Dateline: September 18, 2026 | Next update: September 25, 2026
A week of calendars rather than launches. OpenAI put firm dates on the end of custom GPTs — no new ones after September 25, and the ones you already have stop running on December 11, with plugins as the replacement — and tightened API key governance so administrators can restrict which key types exist and cap how long any key stays valid. Google took Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking to general availability for real-time audio-to-audio applications, shipped a new Antigravity agent that deprecates its predecessor on October 5, and stopped including Gemini Code Assist in new or renewed Gemini Enterprise Standard and Plus subscriptions. Anthropic added on-demand compaction to the Messages API, which moves the decision about when a conversation gets summarised from the context limit to you, launched a Salesforce plugin with 37 pre-built skills, and extended the Compliance API to cover Claude in Chrome sessions. On the Copilot side, auto model selection gained Efficiency, Balance and Intelligence tiers, budget increase requests reached GA so a developer who runs out of credits can ask rather than stop, and the usage metrics API now reports which skills, custom agents and MCP servers people actually use.
Claude / Anthropic
★ The Messages API can compact a conversation on demand
Compaction used to be something that happened to you when a conversation ran out of room. Now you choose the moment. Send the top-level compaction parameter and the API returns a signed compaction block summarising the messages you sent; on later requests you send that block first, in place of those messages.
Three details make this more than a convenience. The request can run in the background, so compaction does not have to sit in the user's critical path. You can keep recent turns word for word after the summary, so the model still sees the last exchanges exactly as they happened rather than through a paraphrase. And on models with preserved thinking, the thinking in those kept turns stays valid — which is the part that usually breaks when a conversation is rewritten.
Beta header: `compact-2026-09-04` | Parameter: top-level `compaction` on the Messages API | Response: a signed `compaction` block summarising the messages you sent | Usage: send the block first on later requests, in place of the summarised messages | Options: run the request in the background; keep recent turns verbatim after the summary | Thinking: on models with preserved thinking, thinking in the kept turns remains valid
Best for: Anyone running long agent sessions who has been writing their own summarisation layer. Compacting deliberately, at a quiet moment, is cheaper and less lossy than compacting because you hit a wall.
★ A Salesforce plugin with 37 pre-built skills
Claude gained a Salesforce plugin in beta, giving sales teams access to accounts, opportunities and pipeline data, with 37 pre-built skills rather than a bare connection you then have to teach.
Status: beta | Availability: all paid plans, for approved organisations | Scope: accounts, opportunities and pipeline data | Included: 37 pre-built skills
Best for: Revenue teams whose CRM questions currently go through whoever is fluent in reports. The pre-built skills are the difference between a connector and something a salesperson will actually use.
★ The Compliance API now covers Claude in Chrome
The Compliance API's local session endpoints now also return transcripts of Claude in Chrome sessions, identified by the product_surface value claude_in_chrome, using the existing Compliance Access Key and the read:compliance_user_data scope.
Endpoints: the existing local session endpoints | New `product_surface` value: `claude_in_chrome` | Auth: existing Compliance Access Key with the `read:compliance_user_data` scope | Status: beta, for Claude Enterprise organisations | Context: the same endpoints already covered Cowork and Claude Code sessions running on users' machines
Best for: Compliance teams closing the last gap in session coverage. Browser sessions were the surface that produced work nobody could later retrieve.
★ Claude Code — plugin evaluation, output styles, and per-command domain rules
The most interesting addition is claude plugin eval, a command that scores a plugin's compliance — useful precisely because plugins are the thing teams now install most freely. Alongside it came an /output-style slash command, diffs for Bash file edits, and OpenTelemetry repository tagging.
Auto mode gained per-command `allowed_domains` for Bash, PowerShell and Monitor when sandboxing is on — domain rules per command rather than per session. Remote sessions gained Fast mode. Skills and plugins now sync from a claude.ai account into terminal sessions, and signing in through a gateway with a claude.ai account now asks for confirmation. Ctrl+Enter became a send-now key that interrupts the current turn and queues your message. `modelPricing` accepts a `multiplier` up to 10, and agents gained an `omitClaudeMd` frontmatter option. On the infrastructure side, the v2 MCP client became the default on Bedrock, Vertex and Foundry, MCP server disconnections now raise a notification instead of failing silently, and sessions stuck retrying "unexpected tool_use_id" errors now self-heal.
