Some of your agents keep working after you close the laptop, and ChatGPT dots now text you on their own. Last week Anthropic and OpenAI shipped the plumbing to put a meter and a clock on each one.
I think any agent that runs unattended needs three things before it starts: its own key, so the bill names it; a dollar cap per run; and a time box. A shared key turns the bill into one number nobody can explain, so that's what I'd fix first.
The Signal
Anthropic's Usage and Cost Analytics API went into beta on Oct 29. GET /v1/organizations/analytics/usage returns token and request counts, and GET /v1/organizations/analytics/cost returns cost in USD. Both break down by model, workspace, API key, service tier, or "the user or service account behind each request," and the docs list Claude Managed Agents dimensions like agent_id and session_id. You can rank results ("which API keys spend the most this month?") in daily, hourly, or per-minute buckets.
Check the limits first. It needs the beta header anthropic-beta: analytics-usage-cost-2026-08-24, Console orgs need an Admin API key, and "The Admin API is unavailable for individual accounts." The docs advise giving the collecting service its own key and rotating it. (Claude release notes, analytics API docs)
openai-python v3.42.0 shipped Oct 30 PT (Oct 31 UTC) with new Agents API session settings. Time awareness gives the model "periodic current-time reminders," and it's experimental and off unless you set it. Compaction runs in the background by default, which the SDK docs say may use more tokens.
The cap isn't new. Sessions already accepted spend_control: "Optional spending limit in USD cents. Omission or null creates an unlimited session." So 50000 means $500.00. The SDK types don't say what happens when a session hits it, so test it before you trust it. (v3.42.0 release, session params)
ChatGPT dots, the always-on GPT-6 Astra agents with their own cloud computer, started texting on Oct 28. It's a beta rolling out to U.S. Pro users over iMessage or RCS, and "Your dot will also text you proactively with updates." Dots usage didn't count toward plan allowances for the first month, and OpenAI said it will share usage terms after that. (ChatGPT release notes)
The Stack
- List every agent or scheduled run that works while you're away: cron jobs, n8n flows, dots, Agents API sessions, Claude Code loops.
- Give each one its own API key or service account, so the analytics can name it.
- Set a dollar cap per run where the platform supports it (
spend_control, in USD cents) and a time box in the instructions. Try time awareness on a test agent first, since it's experimental. - Hit the cap once on a throwaway session so you know what happens. The SDK types don't say.
- Once a week, pull cost grouped by API key (Anthropic:
GET /v1/organizations/analytics/cost; OpenAI: the Usage dashboard) and post the top three spenders somewhere you'll read them. - Kill or re-scope any agent whose cost rose and whose output you didn't use.
One decision: pick the one agent that runs most often while you sleep and give it its own key and a cap this week.
Prompt of the Day
Paste this with a list of your agents and scheduled runs:
Below is every agent and scheduled run I have, with its name, what it does, how often it runs, which API key it uses, and any dollar cap or time limit.
Return a table with these columns: agent, shares a key (yes/no), has a dollar cap (yes/no), has a time box (yes/no), suggested cap, suggested time box. State the reasoning for each suggestion as an assumption I must confirm.
Then name the one agent I should fix first, and why.
Use only what I give you. Do not guess prices, model costs, or platform features I didn't list. If you can't tell whether an agent has a cap or a time box, write "unclear" and move on.
If it suggests a cap without stating the assumption behind it, throw out the whole answer and run it again.
Bring your table to AI Freedom Lab, where builders launch companies and web and mobile apps with AI.
Tool of the Day
n8n, featured in the AI Tools Index.
n8n is open-source workflow automation, and self-hosting is free. n8n Cloud bills by monthly workflow executions "regardless of complexity," not per step. Starter is 20 EUR a month billed annually for 2.5K executions, and AI Tools Index lists Pro at 50 EUR. The index's caveat: it takes more technical comfort than beginner no-code tools. (AI Tools Index, n8n pricing)
A build to try: a Schedule Trigger on Mondays, an HTTP Request to your provider's cost endpoint grouped by API key, then the top three keys by spend to Slack or email. One execution a week barely touches your count.
Quick Hits
On Oct 28 OpenAI let eligible ChatGPT Pro and Business subscribers claim free monthly API credits by linking an API organization they own. Pro tiers get $100, $200, or $500 to match their plan, and Business gets $20 per paid Standard seat or $100 per Premium seat, capped at $500 per workspace per month. When they run out with no purchased credits, requests fail with HTTP 429 credit_balance_exhausted, so handle that error in any agent running on them. (OpenAI docs)
On Oct 28 Anthropic raised the max request size for the Messages and Token Counting APIs from 32 MB to 128 MB on the first-party Claude API, so bigger PDF and image batches fit in one call. (Claude release notes)
On Oct 30 Anthropic deprecated fast mode for Claude Opus 4.8, with removal on Nov 30, 2026. If a scheduled job pins it, move to Claude Opus 5.5 before then. (Claude release notes)
Closing the loop
Pick the agent that runs most often overnight. Give it its own key and a cap this week, and read next Monday's receipt. If nothing runs unattended yet, set up all three before anything does, and keep building.
Join AI Freedom Lab, where builders launch companies and web and mobile apps with AI: skool.com/aifreedomlab