For businesses that use AI at work, these were the October 2026 changes worth knowing: small models got much cheaper, Anthropic updated its usage policy effective 12 November 2026, the Agent Skills format moved to an independent foundation, and Google and OpenAI added a new model and cheaper repeated requests. This roundup covers announcements up to 24 October 2026; anything later goes into the November edition. I only use official pages from the vendors and the European Commission, and each source is linked next to the change it supports. Prices are in USD as the vendors state them and apply on the day I checked them, not the day you read this.
I use AI every day to build websites and games, and I teach AI training for companies (over 20 AI courses, more than 10 of them for the Czech Labour Office in cooperation with Softinum). So for each change I look beyond the announcement at what an ordinary company should actually do about it.
Three things from October matter for most companies. Claude Haiku 5.5 (7 Oct) costs USD 0.10 per million input tokens on the API for shorter requests, a tenth of Haiku 4.5, and suits sorting, summaries and customer support. Claude's new usage policy applies from 12 November 2026: if you use Claude in healthcare, law, finance or to control physical equipment, review it before then. And Agent Skills (22 Oct) is now an open standard under the Linux Foundation, so the company procedures you write down for AI are not tied to one vendor. Also: Gemini 4 Argon has introductory pricing until 30 November 2026, and OpenAI now keeps prompt caches for up to 12 hours.
Which October changes matter for businesses?
| Change | Date | Who it affects | What to do now |
|---|---|---|---|
Claude Haiku 5.5 |
7 Oct |
high volumes of simple tasks via the API |
recalculate the cost of sorting and summaries |
Anthropic 2026 usage policy |
8 Oct, effective 12 Nov |
sensitive sectors, physical equipment |
review the policy before 12 Nov |
Agent Skills under AAIF |
22 Oct |
anyone writing procedures for AI agents |
write procedures as portable skills |
Gemini 4 Argon |
14 Oct |
software, finance, law, security |
test it before introductory pricing ends |
OpenAI: 12-hour cache, Decisions API |
20 and 22 Oct |
repeated requests via the API |
put fixed instructions at the start of the prompt |
AI Act: transparency |
applies since 2 Aug 2026 |
anyone publishing AI content |
have a labelling rule |
What does Claude Haiku 5.5 bring, and whose costs does it cut?
On 7 October 2026 Anthropic released Claude Haiku 5.5, the smallest model in the 5.5 family. Anthropic aims it at high-volume, low-cost work: summaries, classification, database queries, live customer support, and as a helper for the larger Opus 5.5 and Sonnet 5.5. It is the first Haiku with an adjustable effort setting (how much the model thinks). It is available on all platforms including AWS, Google Cloud and Microsoft Azure, with the API ID claude-haiku-5-5.
Prices from the announcement (USD per million tokens, requests up to 100k tokens / over 100k):
| Haiku 5.5 | Haiku 4.5 | Sonnet 5.5 | |
|---|---|---|---|
Input |
0.10 / 0.50 |
1.00 |
2.00 |
Output |
0.50 / 2.50 |
5.00 |
10.00 |
Cache reads |
0.01 / 0.05 |
0.10 |
0.10 |
In the same announcement Anthropic halved Sonnet 5.5 cache-read pricing to USD 0.10 and gave Max and Team subscribers a monthly API credit (USD 100 for Max 5x, USD 200 for Max 20x, up to USD 500 pooled for Team).
What it means: tasks you ran on a larger model only because the small one wasn't good enough may now cost a fraction. The benchmark results in the announcement are Anthropic's own, not independent. I have not tested Haiku 5.5 on my own work yet.
What to do now: take one repeated task (sorting emails, summarising enquiries), run the same 20 to 50 inputs through Haiku 5.5 and through the model you use today, and compare both the results and the cost. Only then switch.
What changes in Claude's usage policy from 12 November 2026?
On 8 October 2026 Anthropic published a usage policy update that takes effect on 12 November 2026. The points that matter for businesses:
- High-risk use (healthcare, legal, financial, essential services): the policy clarifies which recommendations trigger the requirements. The requirements themselves stay the same: a qualified human in the loop and disclosure to the people affected.
- Physical equipment (new): when Claude controls hardware that can act on its own and injure someone, a qualified operator must be able to observe and stop it, and the equipment must fall back to a safe state if Claude disconnects.
- Tracking people without consent is prohibited; consent-based tracking such as fraud monitoring or content moderation stays allowed.
- Deception: a new section brings together the bans on fake account networks, fabricated news sites and influence operations, political or commercial.
- Supported regions: the rules now explicitly take a company's ownership structure into account.
What to do now: if you use Claude in a sensitive sector or connect it to equipment, read the full text before 12 November and write down who checks AI output in your company. For everyday writing, analysis and coding, practically nothing changes.
Why does it matter that Agent Skills moved to an independent foundation?
On 22 October 2026 Anthropic donated the Agent Skills standard to the Agentic AI Foundation under the Linux Foundation, which already received MCP from Anthropic. A skill is a folder with a SKILL.md file that teaches an agent a specific task or domain, optionally with scripts and reference material. According to the announcement, more than 70 platforms already support it, including ChatGPT, Cursor, GitHub Copilot and VS Code. Engineers from Anthropic, Google, OpenAI and other companies maintain it, and the roadmap includes versioned releases, a formal proposal process and security guidance for skill authors.