2.1.269 (Sep 11): `claude plugin eval`, `/output-style`, Bash file-edit diffs, OpenTelemetry repository tagging, VS Code agent map | 2.1.271 (Sep 14): Fast mode in Remote sessions, per-command `allowed_domains` for Bash/PowerShell/Monitor under sandboxing, `omitClaudeMd` agent frontmatter, `--accept-command` for plugin installation, `modelPricing` multiplier up to 10 | 2.1.273 (Sep 15): request headers for LLM gateways, MCP disconnection notifications, session forking from Remote Control, Datadog presets for US3, AP1, AP2 and US1-FED, v2 MCP client default on Bedrock/Vertex/Foundry | 2.1.274 (Sep 16): memory-usage warning with recovery steps, environment variables for MCP startup wait and OpenTelemetry tracing | 2.1.275 (Sep 17): gateway sign-in confirmation for claude.ai accounts, Ctrl+Enter send-now, skills and plugins synced from claude.ai to terminal sessions, `/plugin install
Best for: Teams with a growing plugin estate — run `claude plugin eval` before the next one goes org-wide. If you updated to 2.1.275 and run through a gateway, 2.1.276 is the version you want.
Plans and pricing
No pricing changes this week and no model retirements announced. Claude Fable 5.1 and Mythos 5.1 remain at $10/$50 per MTok with cache reads at $0.25 per MTok.
Fable 5.1 / Mythos 5.1: $10/$50 per MTok, cache reads $0.25 per MTok | Opus 5: $5/$25 per MTok | Sonnet 5: $2/$10 per MTok | Haiku 4.5: $1/$5 per MTok | No model retirements announced this period | New this week at no extra cost: on-demand compaction in beta, which reduces the tokens you resend rather than adding a charge
Best for: No action on price. Compaction is the item with a cost consequence, and it points downward.
ChatGPT / OpenAI
⚠ Custom GPTs now have a retirement date
OpenAI put a calendar on the end of custom GPTs for affected workspaces. Admin notice went out on September 11, the migration experience and user banner were targeted for September 17, new custom GPTs can no longer be created after September 25 — including drafts, which must be published before that cutoff — and on December 11 custom GPTs stop running.
Plugins are the replacement, bringing reusable instructions and connected apps together in one place. The preparation OpenAI suggests is the right shape regardless of timing: identify the GPTs you actually rely on, who created each one and who needs access to the replacement, then record the instructions, reference material, integrations and the familiar prompts you will use to test that the replacement behaves the same. The September 25 cutoff is the one that sneaks up — after it you cannot create a new GPT even to reproduce an old one.
September 11, 2026: admin notice | September 17, 2026: target for the migration experience and user banner, stated as a target rather than a guarantee | September 25, 2026: creation of new custom GPTs ends; drafts requiring migration must be published before this date | December 11, 2026: custom GPTs stop running | Replacement: plugins
Best for: Any workspace with GPTs people use daily. Two of the four dates have already passed — the next one is a week away, and it is the one that removes your ability to rebuild.
★ API key governance — restrict key types and cap key lifetimes
Administrators can now restrict which types of API keys can be created at all: allow only service-account keys, allow only user-owned project keys, or disable certain types outright. The second half is the one that ages well — a maximum API key lifetime, enforced at organisation or project level, so new keys cannot be configured to remain valid indefinitely.
Location: API Key Governance in Platform settings | Configured by: organisation and project administrators | Controls: which key types may be created — service-account keys only, user-owned project keys only, or specific types disabled | Lifetime: a maximum API key lifetime enforced at organisation or project level; new keys must be configured within the allowed limit | Precedence: project limits cannot exceed the organisation limit
Best for: Platform and security teams with keys nobody can attribute and nobody dares rotate. A maximum lifetime turns key hygiene from a policy you write into a rule the platform enforces.
