Dateline: October 28, 2026 | Next update: November 6, 2026
A double issue covering two weeks, and almost nothing in it is a new capability. Anthropic spent the fortnight on governance: who an API request is for, which domains an agent may reach, which endpoint a developer's machine may call. OpenAI cut the price of its computer-use model in half and graduated the Decisions API. Google changed a default that quietly affects what gets stored. GitHub shipped the API that makes model policy manageable at scale — the week after it retired six models.
Two deadlines from the last issue resolved: Copilot retired its six models on October 19 as scheduled. Google's deep-research-pro-preview-12-2025 was listed for shutdown on October 23 but, as of the deprecations page's last update on October 22, had not yet been switched off — treat it as gone and verify before relying on it.
Claude / Anthropic
⚠ New API version, new SDK majors — and a licence change
API version 2026-10-21 is out, with a migration guide. Eight new major releases send it by default: Python SDK 2.0.0, TypeScript 1.0.0, C# 13.0.0, Go 2.0.0, Java 3.0.0, PHP 1.0.0, Ruby 2.0.0, and ant CLI 2.0.0.
Two things here that are easy to conflate and should not be. First, the SDK major versions carry breaking changes independent of the API version — upgrading the library is its own migration, separate from moving to the new API version. Second, and more quietly: the new SDK and CLI releases are licensed under Apache 2.0 instead of MIT. If your organisation maintains a dependency licence allowlist, or produces an SBOM for customers, that is a change your legal or procurement side needs to see before the upgrade lands, not after.
API version: `2026-10-21`, migration guide at `platform.claude.com/docs/en/api/migration-2026-10-21` | New majors sending it by default: Python SDK 2.0.0, TypeScript SDK 1.0.0, C# SDK 13.0.0, Go SDK 2.0.0, Java SDK 3.0.0, PHP SDK 1.0.0, Ruby SDK 2.0.0, `ant` CLI 2.0.0 | Breaking changes in the SDKs and CLI are independent of the API version change | Licence: Apache 2.0, previously MIT | In Claude Code 2.1.307: `/claude-api api-version` moves Claude API code to version `2026-10-21`
Best for: Plan the SDK upgrade and the API version migration as two pieces of work. Raise the licence change with whoever signs off on dependencies.
★ User profiles — enforcement that lands on one customer, not your whole org
You can now tell Anthropic which end customer each API request is for. The point is containment: a safety enforcement action applies to that one customer rather than to your organisation.
For anyone reselling or embedding Claude, this removes a genuinely frightening failure mode — one abusive end user triggering a restriction that takes down the service for everyone else. Profiles expose `enforcement_status` and `enforcement_actions` so you can see whether a customer is restricted or banned, and the profile list can be filtered by `enforcement_status[]` to find open cases. Organization webhooks entered beta alongside it, firing when a profile is restricted or banned and again when that ends — so you learn from an event rather than from a support ticket.
Endpoints: `/v1/user_profiles` | Fields: `enforcement_status`, `enforcement_actions` (added October 21) | Filtering: `enforcement_status[]` on the profile list finds profiles with an open restriction or ban | Organization webhooks: beta, header `webhooks-2026-10-07`, endpoints `/v1/organizations/webhooks`, two event types — a profile is restricted or banned, and that restriction or ban ends; access requires contacting your Anthropic account team | Also October 20: inference hooks left beta for Claude Enterprise organizations
Best for: Any product with Claude behind it and customers in front of it. This is the difference between one account being suspended and all of them.
★ Administrators get an off switch for web fetch
Organisation administrators can now turn web fetch off entirely, or leave it on with allowed and blocked domain lists. The enforcement is hard: with web fetch off, a Messages API request that includes the web fetch tool returns a 400 error, and in Managed Agents a web_fetch call returns an error result to the agent.
The domain enforcement is layered, which is the detail worth understanding. A URL reached by a Managed Agent must pass both the organisation list and the agent tool's own `allowed_domains` or `blocked_domains` — and that includes URLs arrived at through a redirect, which is exactly the hole this kind of control usually leaves open. Web search behaves the same way: the agent only receives results that clear both the organisation and agent-level lists.