What it means: company know-how written down for AI ("how we write a quote", "how we check an invoice") doesn't have to be rewritten for every tool. My own library of procedures for AI agents has more than 100 skills in this format, and I share it between Claude Code and Codex.
What to do now: pick one procedure your staff keep explaining to AI over and over, and write it as a single skill: what it is for, the steps, an example of good output, and what the AI must not do. Only install skills from other authors after reading them, since they can contain scripts. How agents work with skills and tools is covered in my article on AI agents and harnesses.
What does Gemini 4 Argon offer and what does it cost?
On 14 October 2026 Google launched Gemini 4 Argon, which it calls its new frontier model. It is in Gemini for business on the web, desktop and mobile, in the API as gemini-4-argon, and in Google AI Studio and Antigravity. Google targets it at complex software builds, finance, law and security; the benchmark results are Google's own.
Introductory pricing until 30 November 2026 is USD 2 per million input tokens and USD 10 per million output tokens, then USD 4 and USD 20. A week earlier, on 8 October, Google introduced the Gemini agent as a "universal agent for work" in Gemini Enterprise; the announcement gives no availability or pricing, so for now just keep an eye on it.
What to do now: if you want to try Argon, do it before 30 November at the introductory price, but budget and sign contracts at the standard price. Which model suits what is covered in my comparison of ChatGPT, Claude and Gemini.
What did OpenAI change for developers?
According to the OpenAI developer changelog, October brought two practical additions:
- 12-hour prompt caching (20 Oct): shared instructions, tools and context stay cached longer, so requests hours apart can still get cached-input pricing. Writing to the cache costs 2.5 times the input price.
- Decisions API (beta on 6 Oct, generally available on 22 Oct): turns text and images into typed answers, meaning structured data instead of free text, using gpt-6-luna.
On 6 October OpenAI also simplified API usage tiers from five to three (Build, Launch, Grow). My own test of GPT-6 Astra on game work is in the article GPT-6 Astra in practice.
What to do now: if you keep sending the same long instructions to the API (rules, a price list, product descriptions), put them at the start of the prompt and the variable part at the end, so the repeated start can come from the cache.
What applies under the AI Act that you must not miss?
This is not October news, but the European Commission's AI Act page (updated 19 October 2026) is a useful reminder: the transparency rules have applied since 2 August 2026. AI-generated content should be identifiable, and deepfakes and AI-generated text on matters of public interest should be labelled. After the amendment, high-risk systems under Annex III follow the rules from 2 December 2027 and under Annex I from 2 August 2028. The Commission has published a voluntary code of practice on labelling (10 June 2026).
What to do now: have a simple internal rule on when and how you label AI content and who checks it before publication. What the AI Act means for an ordinary company is covered in The AI Act in practice. I'm not a lawyer; for borderline cases, ask one.
Which of these have I tried myself?
Of the recent releases, I have tested Claude Opus 5.5 (released 22 September) and GPT-6 Astra on my own work; both have their own articles, Claude Opus 5.5 and GPT-6 Astra above. I use skills every day. As of 24 October 2026 I had not tried Haiku 5.5, Gemini 4 Argon or the Decisions API, so for those I only report what the vendor states.
What to do in November: a short list
- By 12 November: review Anthropic's new policy if you use Claude in a sensitive sector.
- By 30 November: test Gemini 4 Argon at the introductory price, if it interests you.
- Any time: compare one repeated task on a small model (Haiku 5.5) against your current one.
- Any time: write one company procedure as a skill and check it in two tools.
- Have an internal rule for labelling AI content.
If you want to choose models and procedures on your own tasks, I'll show you in AI training. Connecting AI to your processes and systems is part of AI automation.
FAQ
Which October model is the cheapest for high volumes?
Of the announcements in this roundup, Claude Haiku 5.5: USD 0.10 per million input tokens and USD 0.50 per million output tokens for requests up to 100k tokens (checked 24 October 2026). You still have to verify the quality on your own data.
Do I need to change anything because of Anthropic's new policy?
If you use Claude for writing, analysis or coding, practically no. Reviewing it before 12 November 2026 makes sense for use in healthcare, law, finance, for controlling equipment, and for companies with owners in unsupported regions.
What is Agent Skills?
An open standard for packages of knowledge and procedures for AI agents: a folder with a SKILL.md file and optional scripts. Since 22 October 2026 it has been managed by the Agentic AI Foundation under the Linux Foundation, and more than 70 platforms use it.
Why is the AI Act in the roundup if it didn't change in October?
Because the transparency rules have applied since 2 August 2026 and many companies haven't accounted for them yet. The Commission's page restates them in its 19 October 2026 update.
When is the next roundup?
The November roundup comes out on the first Tuesday of December and covers announcements from 25 October 2026.
Sources and links
- Anthropic: Introducing Claude Haiku 5.5 (7 Oct 2026)
- Anthropic: 2026 Usage Policy update (8 Oct 2026, effective 12 Nov 2026)
- Anthropic: Donating Agent Skills to the Agentic AI Foundation (22 Oct 2026)
- Google Cloud: Gemini 4 Argon on Google Cloud (14 Oct 2026) and Google: Gemini agent (8 Oct 2026)
- OpenAI: developer changelog (entries of 6, 20 and 22 Oct 2026)
- European Commission: AI Act (updated 19 Oct 2026) and code of practice on labelling AI content (10 Jun 2026)
I opened and checked all sources on 24 October 2026.