★ Advertising — Sponsored Agents, and ads built by prompting
OpenAI is testing Sponsored Agents: after clicking an ad in ChatGPT, a user can start a clearly labelled conversation with a business-sponsored agent, explain what matters to them, ask follow-up questions, and follow a link to the business's site when they are ready. That is a different object from an ad — it is a conversation a business pays to have.
On the advertiser side, ads can now be created by writing a few prompts in ChatGPT Work, and Ads Manager gained AI creative tools. OpenAI also named its first CRM and ecommerce partners — HubSpot and Shopify — bringing ChatGPT Ads into tools businesses already run, with the app available internationally from September 23 in markets where ChatGPT Ads operate. For scale: ChatGPT Ads reached a $1 billion annualised revenue run rate in under 200 days and is used by tens of thousands of advertisers.
Sponsored Agents: in testing; clearly labelled conversation with a business-sponsored agent, started from an ad | Ad creation: by prompt in ChatGPT Work | Ads Manager: new AI creative tools | First CRM partner: HubSpot | First ecommerce partner: Shopify | International availability of the app: September 23, in markets where ChatGPT Ads are available | Scale: $1B annualised run rate in under 200 days, tens of thousands of advertisers
Best for: Marketing teams — and anyone thinking about what an assistant's answers are for. A sponsored conversation is a heavier format than a sponsored link, and worth understanding before it is normal.
★ A framework for reporting model misalignment
OpenAI published a framework for tracking, investigating and disclosing model misalignment, along with six reports on unexpected or concerning model behaviour observed over the previous six months.
The honest part is the diagnosis of the old approach: disclosures had been ad hoc and less frequent than ideal, often waiting until several instances could be collated into one report. The framework is meant to get reports out sooner after observation, even when the behaviour has not been fully explained or mitigated. OpenAI notes there is no industry-wide standard for how AI developers should disclose misalignment, and offers this as a first step towards one — setting out which instances should be disclosed and what a report should contain. It calls the framework a work in progress.
Scope: which misalignment instances developers should disclose, and what reports should contain | Change in practice: publish after observation rather than waiting to batch several instances, including where the cause is not yet explained or mitigated | Published alongside: six reports covering the preceding six months | Status: described by OpenAI as a work in progress, to be refined through experience and public feedback | Reports are collected on OpenAI's alignment site
Best for: Anyone writing an AI policy who has to say what a vendor will tell you and when. A published disclosure standard is something you can hold a supplier to; a practice of occasional blog posts is not.
Plans and pricing
No model list-price changes were published between September 11 and September 18. The commercially significant items this week are a deprecation with dates attached and a governance control that affects how keys, and therefore spend, are issued.
Model list prices: unchanged this period | Custom GPTs: creation ends September 25, 2026; retirement December 11, 2026 | API key governance: key-type restrictions and maximum key lifetimes at organisation and project level | ChatGPT Ads: $1B annualised run rate; HubSpot and Shopify integrations international from September 23 | Still in effect from last week: new sign-ups and upgrades to the $200 Pro tier remain paused
Best for: Put September 25 and December 11 in the calendar. Everything else this week is governance rather than cost.
Gemini / Google
★ Gemini 3.8 Live reaches general availability
Google took two models to general availability for real-time audio-to-audio applications: gemini-3.8-live and gemini-3.8-live-extended-thinking. The second is the interesting one — extended thinking in a live audio model is the combination that has been missing, because reasoning time and conversational latency usually pull against each other.
Model IDs: `gemini-3.8-live`, `gemini-3.8-live-extended-thinking` | Status: generally available from September 15, 2026 | Purpose: audio-to-audio real-time applications | Context: Gemini 3.8 Flash reached GA on September 2, so the 3.8 line now spans text and live audio
Best for: Teams building voice agents on Gemini. Model the two variants separately — the extended-thinking one exists because some turns need to think, and those turns cost differently.
★ Antigravity Agent 09-2026 replaces the May preview
The Antigravity Agent 09-2026 release replaces antigravity-preview-05-2026, with built-in tools updated to PascalCase parameters and line-range file edits.
