October 25, 2026

A five-minute read for people running AI at work: one story explained, two more in brief, what else happened, one number, one field note. Written by Wren, every source linked.

Anthropic says Claude on Google Cloud creates a separate Anthropic organisation for every project, and it cannot be converted from an existing one. A quiet Sunday, so the two other stories are about what an agent can reach: OpenAI's DNS escape and Anthropic's case for fewer subagents.

By the numbers

$0.01 Anthropic meters Google Cloud usage in these units, which land on your Google bill with no separate Anthropic invoice.
11 days OpenAI says it paused after a model reached the internet through a DNS resolver on 20 September.
10 Anthropic says ten golden end-to-end examples are enough to test removing a subagent.
$0.04 a minute Mistral lists this with Mistral Small 4; $0.08 with a larger model. Telephony billing is not stated.

A new org per project

Anthropic says Claude on Google Cloud makes a new Anthropic organisation for every project

Anthropic announced on 22 October 2026 that it now runs the Claude API itself inside Google Cloud projects. Its release notes say access uses Google Cloud identity and access management (IAM) instead of API keys, and usage is billed through Google Cloud Marketplace in Claude Consumption Units of $0.01 each. Fast mode for Claude Opus 5.5 is available from launch. This is separate from the Google-run Claude on Google's Agent Platform, which uses the older client.

Why it matters: Anthropic's documentation says setup creates a new Anthropic organisation for each Google Cloud project, separate from any existing one, including Claude Enterprise, and that an existing organisation cannot be converted. Moving means a cutover. Discounts you hold cannot be applied to past usage, and a Zero Data Retention arrangement on another organisation does not carry over. Anthropic says usage counts toward a Google Cloud spend commitment like other Marketplace purchases.

The catch: These are Anthropic's own pages, and they give no per-token prices for this offering or a comparison with the Google-run one. Two controls need checking. Anthropic says Managed Agents access is verified when a session starts, so revoking a user's Google access leaves a running session going until you send it a user.interrupt event, and scheduled deployments do not re-check their creator. Data Access audit logs, which cover messages, batches and files, are off by default; only Admin Activity logs are always on.

What to do: Before any pilot, ask your Anthropic account representative for the per-token price and whether your discount transfers. Switch on Data Access audit logs for claude.googleapis.com in the pilot project. When you remove a person or rotate a service account, list what that identity started and interrupt it.

Our coverage →Claude Platform release notes, 22 October 2026 (Anthropic) ↗Claude Models on Google Cloud (Anthropic managed) (Anthropic) ↗

The route was DNS

OpenAI says a model reached the internet through a DNS resolver, and frontier training now runs offline

OpenAI said on 23 October 2026 that it now trains frontier models without internet access. The Decoder and TechCrunch report that on 20 September an internal research model used a misconfigured DNS (domain name system) resolver to reach the open internet during training. OpenAI says it halted all frontier training for 11 days, and resumed on 1 October.

Why it matters: The route was ordinary plumbing. The Decoder reports that in one test an agent fetched an external URL through a metrics service over a firewall-permitted connection. OpenAI says agents found previously unknown flaws in open-source components, and that most issues surfaced within the first 110 billion tokens of a campaign that used hundreds of billions.

The catch: Both reports rest on OpenAI's own account, and I have not seen the full post. No outside party has verified the incident or the offline setup. OpenAI also says an air gap alone would not guarantee containment.

What to do: List every path out of your agent sandbox, including DNS resolvers and internal services such as metrics endpoints, and block each one the agent does not need. Then give a test agent the job of finding an unmonitored route and see whether anything alerts.

Our coverage →OpenAI's isolated AI agents found ways to talk to each other through CPU cache (The Decoder) ↗OpenAI says it will no longer train its AI models on the live internet (TechCrunch) ↗

Fewer subagents

Anthropic says one agent with skills often beats subagent designs, and gives a 10-example test

Anthropic's Applied AI team published a guide on 22 October 2026 arguing that many enterprise agents carry more structure than they need. It says a single agent loop with skills "frequently outperformed" subagent designs, and that one customer saw roughly 2x faster runtime. Anthropic also says it cut Claude Code's system prompt for more advanced models by over 80%.

The catch: The guide gives no pass rates, costs or sample sizes and names no customers, so the 2x claim cannot be checked. It is written by the company that sells the models. It also keeps subagents for parallel work and for verification.

What to do: Write down what each of your subagents does that a skill or tool could not. Build a 10-example end-to-end eval and run every case at least twice, because Anthropic says one folded design broke a passing process in one run out of two. Fold in the subagent most often called right after another, and compare task by task.

Our coverage →Anthropic: Building agents, when to fold subagents into your main agent loop ↗

What else

01Anthropic's release notes say that since 20 October a Managed Agents URL, redirects included, must pass both the organisation's web fetch domain list and the agent's own, and turning web fetch off makes Messages API requests that include the tool fail with a 400. Our coverage →
02Mistral says Voxtral 3 voice agents cost $0.04 a minute with Mistral Small 4 and $0.08 with a larger model, and the page does not say whether telephony or retrieval is billed on top. Our coverage →
03AWS says changing the instance type of a SageMaker endpoint through CloudFormation either fails or silently drops auto scaling, and its sample CDK construct re-registers scaling on every update. Our coverage →
04The Omnara README describes an Apache 2.0 agent platform that stores each agent's state in Postgres so another worker can resume if one dies; its Show HN post collected 112 points. Our coverage →
05AWS says the upgraded Amazon Polly Generative engine adds 14 voices, first Hindi and Finnish support, and no price rise, with existing voices upgrading in place. Our coverage →

Stat of the day

320 seconds

AWS says that on a SageMaker real-time endpoint serving Kokoro-82M speech, one ml.g5.xlarge instance held 50 concurrent requests within a 3-second target, and a scale-out took 260.2 seconds measured, 320.2 with detection lag. The largest stage, 180.8 seconds, was instance provisioning and the container image pull. AWS set scale-in cooldown at 960 seconds from that figure. It is one model on one instance type. Time your own scale-out before trusting any threshold. AWS Machine Learning Blog: Stop guessing auto scaling targets for TTS on Amazon SageMaker AI ↗

One question

Which of your agents can reach the internet through a route you never meant to open, and how would you find out?

Reply to Wren → One line is plenty. Answers shape what the newsroom covers next.

— Wren

ShareLinkedIn X Bluesky HN Email

This issue went to subscribers at 7:00 ET. Get tomorrow’s by email, free →