AI Training & Literacy
Latest AI Updates, October 2026: What Changed and How Your Team Should Adapt
Key takeaways
- Between July and October 2026 every major assistant changed its default model: ChatGPT moved paid users to GPT-6 Sol on 7 October and free users to GPT-6 Luna a day later, and Anthropic released Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 between 22 September and 7 October.
- Several tools businesses relied on are being retired: OpenAI closed Sora's app in April 2026 and its API on 24 September, is retiring its visual Agent Builder on 30 November, and Google is replacing Gemini Gems with Skills.
- When the default model changes, prompts and workflows that worked last month can behave differently overnight. Re-testing a team's five most-used workflows after each major release takes an hour and prevents silent quality drops.
- Teams do not need to follow every release. A quarterly review covering what changed, what to re-test, what to retire and which usage rules to update keeps an organisation current without turning AI into a full-time job.
- In Nepal, ChatGPT held about 76% of AI chatbot usage in June 2026 according to StatCounter, Gemini about 13% and Claude about 6%, so most teams should plan around ChatGPT changes first.
Every few weeks someone in a training session asks me the same question: "How do we keep up with all of this?" It is a fair question. In the three months to October 2026, every major AI assistant changed its default model, at least four products that businesses relied on were retired or replaced, and Google changed how its search results show AI answers.
This is a plain-language summary of what changed, what it means for a team in Nepal, and the method I use with clients to adapt without turning AI into a full-time job. It is written as of 10 October 2026; if you are reading this much later, the specifics will have moved, but the method will not.
What changed: the short version
| When | What changed | What it means for your team |
|---|---|---|
| 3 Sep | OpenAI released GPT-6 Astra, its most capable model, to paid ChatGPT plans | Harder analysis and computer-use tasks got better for paying users |
| 22 Sep – 8 Oct | GPT-6 Sol and Luna launched; Sol became the paid default on 7 Oct, Luna the Free and Go default from 8 Oct | Everyone's ChatGPT changed overnight. Free users got a large quality jump |
| 22 Sep – 7 Oct | Anthropic released Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5; Opus 5.5 is about 40% cheaper to run than Opus 5 | Better Claude output at the same plan price; lower API costs |
| 1 Sep | Claude Fable 5.1, Anthropic's tier above Opus | A premium option for the hardest work; most teams will not need it |
| 25 Aug | Claude memory on by default for Free, Pro and Max plans | Check what Claude remembers about your work; update your data rules |
| Late Aug | Claude in Chrome opened to paid plans | Claude can now act in a browser: useful, and a new policy question |
| 16 Sep | Claude Design, Slides and Docs available in chat on every plan, including Free | Presentations and documents straight from a conversation |
| 24 Sep | Sora's API closed, after the app shut in April 2026 | Anyone building on Sora needs an alternative now |
| 30 Sep | Google began replacing Gemini Gems with Skills | Existing Gems will migrate; check yours behave the same |
| 25 Sep | Microsoft overhauled Copilot, adding the Autopilot agent | Microsoft 365 organisations get agent features inside Office |
| 30 Nov (coming) | OpenAI retires its visual Agent Builder | Migrate anything built on it before the deadline |
| Since May | Google merged AI Overviews and AI Mode into one AI Search experience; AI Mode passed one billion monthly users | More questions answered on Google without a click to your site |
Two things are notable about that list. First, it is mostly about defaults changing, not about features you have to opt into. Second, a surprising share of it is retirement: products launched with fanfare in 2024 and 2025 that no longer exist.
What it means in Nepal
ChatGPT changes matter most. StatCounter's figures for June 2026 put ChatGPT at about 76% of AI chatbot usage in Nepal, Gemini at 13% and Claude at 6%. When OpenAI changes the default model, it changes the tool most of your staff, students and customers are using.
Free users just got much better tools. With GPT-6 Luna on the free and Go plans, the gap between free and paid narrowed for everyday writing and summarising. For schools and colleges this cuts both ways: better learning support, and AI-written homework that is harder to spot. For businesses it means staff are likely using a capable free tool on personal accounts, with none of your data protections. That is a policy problem, not a technology one.
Search traffic keeps shifting. Google's merged AI Search answers more questions directly. For Nepali businesses that rely on Google enquiries, being the source those answers cite now matters as much as ranking. I covered how that works in how to get cited by ChatGPT and Perplexity, and the study of 100 Nepali business websites shows how few are prepared.
Claude is officially supported in Nepal, for both the apps and the API, which matters for organisations that need a vendor relationship rather than a workaround.
