What changed in AI this quarter that matters at your desk
Three things changed in the last six weeks that reach a non-technical leader. The default model behind everyday ChatGPT chats was replaced. New Claude models began watermarking their text output so it can be detected later. And a new frontier model arrived. Two of those three change how you work next week.
As of 13 September 2026.
The default model changed and effort became a setting
On 6 August 2026 OpenAI replaced GPT-5.5 Instant, the model that had been running everyday chats for Free and Go users of ChatGPT. The two August models are GPT-5.6 Sol and GPT-5.6 Luna. Plus and Pro users got an updated GPT-5.6 Sol with a slider that sets how much effort goes into a response. All of this is recorded in OpenAI's own August update page: https://deploymentsafety.openai.com/gpt-5-6-august-update
What this means at your desk is that effort is now something you choose.
A reply to an internal mail, effort low. A market note going to your regional head, effort high, and expect to wait for it. Your team on the free tier and you on a paid tier are no longer running the same thing. When a junior shows you an output you cannot reproduce, that is why.
Claude's text output now carries a watermark
On 14 August 2026 Anthropic said new Claude models carry a watermark in their text output by default, and that it is working to add watermarking to models launched before 2 August 2026. Anthropic describes it as a version of the SynthID-Text approach published by Google DeepMind. The pattern sits in how words get selected, so nothing visible is added to the text. Image files get C2PA content credentials in the metadata instead. Detection runs through an API in limited preview: https://www.anthropic.com/news/claude-text-watermark Access is limited to bodies with obligations under EU law. That list includes regulators, law enforcement, media, fact-checkers, researchers, educational organizations, and enterprises that have to verify compliance.
The stated reason for the rollout is the EU AI Act's requirement to mark AI-generated content.
Provenance is now checkable by someone other than you. If you run a country subsidiary, that is a reporting question before it is a technology question. A board paper, a regulatory submission, a published claim: assume a passage drafted in Claude can be checked later by one of those bodies. The first draft can come from a tool. The thinking, the numbers and the sign-off have to be yours. That was already true. It is now auditable.
What to ignore this quarter
Ignore GPT-6 Astra for now. OpenAI published its system card on 3 September 2026. It is the first OpenAI model to reach the Critical level of cybersecurity capability under the company's Preparedness Framework. That rating brought stricter internal isolation and monitoring with it: https://deploymentsafety.openai.com/gpt-6-astra
None of that reaches your desk this month. Before planning around it, open your account and check what you have. If it is not in your model picker, it is news, not a tool.
Ignore version numbers generally. Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 in early September 2026: https://www.anthropic.com/claude-fable-and-mythos-5-1 That was less than two weeks before I wrote this. A quarterly plan built on a version number is out of date before it is approved. Plan around the task instead.
How do I learn AI when the model names keep changing?
Learn the work, not the release notes. The version numbers change every few weeks and none of what I wrote above required knowing how any of it was built. I can't code and never could. What I did in 2023 was take one real problem off my desk each day and ask whether a tool could do part of it better or faster.
Pick the piece of your own week that is assembly rather than judgment. The monthly review pack. The competitor note nobody has time to write. Run it once, end to end, on the real file. Then mark up the output against what you know: every figure that traces to nothing, every cause asserted with no number behind it, every sentence that sounds right and says nothing. That mark-up is what you are actually learning, and no release note will do it for you.
How do I start with AI?
Start with one recurring task you already own, this week, on the real file. Not with a model name, a course, or a release note.
A junior showed me an output I cannot reproduce. What now?
Check which model each of you is on before you question the output. Free, Go, Plus and Pro are not running the same thing this month, and the effort slider changes the answer again. Ask for the prompt and the tier, then run it yourself.
Can I put company documents into these tools?
Ask whoever owns the AI policy to give you the line in writing, because until somebody states it you are guessing on the company's behalf. Until then, practice on material you would already send outside the building, or strip the names and codes and keep the structure. The method works on anonymized data.
Should I wait for the next model before starting?
No. The gap between you and a junior who uses this daily is reps, not account access. A new model will not teach you to catch an answer that reads well and is false about your own market. That comes from marking up your own work, week after week.