of per token

AWS lays out a practical framework for cost per task. Four drivers set the real bill: tokenizer efficiency, output obedience instead of filler, the efficiency of agent loops over multiple turns, and the cost of enabled reasoning. A higher priced model can be cheaper per completed task. You will measure cost and success per task, not just price per token, and include agent success rate in the math. The numbers in the example come from the provider and illustrate the principle.1

Primer turns spoken lessons into playable simulations

Ethan Mollick’s Generative AI Lab releases Primer as an open system. Teachers speak their material and get a simulation that learners can play through. Paper, prompts, and code are published. This can also work for onboarding and process training if you want to make flows tangible fast.2

Anthropic prepares for an IPO

Anthropic filed a registration with the SEC and points to the prospectus. This is not yet an offer to sell securities. For you this means: more insight into numbers, costs, and infrastructure should follow. Daily life will not change much yet.3

Highlights for your workday

  • Make Bedrock costs visible per user or team and cap them automatically. AWS shows the full path with built-in tools: tag identity, report in CUR 2.0 or Cost Explorer, and limit access above a threshold with AWS Budgets plus a budget action. No custom enforcement loop. Clear requirements in billing access.4
  • Quick answers right in Chrome with the Gemini sidebar. You summarize articles, compare products, and get hard parts explained without switching tabs. This saves clicks in research and shopping.5
  • One prompt in, one YouTube Short out. The n8n template writes a script, generates a voice-over, pulls matching stock footage, and publishes the Short on your channel. You connect n8n to your YouTube account for this.6

Tools and updates

  • LLM gateway for developer teams: put LiteLLM in front of Amazon Bedrock, expose an OpenAI-compatible API, and centrally manage latency, errors, and cost. For small pilots the open source tier is free up to five users. Self-service keys need an enterprise license.7
  • Connect multi-agents across boundaries: the A2A protocol is an open agent-to-agent standard. Agents can discover each other, negotiate tasks, and work together securely, even across organizations. AWS shows a reference with separate identity sources and no shared secrets.8
  • Java teams build autonomous agents with Spring AI and Bedrock AgentCore. Planning, tool calls, and memory follow familiar Spring patterns. The AgentCore runtime handles operations, identity, and guardrails. For existing JVM setups this is a short path from prototype to production.9

Try it in five minutes

1) Open a long technical article in Chrome. Start the Gemini sidebar and ask: “Summarize the text in five points. Name two opportunities and two risks.”5 2) Open a product page and ask: “Compare product A and product B on five criteria and name one clear advantage each.”5 3) Ask for a two-sentence explanation of a technical term from the text for non-specialists.5

Expected result: a crisp list of key points and a usable side-by-side comparison of the products in the sidebar. Check criterion: two spot checks from the summary match the source text, and every named criterion is findable on both product pages.5

A good find

Four days ago Simon Willison showed how Claude Opus 5.5 ports a Three.js 3D effect to vanilla JS/WebGL. The result is not identical, but it works and needs no external library. Handy if you want small embeddable 3D elements or to cut load times. Next step: have your chat model rebuild a tiny example without a framework and compare size and clarity.10

Sources

  1. Optimizing cost per task and proving agent ROI on Amazon Bedrock (aws.amazon.com)
  2. @emollick: 🚨 Our Generative AI Lab is releasing Primer: an agent-driven system where a teacher talks through lessons & gets back a playable simulation for their students, grounded in peda (x.com)
  3. @anthropicai: Our initial public offering: https://t.co/Ah4BkI2loD

Legal disclaimer: Anthropic has filed a registration statement (including a prospectus) with the SEC for the offering to which (x.com) 4. Tokenomics at scale: Track and cap Amazon Bedrock spend per IAM principal with CUR 2.0 and AWS Budgets (aws.amazon.com) 5. Get quick answers from Gemini on any website. (blog.google) 6. @n8n_io: One prompt in, one YouTube Short out. This featured template writes the script, adds an AI voiceover, finds matching stock footage, and publishes the finished Short to your channel (x.com) 7. Building a self-service LLM gateway with Amazon Bedrock and LiteLLM (aws.amazon.com) 8. Build composable, interoperable multi-agent systems with A2A on AWS (aws.amazon.com) 9. Build autonomous AI agents with Amazon Bedrock and Spring AI (aws.amazon.com) 10. The Web Is More Than Enough, by Paul Ford (simonwillison.net)