Location: Capabilities page in the Claude Console | Web fetch off: Messages API requests including the web fetch tool return 400; Managed Agents `web_fetch` calls return an error result to the agent | Web fetch on: admins set allowed or blocked domains | Managed Agents `web_fetch`: a URL, including one reached by redirect, must pass both the organization list and the agent tool's `allowed_domains` or `blocked_domains` | Managed Agents web search: respects organization web search settings and domain restrictions; the agent receives only results passing both lists
Best for: Regulated environments where "the agent browsed somewhere it shouldn't" is an incident. Note the 400 — switching this off breaks code that assumes the tool is available.
★ Claude on Google Cloud, operated by Anthropic
A third path onto Google Cloud, distinct from Vertex AI. This is the Claude API operated by Anthropic, reached through your Google Cloud project: Google Cloud IAM controls access, and Google Cloud Marketplace handles billing. Fast mode for Opus 5.5 is available on it.
The split is worth being precise about, because "Claude on Google Cloud" now means two different things. Vertex AI serves Claude as a Google-operated model with Google's own lifecycle dates. This new option is Anthropic's own API — Anthropic's release schedule and retirement commitments — with Google handling only identity and billing. If your procurement wants everything through the Marketplace but your engineering wants Anthropic's release cadence, this is the option that gives you both.
Product: Claude Models on Google Cloud (Anthropic managed) | What it is: the Claude API operated by Anthropic, reached through your Google Cloud project | Access control: Google Cloud IAM | Billing: Google Cloud Marketplace | Fast mode: available for Claude Opus 5.5 on this platform | Distinct from: Claude on Vertex AI, which Google operates with its own lifecycle dates
Best for: Teams on Google Cloud who want Anthropic's model lifecycle rather than Google's, without opening a separate billing relationship.
★ A Claude Tag API for Slack settings
Endpoints under /v1/organizations/claude_tag/ read and change Claude Tag settings on Slack channels and workspaces, show which settings apply in a given channel and where they come from, and attach Access bundles.
Beta header: `experimental-claude-tag-2026-10-21` | Endpoints: `/v1/organizations/claude_tag/` | Capabilities: read and change Claude Tag settings on Slack channels and workspaces; show which settings apply in a channel and their source; attach Access bundles | Authentication: an Admin API key created by the primary owner in claude.ai, with the `read:claude_tag` or `write:claude_tag` scope | Not available to Claude Platform (Claude Console) organizations
Best for: Enterprises managing Claude across many Slack channels who have been clicking through settings one channel at a time.
Plans and pricing
No pricing changes were published in this window. The commercial news is structural: a new billing path through Google Cloud Marketplace, and per-customer enforcement that protects an organisation's own access.
No published Claude API price changes between October 16 and 28, 2026 | Opus 5.5: $4/$20 per MTok | Sonnet 5.5: $2/$10 per MTok, cache read $0.10 | Haiku 5.5: $0.10/$0.50 per MTok up to 100K-token prompts, $0.50/$2.50 above | Fable 5.1: $10/$50 per MTok | New billing path: Google Cloud Marketplace, via Claude Models on Google Cloud (Anthropic managed) | No model retirements announced this period
ChatGPT / OpenAI
★ GPT-6.1 Terra halves the price of computer use
OpenAI describes Terra as a fast, efficient model for visual reasoning, precise computer use and clear writing. Standard pricing is $1 per million input tokens and $5 per million output, with cached input at $0.05 — which OpenAI puts at 50% below both GPT-5.6 Terra and GPT-6.1 Sol.
Computer use is the expensive kind of agent work: it reads screens, so it burns input tokens at a rate that text workflows never approach. Halving the input price changes which processes are worth automating that way — a task that cost too much to run against a GUI last month may not this month. Watch the long-context rule, though: prompts over 272K input tokens are billed at 2× input and cache rates and 1.5× output for the whole request, and screen-reading prompts get long quickly.