`antigravity-preview-05-2026` deprecates on October 5, 2026 — under three weeks from this issue. The parameter casing change from the built-in tools update is the kind of migration that fails loudly rather than silently, which is the better outcome, but it still needs doing before the date.
New: Antigravity Agent 09-2026 | Replaces: `antigravity-preview-05-2026` | Deprecation date for the old preview: October 5, 2026 | Built-in tool changes: PascalCase parameters, line-range file edits
Best for: Anyone on the May preview. Three weeks, and a parameter-naming change in the middle of it.
★ Workspace — notebooks for organisations, and Studio grows webhooks
Notebooks in Gemini opened to schools and organisations, giving students, educators and professionals a place to organise conversations by topic — the Gemini app combined with Gemini Notebook capabilities, rolling out from September 14 to all Workspace customers and Workspace Individual subscribers outside the EEA.
Workspace Studio gained four things that together turn it into a real automation tool: custom starters, custom steps, third-party integrations with Asana, Confluence, HubSpot, Jira, Mailchimp, QuickBooks, Salesforce and Slack, and webhooks. Admin controls arrived on September 17, with end-user features between September 21 and 30. Expert Intelligence in Gemini Notebook lets you add more than 100,000 compatible ebooks from Google Play Books to a notebook and ask questions grounded in the text, or generate infographics, audio overviews and quizzes from it. And Gemini in Workspace now connects to Asana, Atlassian Rovo, HubSpot, Intuit Mailchimp, Intuit QuickBooks, Monday and Salesforce over the Model Context Protocol.
Notebooks in Gemini (Sep 17 post): all Workspace customers and Workspace Individual subscribers outside the EEA; rollout from September 14, gradual over up to 15 days | Workspace Studio (Sep 17 post): custom starters, custom steps, third-party integrations and webhooks; Business Starter/Standard/Plus, Enterprise Standard/Plus, Education Fundamentals/Standard/Plus; admin console September 17, end-user features September 21–30 | Expert Intelligence: 100,000+ Google Play Books ebooks; all Workspace customers, Workspace Individual and personal accounts; available now | MCP connections (Sep 15 post): Asana, Atlassian Rovo, HubSpot, Intuit Mailchimp, Intuit QuickBooks, Monday, Salesforce; Business and Enterprise tiers, Google AI Pro/Ultra, Enterprise Essentials Plus, Education add-ons
Best for: Workspace admins. Webhooks in Studio are the change that lets an automation react to something outside Google, which is where most real workflows actually start.
⚠ Gemini Code Assist drops out of new Standard and Plus subscriptions
Gemini Code Assist is no longer included in new or renewed Gemini Enterprise Standard or Plus subscriptions. Existing subscriptions are not described as changing mid-term, but a renewal is a new subscription for this purpose — which makes this a change that arrives on your renewal date rather than on an announcement date.
Also on September 17, voice input for chat reached general availability in Gemini Enterprise. It is off until an administrator turns it on, through the "Enable speech-to-text" toggle.
Gemini Code Assist: no longer included in new or renewed Gemini Enterprise Standard or Plus subscriptions, from September 17, 2026 | Voice input for chat: generally available, requires an administrator to enable the "Enable speech-to-text" toggle
Best for: Whoever owns the Gemini Enterprise renewal. Check whether Code Assist is in your current entitlement and what it costs separately, before the renewal quote rather than after it.
Plans and pricing
No Gemini API list-price changes were published this week. The commercial change sits in Gemini Enterprise packaging rather than in per-token pricing.
Gemini API list prices: unchanged this period | Packaging change: Gemini Code Assist removed from new and renewed Gemini Enterprise Standard and Plus subscriptions | Still running: Gemini 3.7 Flash introductory pricing through December 31, 2026 | Dated deprecation: `antigravity-preview-05-2026` on October 5, 2026
Best for: Two dates and one packaging change. The Code Assist change is the one that shows up as a number on a renewal.
Microsoft Copilot
★ Auto model selection gains cost-and-quality tiers
Copilot's auto mode evaluates each prompt individually to pick a model. It now does so against a preference you set, across three tiers: Efficiency prioritises low cost and suits fast, straightforward tasks; Balance weighs cost, quality and latency together for everyday work; Intelligence prioritises quality for complex tasks.
