The Signal: AI You Can Use This September

A September 7 briefing on Astra, Fable 5.1, Gemini, Copilot, memory, learning tools and AI search, with practical experiments for work, study and everyday life.

September 7, 2026. Choose one thing to try.

The AI news cycle can make ordinary users feel permanently behind. A new model arrives, another app learns to act, and a feature you were about to understand changes its name. The useful question is smaller: which change makes something you care about possible now? This edition follows the current announcements into a study session, a work task, a business page and a family activity. The experiments are original suggestions, not results from tests we have run.

The bigger models need a real task to prove themselves

OpenAI’s current model reference describes GPT-6 Astra as its most capable model for difficult work from beginning to end. Anthropic announced Fable 5.1 and Mythos 5.1 on September 1, emphasizing coding, knowledge work and research. Mythos access remains restricted; its announcement is not an invitation for every reader to select it in a chat. These are provider descriptions and availability statements, not our independent ranking.

The practical opportunity is to revisit one task whose earlier attempt got stuck. Perhaps the assistant made a working page but could not fix an awkward interaction, or summarized sources without resolving which passage supported a claim. Keep the original inputs and define the failure you want the new attempt to address. Changing the task, the files and the model at once makes it hard to know what improved.

Try a small comparison: give the same fictional meeting notes to two assistants you already have access to. Ask both to separate definite actions from conditional offers and leave missing owners blank. Inspect the actual assignments. The useful result is a record of what each did on your task, including mistakes, rather than a verdict about every possible use of the model.

Sources: OpenAI: GPT-6 Astra model reference · Anthropic: introducing Fable 5.1 and Mythos 5.1

A request you can adapt

Use these same fictional meeting notes and the stated extraction rules. Return decisions, explicit actions, conditional offers and unresolved questions. Do not infer owners or deadlines. I will compare your output with a reviewed reference. Do not claim that your result establishes a general model ranking.

Use the complete meeting example

A faster release cycle changes how to choose

Google’s August roundup, published September 1, includes Gemini 3.7 Flash and Gemini 3.5 Transcribe. On September 2 it announced Gemini 3.8 Flash and the specialist 3.8 Flash Cyber variant. That sequence is a useful reminder that a monthly recap and a current model listing answer different questions. Access to a specialist variant also needs to be checked separately from a general model announcement.

For most readers, a new version number is a reason to inspect the relevant capability, not to rebuild a personal workflow immediately. If your task is a clear two-paragraph reply, changing models may matter less than giving the assistant the actual policy. If a tool repeatedly fails to handle your files or finish a multi-step task, a new capability may be worth a bounded retry.

Keep a short working note: task, environment, model selected, date and the check that mattered. You can then revisit the task after a release without reconstructing it from memory. The point is to make the rapid change easier to evaluate, rather than letting each announcement create another subscription decision.

Sources: Google · August 2026 AI roundup, published September 1 · Google: Gemini 3.8 Flash and Flash Cyber, September 2

Read the dated model radar

Memory is becoming something you should inspect

Anthropic’s August 25 release note says memory now works across chat and Cowork in the cloud. It describes editable topics in Settings, a separate setting for sensitive topics, and different defaults for individual and organizational plans. That is a product change with a practical consequence: the context influencing a new task may come from more than the messages currently on screen.

Our suggested experiment is deliberately ordinary. In an account where the feature is available, inspect the memory settings and the topics the product shows. Check whether a harmless preference is accurate and whether an old project assumption still belongs there. Use the product’s controls to correct a stored item if you choose. Do not infer that asking the assistant to forget something has changed every storage or retention setting.

For an important task, still supply a compact current brief. “This is the draft for the autumn workshop; the spring venue and dates no longer apply” is useful even if the assistant remembers the project. Memory can reduce repetition, but current facts should remain explicit enough for you to inspect.

