This Month in AI — November 2026

The 60-second version

October brought few headline launches and many changed defaults. Several settings now favour more spending and more agent access unless an admin turns them back. Five things changed:

1. Copilot is being repackaged. On 14 October Microsoft announced that current Copilot offers get "(add-on)" names. From 1 December 2026, new SKUs no longer need a base subscription, and the add-on offers reach end of sale in CSP. Existing subscriptions keep their prices. CSP discounts on Copilot Business (15% off, 1–300 seats, annual) and Business Basic + Copilot Business (25% off) run to 31 December.
2. AI Builder credits ended on 1 November. New and renewed Power Platform and Dynamics 365 licences no longer include them (MC1526652). When existing credits run out, usage falls back to Copilot Credits. With neither, flows that read invoices and forms are blocked.
3. Four agent defaults opened up. Any user can upload agents that have actions or use MCP servers (MC1472007, from 25 October). Any Entra user can create agent identities (MC1506733, from 4 November). Premium models are billable in Office apps, off by default (MC1519122). And SharePoint's KnowledgeAgentScope exclusions stopped being honoured on 1 November.
4. The cheap tier collapsed. Claude Haiku 5.5 costs $0.10/$0.50 per million tokens, the same as GPT-6 Luna. GPT-6.1 Terra is $1/$5. OpenAI shuts down gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini on 14 January 2027.
5. Open-model scores reset. Artificial Analysis moved to Index v5 on 27 October. The best downloadable open model now scores 32, against 58 for Claude Opus 5.5. These figures don't compare with last issue's.

Microsoft: new SKUs, and a meter by default

Usage billing on new Copilot Business seats moved again. Direct purchases now get it from mid-November, Business Premium tenants first and others "at a later date". CSP purchases get it from 1 December 2026. The default limit is 4,000 Copilot Credits per user per month, which at $0.01 a credit is up to $40 per user. Set a spending policy before you buy seats. GPT-6 Astra and Claude Fable 5.1 reach the model picker in Chat, Word, Excel and PowerPoint from early to mid-November, billed per use with no published credit rates. They are off by default.

Business Central agent rates were unchanged. The Payables Agent costs 50 credits per invoice plus 5 per line, so a 10-line invoice is about $1.00 and 300 of them a month is about $300. Power Apps Premium rises from $20 to $22 per user per month from 1 January 2027.

SharePoint after KnowledgeAgentScope

Copilot in SharePoint is GA, and sites you excluded with PowerShell are now in scope. The replacements are per-site Site AI settings and Restricted Content Discovery. Neither changes who can open a file, so fix permissions first. Sources disagree on whether Restricted Content Discovery needs SharePoint Advanced Management. SharePoint ".agent" agents can't be created from mid-November and are cut off in early April 2027 (MC1511052).

Business Central v29

v29 shipped at the start of October. Partner blogs say 1–2 October; Microsoft's page gives no date. Payables Agent PO matching is GA. The new Copilot chat experience and Expense Agent additions are public preview. You can schedule the update through February 2027. SOAP endpoints on Microsoft's own pages stop working in v29, so check older integrations first.

Governance

• Targeted Release retires in January 2027, replaced by Frontier, Standard and Deferred (MC1490899).
• Purview DLP for Copilot Cowork was planned for GA from mid-October (MC1477182). We could not confirm the rollout finished.
• The App Builder agent and vibe.powerapps.com retire on 30 November.
• Microsoft cancelled moving Copilot out of the Teams chat list (MC1477999).

OpenAI

GPT-6.1 Terra (26 October) is a mid-tier model for screen and cross-app work at $1 input, $0.05 cached and $5 output per million tokens, with a 1.05M-token context. Prompts over 272K tokens pay 2× input and 1.5× output on the whole request. Deprecations announced in October: tts-1, tts-1-hd and gpt-4o-mini-tts shut down 6 January 2027. gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini shut down 14 January 2027. gpt-5.1, gpt-5.3-codex and gpt-5.4-nano shut down 1 April 2027. OpenAI's data-controls page now also lists /v1/agents, /v1/chatkit/threads, /v1/evals, /v1/batches and /v1/audio/voices as not eligible for Zero Data Retention.

Anthropic

Claude Haiku 5.5 (7 October) costs $0.10/$0.50 per million tokens for prompts up to 100K tokens. Longer prompts bill the whole request at $0.50/$2.50. It drops temperature, top_p, top_k and prefill. Sonnet 5.5 cache reads fell from $0.20 to $0.10 (release notes). API version 2026-10-21 is the first new version since 2023 and rejects sampling parameters, so pin your SDK versions. Team plans now include up to $500 a month in pooled API credits (pricing). The BAA article now also excludes fast mode, Claude Tag, Design/Slides/Docs and Claude Code via cloud providers. The Computer Use API now appears as covered. The article is undated, so confirm it in writing.

Open-weight models

October releases: Meta's Muse Spark Open (816B, 42B active; attribution plus a separate agreement for hosted-service use), Reflection AI's Beam 1.0 (Apache 2.0), DeepSeek V4.1 Flash 1017 (MIT), MiniMax M3.1 Flash ($20M revenue line, attribution), Mistral Large 4 (hosted service needs a commercial licence), and NVIDIA's Nemotron 3.5 Bolt 74B-A9B (OpenMDW-1.1), which needs about 28 GB free, so a 32 GB card. These hosted-resale clauses don't apply to running the models in-house for your own staff.

Automation of the Month: the Monday renewal-risk watch

The client is a 96-person commercial cleaning company in Westchester and Fairfield County with about $14M in revenue and 212 recurring contracts. It had learned about 11 lost contracts a median of 9 days before they ended. Every Sunday night a job pulls Jobber, Zendesk, HubSpot, QuickBooks Online and Microsoft 365 data into Postgres. SQL computes eight risk signals per account. Claude Haiku 5.5 classifies email threads, and one Claude Sonnet 5.5 call ranks about 40 flagged accounts into nine tiered actions by 06:30 Monday. Over three quarters, gross revenue retention rose from 86.4% to 94.1%, surprise losses fell to 2, and median warning time rose to 71 days. It took 41 build hours and costs $15.92 a month to run. The client is anonymised, and these are that engagement's figures, not a benchmark.

Every item is dated and attributed, and features are marked GA only where the vendor says so. Some items we could not verify. The full text of MC1476200 and the mid-November date come from a mirror. Business Central v29's exact GA date isn't on Microsoft's page. Purview DLP for Cowork may not have finished rolling out. ChatGPT product updates came from aggregators because openai.com did not load. Some open-model parameter counts conflict between sources. The full 15-page PDF edition is free below.