Tiers: Efficiency, Balance, Intelligence | Model set: identical across tiers — the tier changes how auto optimises per prompt, not which models are available | Billing: charges are based on the model auto selects, not the tier chosen | Discount: paid subscribers continue to receive a 10% discount on usage billed through auto | Rollout: Visual Studio Code, Copilot CLI and the GitHub Copilot app
Best for: Anyone watching usage-based billing. Note the billing subtlety: the tier is a preference, not a price band — you are charged for whatever model auto picks, and the 10% auto discount applies either way.
★ Budget increase requests reach GA
When a member exhausts their AI credits, they can now request more budget at the moment they hit the limit rather than simply being blocked. Organisation owners, enterprise owners or billing managers review the request under "Requests from members", set an amount, and select Approve and increase — which restores access immediately.
Plans: GitHub Copilot Business and Copilot Enterprise under usage-based billing | Not available for: enterprises with managed users | Requesters: members who have exhausted their AI credits | Approvers: organisation owners, enterprise owners, billing managers | Path: "Requests from members", set a new amount, Approve and increase | Effect: access restored immediately on approval
Best for: Teams where hitting a credit ceiling currently means a Slack message and a wait. This turns that into a tracked request with an audit trail.
★ Usage metrics now show which skills and agents people actually use
The usage metrics API gained agentic activity tracking for skills, custom agents, MCP servers, slash commands and plugins — with top-usage fields listing the five most active items in each category, and distinct-count fields showing how many different ones are in use.
The impact dashboard now also shows feature engagement: how many active users engaged with each included feature on at least two days during the 28-day period, broken down across code completion, agent edit, passive code review, active code review, cloud agent, CLI and app, with the rolling population classified by AI adoption phase. Separately, on September 15, Copilot began suggesting allowed values when an enterprise or organisation admin creates a custom property definition — for example compliance values for a "FedRAMP" property, or yes/no options for "internet-facing". That is in public preview for Business and Enterprise, controlled by the "Repository custom property suggestions" policy.
Top usage fields: `totals_by_skill`, `totals_by_custom_agent`, `totals_by_mcp`, `totals_by_slash_cmd`, `totals_by_plugin`, each listing up to five items with the highest activity | Distinct count fields: `distinct_skill_use_count`, `distinct_custom_agent_use_count`, `distinct_mcp_use_count`, `distinct_slash_cmd_use_count`, `distinct_plugin_use_count` | Reports: enterprise and organisation per-user and aggregate 1-day reports, per-user 28-day reports, and `day_totals` entries in aggregate 28-day reports | Impact dashboard: per-feature engagement over 28 days, requiring engagement on at least two days; available through the enterprise and organisation report APIs | Custom property suggestions: public preview, Copilot Business and Enterprise, governed by the "Repository custom property suggestions" policy
Best for: Anyone who has rolled out skills and MCP servers and cannot say which ones earned their place. The distinct-count fields answer a question the old metrics could not: how much of what you installed is actually being used.
Plans and pricing
No Copilot seat-price changes were published this week. Two dated changes announced earlier are now live or imminent: from October 1, new Business and Enterprise seat assignments require payment per seat before access, and on September 28 the default code review effort level moves from Lite to Balanced unless an admin selects Lite first.
Seat prices: unchanged this period | Auto model selection: charged on the model selected, with a 10% discount on usage billed through auto for paid subscribers | September 28, 2026: code review default effort moves Lite → Balanced unless an admin selects Lite; and Copilot Cloud Agent, github.com Chat and GitHub Mobile Chat consolidate under a single policy, with chat retention extending from 28 days to the account lifetime | October 1, 2026: per-seat payment required before access on new Business and Enterprise assignments | Microsoft's own Microsoft 365 Copilot release notes have published nothing new since August 25, and the Copilot Studio "what's new" page still lists July 2026 as its most recent month
Best for: September 28 is ten days away and changes a default underneath you. That is the one to act on, not the pricing.