Sources: Anthropic · Claude release notes, including August 25 memory update

Understand memory and current context

Study tools are becoming a learning workspace

Google’s June 25 education announcement describes study notebooks that begin with a learning goal or supplied materials, use a diagnostic quiz and build interactive lessons. It also describes a connection with NotebookLM. The launch distinguished personal-account and school-account rollout conditions. Its August back-to-school collection adds current education updates and offers for eligible students. Check your actual account and local eligibility instead of assuming that every announced feature or promotion applies to you.

The feature names are less important than the learning loop. Try an explanation, make an attempt, inspect the error and answer a different question. A beautiful study guide can still leave you unable to solve the next problem. Treat generated quizzes as material to review, and keep your course’s definitions and rules in view.

Start with our fictional cell note rather than uploading a whole course. It gives you enough source material to check whether the assistant corrects the actual misunderstanding. Then close the conversation and try the transfer question. That small step makes the difference between feeling familiar with an explanation and having evidence that you can use the idea.

Sources: Google · Connected student learning tools, June 25, 2026 · Google · Back to School 2026, August 19

Follow the student path and try recall practice

Creative tools need a brief that says what is real

Google’s August roundup also describes new Gemini Live productivity features and expanded video creation controls in Gemini Omni 1.1 Flash. The range is striking: conversation, actions and generated media are appearing in the same broad product family. Availability still depends on the particular surface and account. A capability described in a developer model announcement is not automatically a button in every consumer app.

For a business owner, the useful starting point is a clear service explanation. A generated diagram can show the steps from assessment to quote. A generated customer photo and invented sales claim would communicate something else entirely. Put “illustration” in the brief when the material is illustrative, and keep actual customer evidence separate.

Try preparing three distinct pieces from one approved service question: a website FAQ, a short public explanation and a reply template. The task can be useful without requiring a video at all. Add a visual when it makes the process easier to understand, and inspect the text, image and caption together for accidental promises.

Sources: Google · August 2026 AI roundup, published September 1

Build a useful three-piece content plan

Google’s website-owner announcement, updated August 31, describes a Search Console control for generative AI Search features and reporting to help owners understand that visibility. The announcement also points to updated guidance. These are changes worth knowing about if you run a website; they do not turn a new page into a guaranteed citation or a guaranteed customer.

The content work is concrete. Explain what the service includes, how the quote works and what the visitor should do next. Put an answer on the page because it resolves a customer question. Do not create a dozen nearly identical town pages or an invented case study just to give a crawler more text.

Our suggested exercise is to read one service page as someone unfamiliar with the business. Can you tell whether the service fits, what happens after the form is submitted and which claims have evidence? Improve the missing answer first. Search visibility can then be observed over time; it should not be claimed as the outcome of writing the draft.

Sources: Google · Website-owner controls and insights, updated August 31, 2026

Start with an honest FAQ and service page

A family activity can be bigger than the screen

There is another way to respond to the growing range of AI tools: let an adult use one to prepare a small shared activity, then put the screen aside. A story map, a paper game or a fictional fact-checking packet gives children something to choose, explain and make. This is our editorial suggestion, not a claim about a new child-facing product.

Try the cloud-library story. The family chooses which clue Pip the moth follows and invents an ending that uses it. An adult can prepare additional scenes with a permitted tool, review them and stop when the story is ready. No child account, personal photograph or exact location is needed for the exercise.

The result is easy to keep: a folded booklet with drawings and the family’s decisions. If you want a second activity, play fact detective with a short source card. Ask everyone to distinguish what is supported, what is contradicted and what the card does not establish. Those habits remain useful when the next product name arrives.

Make something together

Copilot is changing the app around the model

Microsoft’s current support guidance describes an updated Copilot app that brings personal and work or school sign-in into the same application while keeping those account experiences separate. It also identifies consumer features being retired from August 18, including Group Chat, Podcasts and Deep Research. The guidance says content from retired features may be handled differently from ordinary chat history. A familiar product name therefore does not guarantee that an older tutorial still describes the current experience.

The useful action is to inspect the account and feature you actually use. Check which account is selected, which files it can access and whether the workflow in your saved instructions still exists. If you rely on a particular artifact, follow the provider’s current guidance for finding or retaining it. Do not assume that every kind of content moved in the same way, or that a work account and a personal account share the same permissions.