Model: `gpt-6.1-terra` | APIs: `v1/responses`, `v1/chat/completions` | Standard: $1 input / $0.05 cached input / $5 output per 1M tokens | Cache writes: $1.25 per 1M with 30-minute retention, $2.50 with 12-hour retention | Batch and Flex: half the standard rate — $0.50 input, $2.50 output | Fast mode: $2 input, $10 output for short context | Long context: prompts over 272K input tokens priced at 2× input and cache rates and 1.5× output, for the full request | OpenAI's comparison: standard API prices 50% below GPT-5.6 Terra and GPT-6.1 Sol | Also added to GitHub Copilot on October 26, the same day
Best for: GUI-driven automation that was priced out at the old rate. Re-run the numbers, and measure your actual prompt length against the 272K threshold.
★ The Decisions API reaches general availability
Three weeks after entering beta, the Decisions API is generally available, with four changes since public beta: shared input can be cached across requests for prompt-cache discounts, images can be included in individual question instructions, structured data can be passed directly with input_json, and each request now supports up to 400 questions — double the beta's 200.
The shared-input caching is the one that changes the economics. A classifier usually sends the same long instruction block with every item; being able to cache that prefix across requests means you pay for the rubric once rather than per document. Combined with 400 questions per request, a batch classification job that previously meant hundreds of calls can now be a handful.
Model: `gpt-6-luna` | Endpoint: `v1/decisions` | Changes since public beta: shared input cacheable across requests for prompt-cache discounts; images in individual question instructions; structured data passed directly with `input_json`; up to 400 questions per request, up from 200 | Status: generally available, from a public beta that began October 6 and an integration with GPT-Live on October 15
Best for: Bulk classification against a fixed rubric. Cache the rubric, batch the questions, and the per-item cost drops twice over.
★ Twelve-hour prompt caching, and per-project service tier controls
The Responses API gained a twelve-hour prompt cache, which keeps shared instructions, tools and context cached even when requests are hours apart. Cache writes cost 2.5× the input token price. Separately, administrators can now choose which service tiers each project may use, and organisation settings can restrict tiers across all projects.
Twelve hours crosses a working day, which is the point. A support workflow that runs from nine to six previously lost its cache between quiet periods and paid full price to rebuild it; now one morning write covers the day. The 2.5× write premium means this only pays off if the prefix is genuinely reused — cache a prompt you send twice and you have spent more, not less.
12-hour prompt caching: `v1/responses`; keeps shared instructions, tools and context cached across hours-apart requests; cache writes cost 2.5× the input token price | Service tier controls (October 21): choose which service tiers each project can use in Platform settings; organization settings can restrict service tiers for all projects
Best for: Workflows that run all day against the same instruction block. Count your reuses before enabling it — below roughly three, the write premium wins.
★ Custom voices from a text description, and richer agent sandboxes
Prompt-to-voice opened to all API customers: describe a voice in text, create it through the API or the voice Playground, then use it in GPT-Live sessions. Creating the voice is free; speech generation bills at the usual GPT-Live session rates. Nine new voices also landed — Nuri, Juni and Flitz in German, Brise and Sillage in French, Nira in Hindi, Harema and Shida in Japanese, and Noeul in Korean.
Two quieter additions the same day matter more for agent work. OpenAI-hosted sandboxes in the Agents API now include the Artifact SDK and skills for creating and editing spreadsheets, presentations and Word documents — so an agent can produce a finished file rather than text describing one. And Ultrafast reached the Agents API: set agent.service_tier to ultrafast with a supported model.
Prompt-to-voice: all API customers; create via `v1/audio/voices` or the voice Playground; use in `v1/live/sessions` with `gpt-live-1`; creation free, speech generation at regular GPT-Live session rates | New voices: Nuri, Juni, Flitz (German); Brise, Sillage (French); Nira (Hindi); Harema, Shida (Japanese); Noeul (Korean) | Agents API sandboxes: now include the Artifact SDK and skills for creating and editing spreadsheets, presentations and Word documents | Ultrafast in the Agents API: `agent.service_tier` set to `ultrafast` with a supported model, endpoint `v1/agents`
Best for: Voice products needing a brand voice without a recording session, and agents expected to hand back a spreadsheet rather than a description of one.
Plans and pricing
One significant cut and one new cost lever. Terra halves the rate for computer-use work; the twelve-hour cache adds a 2.5× write premium in exchange for much longer retention.
