What I'm reading.
What's happening in AI, summed up in two sentences with a link to the original. No padding and no rewriting the story: if the summary is enough, you don't need to click.
The Claude API lets your customer pay for usage from their own account
On November 17 Anthropic opened linked billing in beta: your customer links their Claude organization, sets a monthly spend limit, and on any request where you set billing_policy the cost goes to their account instead of yours. It only works if that organization pays Anthropic directly, and the usage runs in your organization, so their zero data retention and BAA don't apply. If you sell a tool with AI inside, you no longer have to front that cost or pass it on by hand.
Source: platform.claude.comClaude Managed Agents can now run on your own infrastructure
On November 12 Anthropic launched self-hosted Managed Agents: the whole service, including agents, sessions and their history, runs in your cloud or data center, not just the sandboxes. Model calls go to Anthropic or to a provider you choose, such as Bedrock or Microsoft Foundry. Access is requested through your Anthropic account team, and from then on you run the server, the database and the security yourself. If keeping your agents' data in-house was what held you back, there's now a way; if not, the hosted service spares you that work.
Source: platform.claude.comClaude Fable 5.2 reaches the API at a lower price than Fable 5.1
On November 9 Anthropic launched Claude Fable 5.2, with a 1M-token context window and up to 128k output tokens. It costs $8 per million input tokens and $40 per million output, down from $10 and $50 for Fable 5.1. Computer use needs the computer_toolset_20260801 toolset, and the model requires 30-day data retention unless Anthropic expressly authorizes otherwise. If your agents already run on Fable 5.1, check the toolset and retention first; the saving is real, but it takes more than swapping the model name.
Source: platform.claude.comClaude lets you restrict which IP addresses can use your API key
On November 5 Anthropic opened an IP allowlist in beta for organizations in the Claude Console. API requests from an address outside the list get a 403 error, so a leaked key is worth far less. It takes up to 1,000 entries, and the documentation itself notes it doesn't cover Claude on Bedrock or Google Cloud. If your automations call Claude from fixed servers, it's an easy layer to add; before you switch it on, list every network your calls leave from, including jobs that only run once a week.
Source: platform.claude.comClaude adds an invisible watermark to text from three more models
On November 3 Anthropic extended its text watermark to Claude Mythos 5, Opus 4.7 and Sonnet 4.6. Its documentation says it adds no tokens or hidden characters, doesn't change the meaning, and carries no information about you or your organization. On Sonnet 4.6 it may be missing if the request sets temperature, top_p or top_k to a non-default value. If your workflow publishes generated text and you want it marked, leave those three parameters alone.
Source: platform.claude.comThe Claude API stops resending the whole conversation every turn
On November 2 Anthropic opened message threads in beta: the server keeps the conversation and each request sends only the new messages, with a lower and steadier time to first token in long conversations. A thread expires 5 minutes after its last request and is deleted no more than an hour after its last use. It isn't eligible for zero data retention. If your agent loops many times over a long context, it's a way to cut traffic and waiting, but test it first: it's still in beta.
Source: platform.claude.comThe Claude API now accepts requests four times larger
On October 28 Anthropic raised the maximum request size from 32 MB to 128 MB, for both the Messages API and the Token Counting API on its first-party API. That is the limit on the request body, not on the model's context window. If your workflow sends PDFs or base64 images and you kept hitting the size error, a lot more now fits in a single call.
Source: platform.claude.comAnthropic launches an API to see usage and cost per user
On October 29 Anthropic launched the Usage and Cost Analytics API in beta, with one endpoint for token usage and another for cost, for Console and Claude Enterprise organizations. You can filter and group by model, workspace, API key, and even the user or service account behind each request. The older Usage and Cost API becomes legacy, though it keeps working. If you split spend across clients or automations, this gives you a way to measure each one.
Source: platform.claude.comAnthropic hands Agent Skills over to a neutral foundation
On October 22 Anthropic donated the Agent Skills open standard to the Agentic AI Foundation, part of the Linux Foundation, along with the name, the logo, the GitHub organization and the agentskills.io domain. A skill is still a folder with a SKILL.md file, and Anthropic says more than 70 agent platforms already read it. Using it doesn't require joining the foundation. For anyone automating, what you write today no longer hangs on a single vendor, though the roadmap with versioned releases and security guidance is still a proposal.
Source: anthropic.comClaude arrives on Google Cloud with Anthropic's API and Google's billing
On October 22 Anthropic launched Claude Models on Google Cloud (Anthropic managed): the same Claude API, run by Anthropic, but reached from your Google Cloud project. IAM controls access and billing goes through Google Cloud Marketplace. Subscribing creates a new Anthropic organization, separate from any you already have, and Managed Agents and skills are available. If your company buys everything through Google Cloud, you can now use Claude without opening another vendor account, though check your existing discounts first, since they can't be applied to past usage.