Our suggested experiment is a small continuity check. Take an ordinary fictional project note and ask the tool to produce a short handoff containing current state, source references and the next action. Open the result outside the chat where possible. This tests whether you can keep the useful work even while the surrounding interface changes. It does not require moving private work into a personal account.

Sources: Microsoft Support · Copilot app and account changes, checked September 7, 2026

Make a handoff that preserves the work

The spreadsheet is becoming a place to inspect the work

Microsoft’s August 25 Copilot release-note batch describes Python-assisted editing in Excel, with analysis results written into the workbook. The same batch includes Outlook emails as notebook references and changes to the notebook interface. Microsoft describes gradual feature rollout, so the release note should be read alongside the actual account and platform. These are documented capabilities, not results from a Power of AI test.

A practical trial can stay very small. Use fictional rows where two completed requests took 30 and 50 minutes and an open request has no duration. Ask for the average recorded time for completed requests, the admitted rows and the formula or code used. The expected arithmetic is 80 divided by two, or 40 minutes. A chart is useful only after its inputs and denominator answer that question.

Then make every request open. A useful result says there are no completed records to average; it should not substitute zero minutes. Keep the original workbook or a working copy, inspect the actual cells and record any correction. The model’s ability to write code makes the method more inspectable when it exposes that code clearly. It does not make the method correct by default.

Sources: Microsoft Learn · Copilot release notes, August 25, 2026 batch

A request you can adapt

Using these fictional records, calculate the mean recorded duration for completed requests only. Show the admitted rows, denominator and formula or code. Keep missing durations distinct from zero. Then explain what the result should be if no completed records remain. Do not change the original data or claim a business outcome.

Check a spreadsheet question with a small reference

Turn the news into something you can practise here

The expanded audience paths now include local workshops with fictional source packets. A student can explain a library reservation queue; a worker can prepare a handoff; an owner can draft a customer answer; a personal project can be reduced to a workable first version; a family can connect three comic panels. These are authored exercises, not live AI demonstrations.

Each workshop gives you room for a first attempt, a revision and a response to a changed condition. Open the example when you want to compare reasoning, then write an evidence note beside your own work. The site does not grade the result. Copy or download explicitly before leaving, because the drafts remain only in the current page visit.

That gives the news a practical destination. You can understand the task and its checks here, then take a coaching request and the same fictional material into an assistant you are permitted to use. Compare what the assistant does with the work you can already inspect. Keep a useful correction as readily as a successful draft.

Try the explanation workshop

Choose one experiment for your actual week

Pick the path that fits the work in front of you: student, employee, owner, personal project or family activity. Choose one playbook and keep its example small. Save your attempt and write down the check that changed your mind. You can finish with a useful correction, a clearer question or a small working artifact.

This edition was checked against the linked primary sources on September 7, 2026. It combines recent announcements with earlier developments that remain relevant to the practical theme; the source dates are explicit. The Signal is written when an edition is ready. It is not an automatically refreshed news feed, and the suggested experiments are not presented as completed tests.

Choose a small personal project

Take this into your next task

Choose the change that helps one real task, keep the inputs small and inspect the result. A useful experiment is a better way to keep up than trying every announcement at once.

Build a project brief.

Questions people ask

Are these hands-on model test results?

No. Product facts come from the linked primary sources. The experiments and examples are original editorial suggestions for readers to try.

Is every feature available in my account?

No. Check the specific app, plan, region, organization settings and rollout conditions in the provider’s current documentation.

How often does The Signal publish?

When an edition is ready, with an explicit date. There is no automatic weekly publishing promise.

Sources and editorial notes

Reviewed 2026-09-07. Product capabilities depend on the app, plan, region and workspace settings. Workflows and prompts are authored teaching material, not recorded model results or measured time-saving claims.

Keep going

Related Power of AI pages

Keep reading with Start here, everyday uses, the tool guide, the writing workshop, and sources and standards.