GPT-6.1 Terra: $1 input / $0.05 cached input / $5 output per 1M; cache writes $1.25 (30-minute) or $2.50 (12-hour); Batch and Flex at half rate; Fast mode $2/$10 short context; over 272K input tokens, 2× input and cache and 1.5× output for the full request | GPT-6.1 Sol: $2 input / $0.10 cached input / $10 output per 1M | 12-hour prompt caching in the Responses API: cache writes at 2.5× the input price | ChatGPT Work and Codex enterprise token billing for Terra: same $1 / $0.05 / $5; credit billing 25 / 1.25 / 125 credits per 1M
Gemini / Google
⚠ Response storage is now on by default for new paid projects
Generate Content API storage now defaults to on (store=true) for new paid-tier projects, matching the Interactions API. Existing projects keep whatever setting they have.
A storage default is a data-protection question wearing a configuration flag. Any new paid project created from October 26 stores its generated content unless somebody changes it — which is fine for a prototype and a problem if the prompts carry client material you have promised not to retain. You can opt out per request with store=false, or turn Generate Content API logging off in the Logs and Datasets Settings panel. Existing projects are untouched, so this bites on the next project rather than the current one.
Default: `store=true` for new paid-tier projects on the Generate Content API | Parity: matches the Interactions API default | Existing projects: keep their current setting | Per-request opt-out: `store=false` | Project-level: turn off Generate Content API logging in the Logs and Datasets Settings panel
Best for: Check this the next time someone spins up a project for a client engagement. The default changed; your data processing agreement did not.
★ A Gemini API CLI for scripts and pipelines
Google shipped a stateless, zero-dependency command-line client aimed at coding agents, scripts and CI pipelines. It uses the Interactions API to generate text and media, analyse files, and run managed agents.
Zero-dependency and stateless are the operative words. A CLI with no runtime to install is something you can drop into a CI image without negotiating with whoever owns the build container, and statelessness means each invocation is self-contained — no session to leak between pipeline steps. For teams whose automation lives in shell scripts rather than application code, this removes the SDK from the path entirely.
Tool: `gemini-api`, beta | Design: stateless, zero-dependency command-line client | Targets: coding agents, scripts, CI pipelines | Backed by: the Interactions API | Capabilities: generate text and media, analyse files, run managed agents | No model IDs or pricing published with the release
Best for: CI pipelines and shell automation that currently wrap an SDK for a one-line call.
⚠ Five models shut down on October 22, with five more due November 17
Four Veo preview models and gemini-omni-flash-preview reached their shutdown date on October 22, all replaced by gemini-omni-1.1-flash.
A second batch follows on November 17: `gemini-3.1-flash-live-preview` and `gemini-2.5-flash-native-audio-preview-12-2025` move to `gemini-3.8-live`, while `gemini-3.1-flash-tts-preview`, `gemini-2.5-flash-preview-tts` and `gemini-2.5-pro-preview-tts` move to `gemini-3.8-flash-tts` or `gemini-3.8-flash-lite-tts`. That is three weeks away and it is the live-audio and speech stack — if you shipped anything voice-related on a preview model this year, this is the list to check against.
Shut down October 22, 2026: `veo-3.1-lite-generate-preview`, `veo-3.1-generate-preview`, `veo-3.1-fast-generate-preview`, `gemini-omni-flash-preview` — all replaced by `gemini-omni-1.1-flash` | Shutting down November 17, 2026: `gemini-3.1-flash-live-preview` and `gemini-2.5-flash-native-audio-preview-12-2025` → `gemini-3.8-live`; `gemini-3.1-flash-tts-preview`, `gemini-2.5-flash-preview-tts`, `gemini-2.5-pro-preview-tts` → `gemini-3.8-flash-tts` or `gemini-3.8-flash-lite-tts` | Also shut down October 5: `antigravity-preview-05-2026` → `antigravity-preview-09-2026`
Best for: Anyone with voice or video on a preview model. November 17 is a hard date and the replacements are already named.
⚠ Deep Research: scheduled for October 23, not confirmed as switched off
We flagged this last issue. The deprecations page, last updated October 22, still listed deep-research-pro-preview-12-2025 as not shut down, with October 23 given as the earliest possible date. No changelog entry records the shutdown.