Source: platform.claude.comThe Claude API gets a new version and the new SDKs already use it
On October 21 Anthropic released API version 2026-10-21, which replaces 2023-06-01 and is what the new SDKs send by default, such as Python 2.0.0 or TypeScript 1.0.0. It includes breaking changes: temperature, top_p and top_k return a 400 error even at their default value, and so do document blocks inside assistant messages and citations in user messages. If you update the SDK without reading the migration guide, a call that used to work can start failing, so test it before you bump the version.
Source: platform.claude.comThe Claude API lets you identify each end customer
On October 20 Anthropic launched user profiles: you register each of your customers once and send their ID in a header, anthropic-user-profile-id, on every request you make for them. That way, if the safety classifiers act against one customer, the restriction or ban can land on that profile alone instead of your whole organization. If you run a platform or resell access to Claude, one customer who breaks the rules no longer has to take everyone else offline.
Source: platform.claude.comFast mode for Claude Opus 5.5 leaves research preview
On October 15 Anthropic stopped requiring access requests for fast mode on Opus 5.5, and the beta header that used to be needed is now optional, though requests that still send it keep working. On Opus 5 and Opus 4.8 it stays in research preview. If you have a workflow where waiting matters most, like a call with a person on the other end, you can now try it without asking for anything.
Source: platform.claude.comThe Claude API now exports traces to your observability tool
On October 14 Anthropic opened OpenTelemetry export in beta: the API emits a span for every model request and every tool it runs, in both the Messages API and Managed Agents sessions. An admin adds the destination in the console without touching any code, and can choose between sending metadata only or the full content. For anyone automating, you can finally see where time goes or where an agent fails from the dashboard you already use, though delivery is best effort and a span can go missing.
Source: platform.claude.comClaude Managed Agents can now write their own workflows
On October 9 Anthropic opened dynamic workflows for Claude Managed Agents in beta: the agent writes a program that runs other agents in phases and combines what each one returns, for jobs like reviewing hundreds of documents. The server runs it in the background, and a setting on the agent switches it on. For anyone automating, splitting a large job across agents stops being orchestration code you maintain yourself, though it's still in beta and worth testing before you lean on it.
Source: platform.claude.comAnthropic halves the cache-read price on Sonnet 5.5
Since October 7, reading from the prompt cache on Claude Sonnet 5.5 costs $0.10 per million tokens, half of the previous $0.20. Cache writes and every other rate stay the same. If your workflow repeats the same long context on every call, that part of the bill drops by itself, but only if you already use the cache.
Source: platform.claude.comClaude Haiku 5.5 arrives on the API
On October 7 Anthropic launched Claude Haiku 5.5, with a one-million-token context window and up to 128,000 tokens of output. It comes with a breaking change: manual thinking through budget_tokens now returns a 400 error, since the model decides for itself how much to think. If a workflow of yours runs on Haiku 4.5, swapping the model name isn't enough: that call needs a review and a test first.
Source: platform.claude.comAnthropic sets a retirement date for Claude Sonnet 4.5
On September 30 Anthropic announced that Claude Sonnet 4.5 stops working on its API on November 30, and recommends moving to Sonnet 5.5. That leaves two months. If any of your workflows calls that model by name, now is the time to find it and test the replacement, before it fails on some ordinary Tuesday.
Source: platform.claude.comOpenAI switches off the Sora API with no replacement
On September 24 the Videos API and every sora-2 model stopped working. Developers got six months' notice, but the deprecations table, where a migration target usually sits, shows only a dash this time. The lesson holds for any automation: if one piece runs on a rented service, decide in advance what replaces it the day it gets switched off.
Source: developers.openai.comAircall's voice agents don't just answer anymore
Since September 22 they also dial out: upload a contact list, pick an agent, and it carries the conversation from the first question to a booked meeting. English-speaking markets get it first. When a phone provider ships this on every plan, it says something: the voice agent has stopped being an experiment and become one more feature of the phone system.
Source: aircall.ioOpenAI is renting out its agent scaffolding
The Agents API opens the same harness that runs Codex in public beta: long sessions, context compaction and parallel subagents, managed by them, with the environment the agent runs in still your choice. For anyone building automations, what changes isn't the announcement but what stops being worth building yourself — the orchestration plumbing becomes something you rent, at no extra cost beyond the usual tokens.
Source: openai.com