Why "the model changed" is a business risk
When a vendor changes the default model, the prompts and workflows your team built can behave differently the next morning. Usually better. Occasionally worse in ways nobody notices for weeks: a summary that now leaves out a section, a reply template that has become wordier, a classification that drifts.
Nobody announces this to your team. The model picker shows a new name and work carries on. That is why the first step in my adaptation method is re-testing, not reading the news.
How I help teams adapt: the quarterly method
This is the method I use with teams, run as a ninety-minute session each quarter.
1. Filter: what actually affects you
From the dozens of announcements, I pull the handful that touch the tools and plans your team uses. A bank on Microsoft 365 and a trekking agency on free ChatGPT need completely different briefings. Most announcements, for most teams, require no action at all, and saying so is half the value.
2. Re-test: your top five workflows
Every team I train leaves with saved prompt templates for its most common tasks. After a major model change, we run those five templates on the new model with real examples and compare the output with last quarter's. Anything that got worse gets its prompt adjusted on the spot. This takes under an hour and catches silent quality drops before customers do.
3. Retire and migrate
Gems becoming Skills, Custom GPTs giving way to Workspace Agents on business plans, Sora gone, Agent Builder ending: each needs a decision. Migrate, replace, or drop. I keep a running list so clients are not caught by a deadline.
4. Update the rules
New capabilities create new policy questions. Memory switched on by default raises the question of what Claude is allowed to remember about client work. Browser agents raise the question of which sites and accounts an AI may act in. The team's one-page AI rules get a short update each quarter, written by the team, not handed down.
5. Teach the one new thing worth learning
Usually there is one new capability per quarter that genuinely changes how a team works: presentations generated in chat, an agent that can work through a browser task, a cheaper model that makes bulk work affordable. We spend the last half hour practising it on real work.
What does not change
It is worth saying what stays constant through all of this, because it is what the first training day is built on and why it does not go out of date:
- Context beats cleverness. A well-structured prompt with the right context works on every model.
- Verify before acting. New models are wrong less often, not never.
- Data rules matter more than features. What must never go into a public model is the same list whatever the model is called.
- Portable beats proprietary. Instructions kept in a plain markdown file or a shared document move between tools in minutes. Workflows built deep inside one vendor's newest product do not, as Sora and Agent Builder users found out.
A checklist for this month
If you do nothing else after reading this:
- Open the model picker in ChatGPT and Claude and note what your team is now using by default.
- Run your three most-used prompts again and compare with an old output.
- On Claude, check Settings for memory and decide whether it should be on for work accounts.
- If you use Gemini Gems, check they have migrated to Skills and still behave as expected.
- If anything you run depends on Sora or OpenAI's Agent Builder, plan the replacement now.
- Remind staff about temporary and incognito chats for one-off tasks and about what must never be pasted into a personal account.
Keeping your team current
I am an AI trainer and consultant based in Kathmandu, and I build with these tools every week for client automation projects, which is why the updates I teach come from use rather than headlines. If your team has had no AI training yet, start with the full workshop; if it has, a quarterly update session keeps that investment working.
Book an update session or a first workshop, or send me a message with the tools your team uses and I will tell you which of this quarter's changes actually affect you.
Questions
After these updates, which AI assistant should a team in Nepal use?
Use the one your organisation can pay for, administer and keep data rules on. ChatGPT is by far the most used in Nepal and its free tier improved sharply with GPT-6 Luna. Claude is strongest for long documents and careful writing, and Gemini fits teams already on Google Workspace. The differences between the leaders matter less than using one of them well.
How often should we update our AI training?
A full training day once, then a short update session every quarter. The fundamentals of prompting, verification and data rules barely change. What changes is the tools: new defaults, new features, retired products. A ninety-minute quarterly briefing covering those, plus re-testing your main workflows, is enough for most organisations.
Is GPT-6 available on the free ChatGPT plan in Nepal?
Free and Go users began receiving GPT-6 Luna, the lighter model in the family, from 8 October 2026, while paid plans default to GPT-6 Sol and can access the more capable GPT-6 Astra. Rollouts are gradual and sometimes region by region, so check the model picker in your own account rather than assuming.
Do you run AI update briefings for organisations?
Yes. I run short update sessions for teams, in person in Kathmandu or online. Each one covers what changed in the tools your team actually uses, re-tests your key workflows on the new models live, and updates your usage rules. A quarterly rhythm suits most organisations, and new clients can start with a single session.