Google's deprecation page frames published dates as the earliest a model may go, not a guarantee it went. We cannot confirm from Google's own published pages that this agent stopped working on October 23. Plan as though it is gone — the migration is one parameter — but verify against your own calls rather than against this date. Replacement: `deep-research-preview-04-2026`, by changing the `agent` parameter in `interactions.create`.
Plans and pricing
No pricing changes were published in this window. Argon's introductory rate continues to run out on November 30.
No published Gemini API price changes between October 16 and 28, 2026 | `gemini-4-argon`: $2/$10 per MTok through November 30, 2026, then $4/$20; cached input $0.10 then $0.20, implicit caching only | `gemini-3.8-flash`: $0.75/$3.75 through December 31, 2026, then $1.50/$7.50 | Context caching storage: $0.50 per 1M tokens per hour, $1.00 from January 1, 2027
Microsoft Copilot
⚠ The October 19 retirement happened, and a policy API arrived with it
The six models announced on September 18 were retired on schedule: Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini and Grok 4.5, each with a named replacement. On the same day GitHub shipped an API for managing Copilot model policies at enterprise, organisation and team level.
The timing reads as a response to the problem GitHub created for itself. Three deprecation lists in five weeks — six models on October 19, four on October 2, three more due November 16 — is a lot of clicking through settings for an organisation with several teams. A policy API means the model allowlist can live in the same place as the rest of your infrastructure configuration, and change by pull request rather than by someone remembering.
Retired October 19, 2026: Gemini 3.7 Flash → Gemini 3.8 Flash; GPT-5.5 → GPT-5.6 Sol; GPT-5.4 → GPT-5.6 Sol; GPT-5.4 mini → GPT-5.6 Luna; GPT-5 mini → GPT-5.6 Luna; Grok 4.5 → Grok 4.6 | Model policy API: manage Copilot model policies at enterprise, organization and team level | Still scheduled: Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra retire November 16, 2026
Best for: Put the November 16 list into the policy API rather than into a calendar reminder.
★ Assisted permissions and server-to-server authentication
Copilot-assisted permissions reached general availability on October 26, and server-to-server authentication followed on October 27.
Server-to-server authentication is the unglamorous one that unblocks real work. Until now, automating Copilot meant a token belonging to a person — which breaks when that person leaves, and makes every automated action look like it came from them in the audit log. A service identity means a pipeline can call Copilot as itself, which is both more honest in the logs and less fragile in practice.
Copilot assisted permissions: generally available, October 26, 2026 | Server-to-server authentication: October 27, 2026 | Also this period: GPT-6.1 Terra added October 26, the same day OpenAI released it; global default for Copilot features reached general availability (October 22); enterprise managed settings enforcement updates (October 13, previous issue)
Best for: Any Copilot automation currently running on a named person's token.
★ Usage metrics now cover third-party tools, custom skills and MCP servers
Two expansions to measurement in two days. The usage metrics API now reports on custom skills, agents and MCP servers, and separately includes usage from third-party tools such as OpenCode. The Copilot impact dashboard gained industry comparisons.
Counting usage from tools GitHub does not own is the notable half. An organisation that pays for Copilot seats but whose developers also run other agents has, until now, measured only the part GitHub could see — which makes any productivity claim partial by construction. Whether the numbers are flattering is another matter, but at least the denominator is now closer to real.
October 22: custom skills, agents and MCP servers reported in the usage metrics API; industry comparisons in the Copilot impact dashboard; Copilot in Visual Studio September update | October 23: usage metrics API includes usage from third-party tools such as OpenCode; Copilot in Microsoft Teams can turn notifications into code
Best for: Anyone asked to justify Copilot spend with numbers. Check what the new denominator does to last quarter's figure before quoting it.
Plans and pricing
No seat price changes. The month's pattern is governance tooling: a model policy API, a global default reaching GA, server identities, and broader measurement.
No Copilot seat price changes announced between October 16 and 28, 2026 | Model added: GPT-6.1 Terra, October 26 | Retired October 19: six models, as announced September 18 | Scheduled November 16: Claude Opus 4.8, Claude Opus 5, GPT-5.6 Terra | New: model policy API (October 19), global default for Copilot features GA (October 22), assisted permissions GA (October 26), server-to-server authentication (October 27)
Best for: This is a good fortnight to move model policy out of the settings UI and into code, before the November 16 list arrives.