This Month in AI — October 2026

The 60-second version

September turned over a full model generation at the same or lower prices, and it made Copilot's metered bill the default. Five things changed:

1. Copilot's second invoice becomes opt-out. From 19 October 2026 (direct) and 2 November (CSP), new Copilot Business purchases come with an Azure subscription and spending policy already provisioned (MC1476200). (Update: on 1 October Microsoft moved the CSP date to 1 December 2026, with a default limit of 4,000 Copilot Credits per user per month.) "Auto-apply new services" is on by default, so new metered products join the policy without anyone approving them.
2. Copilot in SharePoint and OneDrive reached GA on 30 September, and the bill is split. Q&A, catch-up and file creation are included in the licence. "Advanced work" bills in Copilot Credits at rates Microsoft has not published. On 1 November, Set-SPOTenant -KnowledgeAgentScope retires.
3. New models at the same or lower prices. GPT-6.1 Sol lists at $2/$10 per million tokens, GPT-6 Luna at $0.10/$0.50, Claude Sonnet 5.5 at $2/$10, and Claude Opus 5.5 at $4/$20, down from $5/$25.
4. Custom GPTs are being retired. No new GPTs can be created in Enterprise workspaces from 26 October. Existing ones retire on 11 December 2026, or 11 February 2027 for Enterprise workspaces with a deferral. Custom actions do not migrate.
5. A new open-weight leader. Xiaomi's MIT-licensed MiMo-V2.6-Pro scores 46 on the Artificial Analysis index, 12 points behind Claude Opus 5.5 at 58.

Microsoft: the new Copilot, and the meter

On 25 September Microsoft announced a rebuilt Copilot with three modes. Home and Code are Frontier, and Autopilot is in private preview. Copilot Managed Runtime is in public preview, so nothing in the new app is GA yet. Copilot Cowork has been GA since 16 June, and Teams Phone Agent went GA on 30 September. Both are metered in Copilot Credits at $0.01 each. Microsoft's estimates are 100 to 300 credits for a light Cowork task, 300 to 700 for a medium one and over 700 for a heavy one. A Teams Phone Agent call runs 5–40+ credits. Microsoft has published estimates for Managed Runtime hosting, but no firm credit rates for Code or Autopilot. If you piloted Teams Phone Agent, your number falls back to the old auto attendant until an admin enables usage billing.

The sources conflict on promotions. Microsoft says "Copilot promotions end" on 30 September, the day before CSP growth margins went GA. Three distributors say only the 300+ seat enterprise promos ended, and that Copilot Business promos and Copilot in 30 run to 31 December. Get your price in writing.

Grok models reached Word, Excel and PowerPoint for Frontier tenants from 18 September. They are off by default and not offered in the EU, UK or GCC.

SharePoint: GA, and on everywhere from 1 November

If you used KnowledgeAgentScope to keep Copilot off HR, Finance or client sites, move those sites to Restricted Content Discovery or the per-site "Site AI" settings before 1 November. Then test as an ordinary user. The GA release runs on a Microsoft-managed OpenAI reasoning model. The preview ran on Anthropic.

Business Central v29

2026 release wave 2 went to public preview in early September, and GA is expected in the first week of October. Every new AI feature is labelled preview: unified Copilot chat, Payables Agent PO matching, Expense Agent additions, a new Tax Matching Agent for Shopify, and MCP Server data queries. (Update: Microsoft's GA release notes now list Payables Agent PO matching and MCP Server data queries as generally available, and no longer list the Tax Matching Agent.) The change that is not preview: SOAP endpoints on Microsoft's own pages stop working in v29, and the feature key to re-enable them has been removed. Microsoft's advice for an integration that still needs one is to copy the page into your own extension. Consumption rates are unchanged.

Governance calendar

• 1 Oct: ChatGPT for Word enabled by default for ChatGPT Business and Enterprise.
• Early Oct: Copilot moves to copilot.cloud.microsoft. Allow *.cloud.microsoft.
• 13 Oct: Office LTSC 2021 end of support.
• Mid–late Oct: Purview DLP for Copilot Cowork goes GA. It is in preview now.
• 1 Nov: KnowledgeAgentScope retires.

OpenAI

OpenAI released GPT-6 Astra (3 September, limited rollout, $10/$50), GPT-6 Sol and Luna (22 September), and GPT-6.1 Sol (29 September, $2/$10 with $0.10 cached input). A correction to last month: GPT-5.6 Sol now lists at $4/$20. The $2/$10 we printed was its Batch and Flex rate. Always-on "dots" agents launched 29 September and require Business Premium seats ($125/$100). DevDay called them GA and the release notes say beta, so treat them as beta. ChatGPT Business seat prices are unchanged. New workspaces now cap at 200 paid seats. Sora 2 and the Videos API shut down on 24 September.

Anthropic

Claude Opus 5.5 (22 September) costs $4/$20, and Anthropic says it performs at Fable 5.1 level, at 40% of Fable's list price. Sonnet 5.5 (28 September) stays at $2/$10. Opus 5.5 brings breaking API changes: thinking cannot be switched off, forced tool use is not supported, and computer use needs a new tool version. Sonnet 4.5 retires 30 November. A correction: Mythos 5.1 is "Project Glasswing only", not generally available. Fable 5/5.1 and Mythos 5/5.1 require 30-day retention and are not ZDR-eligible. Cowork is not BAA-eligible in any configuration. (Update: since this edition was published, Anthropic's BAA covers Cowork in local mode once the Primary Owner applies the HIPAA configuration. Cowork in the cloud is still excluded.)

Open-weight models

On the Artificial Analysis Intelligence Index v4.3.2 (read 1 October), the best closed model is Claude Opus 5.5 at 58 and the best open model is MiMo-V2.6-Pro at 46. The index has been re-versioned, so these figures do not compare with last month's. Also released: DeepSeek V4.1-Flash (MIT), Nex-N2.5 (Apache 2.0) and MiniCPM5-2B (Apache 2.0, runs on a laptop). On a single 24 GB GPU, Ternary-Bonsai-2-27B fits in 6–7 GB. Licence traps: FLUX 3 Action allows commercial use only under $5M in annual revenue. Qwen's new Community License requires a separate licence for hosted resale. Anthropic's 10 September report alleges seven labs, Xiaomi among them, trained on Claude outputs. Those are unaudited claims.

Automation of the Month: the weekly margin-leak review

The client is a 58-person specialty food distributor in northern New Jersey with about $22M in annual revenue. Its gross margin had slid from 35% to 31.8%. Every Monday at 05:00 a pipeline pulls 13 weeks of data from Business Central, HubSpot, Shopify, Klaviyo and Stripe into Postgres. SQL then does every join and calculation. One Claude Sonnet 5.5 call ranks about 60 flagged rows into eleven tiered actions, each with an owner and a dollar figure. Over two quarters, gross margin rose to 35.1%, SKUs below target fell from 142 to 37, and stale price agreements fell from 64 to 9. It took 46 build hours and costs $29.62 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: credit rates for Copilot Code and Autopilot, which promotions ended on 30 September, Sherweb's report on on-prem BC licensing, and the GA-versus-beta status of OpenAI dots. The full 15-page PDF edition is free below.

Corrected 8 October 2026: Copilot Cowork has been generally available since 16 June, not 23 September; Microsoft's Cowork estimates are 100 to 300, 300 to 700 and over 700 credits; CSP growth margins went GA on 1 October; the SOAP change covers Microsoft's own pages; and Microsoft does publish Managed Runtime estimates. The price lines now say "same or lower prices", the custom GPT cutoff is marked as applying to Enterprise workspaces, and the note on last month's GPT-5.6 Sol price names the rate we printed in error. Changes since publication, to the CSP date, Business Central's GA release notes and Anthropic's BAA, are marked as updates in the text. The PDF edition still has the original text.