Update since the previous issue

Issue 010 · Period covered 6 October 2026 — 24 October 2026 · as of 14:21 UTC · published 24 October 2026 · 12 min read

What's new in AI since October 5

October 6 to 24: models, documents, workplace systems and agent shopping, with dates and access status.

Issue 01024 October 2026
6-24October: updates since issue 009

What's new in AI since October 5

Official releases · documentation · research

Since the previous issue, new models and tools have arrived for repetitive tasks, document review, spreadsheets, presentations and shopping through AI agents, assistants that can take actions on your behalf. This issue covers the period since our last edition. The selection reflects information available when we prepared it and does not claim to cover every release.

Each brief gives the primary source's date. “Released” means a capability is available, “in testing” means access is limited, and “announced” means a launch is still planned. That distinction matters when choosing a tool.

What actually changed

According to Walmart, its October 21 announcement introduces separate profiles for shopping agents and an Agent Gateway for connecting to the store. Profiles with permissions and spending limits are planned for the following month; the gateway will begin with a beta for companies and developers. General access is still pending, and no sales growth is demonstrated here.

Walmart illustrates the gap between a service description and the ability to use it. Before changing your workflow or planning a purchase, establish what is available to you.

For each update, establish what you can use now, which task it changes and what result would demonstrate its value.

The checks below are editorial suggestions. Choose one that fits your task: familiar work where you can spot an error and compare the result with your current method. You do not need to test every update.

Briefs: the week's signals

01. October 6: Anthropic separates access levels for cybersecurity work

According to Anthropic, its expanded Cyber Verification Program has three access tiers with different checks and restrictions. Eligible professionals and organizations can apply; enhanced permissions are not available to every Claude user. Testing someone else's systems still requires authorization.

If an authorized security task is blocked, describe it and compare it with the program's tiers. Apply when the relevant tier covers the task and your team can document authorization and meet retention conditions. Clarify uncertain terms with the provider first.

02. October 7: Claude Haiku 5.5 is released for frequent, short tasks

According to Anthropic, Haiku 5.5 is designed for high-volume work such as summaries, classification and quick data queries. The company also reduced the price of reusing stored context with Sonnet 5.5. Speed and savings comparisons come from the vendor; they do not guarantee the same gains on your work.

Choose several routine short tasks and one difficult case. Give the same instructions to your current model and Haiku; record errors, time including corrections, and cost. Continue the trial if important errors have not increased and the gains survive review.

03. October 13: Gemini 4 Argon becomes more widely available

According to Google, Gemini 4 Argon is generally available through the API, the interface for connecting a model to software. Its Gemini app rollout starts with Google AI Pro and Ultra subscribers. Google positions it for complex tasks. API availability and a paid app rollout do not mean free access for everyone.

If your model loses requirements in long tasks, choose a completed project with a known result and a requirements list. Repeat it with Argon if you have access. Compare omissions and correction time: continue testing if rework falls without losing important requirements.

04. October 13: Shopify connects product search with payment

According to Shopify, Shop MCP lets compatible agents search products, build carts and complete checkout. MCP connects an assistant to tools and data. To use saved Shop Pay details, the shopper signs in to Shop and confirms the total and payment method. The tool is available; catalog inclusion does not guarantee orders.

If you sell through Shopify and have access to a compatible agent, find your product as a shopper. Check the variant, price and availability against your store; reach the confirmation screen without paying. Save any mismatch for support or your integrator. Test a real order once those mismatches are resolved.

05. October 14: Google opens Deep Review for document checks

According to Google Research, Deep Review is publicly available for complex document feedback, initially accepting PDF and plain text. It grew out of the research Paper Assistant Tool. Model feedback still needs expert review. The public service says uploaded documents are retained for a year.

Use a nonconfidential document containing a known logical error. List valid and false objections and time your review of them. Continue if the service finds the error and review takes less time than your usual check. First make sure a year of retention is acceptable for this text.

06. October 16: OpenAI agents gain tools for documents and spreadsheets

According to the official OpenAI API changelog, hosted Agents API environments now include the Artifact SDK, a set of software tools, and skills for creating and editing spreadsheets, presentations and Word documents. This update is for developers. It does not specify access on particular ChatGPT plans; generated files still need accuracy checks.

Ask the assistant's developer to build a familiar report from test data. Open the file in your usual software, check calculations against a reference, inspect labels and page breaks, and change an input value. If formulas or editing fail, the report is not ready for colleagues.

According to Shopify, compatible cart and checkout integrations can now carry a buyer's privacy and marketing choices. Both sides of the integration must support the protocol extension and agree to use it. Passing on that choice does not turn agreement to a purchase into consent to marketing messages.

If you already use such an integration, ask your integrator to demonstrate consent, refusal and no-choice tests. Check each result against the store's record. If they differ, do not use that record for marketing; fix how the choice is passed on first.

08. October 19: Accenture highlights the gap between use and process redesign

According to Accenture, 43% of surveyed employees use AI daily, while 15% of executives report redesigning end-to-end processes. The survey covered 3,000 executives and 3,000 employees in July and August. These are different groups reporting on their own work. The survey does not measure the effects of October deployments.

Trace a recent report from assignment to acceptance: preparation, generation, review and time waiting for approval. Choose the longest stage you can improve and change it in the next trial. Expand AI use when the whole process becomes shorter at the same quality.

09. October 20: Ampersand is released for group collaboration

According to humans&, Ampersand is available through an API for group conversations, choosing whether to reply or stay quiet. The developer evaluates it partly with simulated users and acknowledges the cost of unwanted interventions. Those tests do not establish an advantage in real teams.

Start in a separate discussion with no customer commitments. Agree on useful interventions and record helpful clarifications and unnecessary messages. Compare with a similar discussion without the bot: if participants spend more time distracted or correcting it, narrow the assistant's role.

10. October 21: Charter examines employee involvement in AI adoption

The new Charter AI Pulse finds that employees who help choose how AI is used rate its benefits more highly. Its public page summarizes survey and interview findings; this selection did not verify the full methodology. An association between answers does not establish that participation caused the benefit.

Ask an employee and manager to choose a task and success criterion together: for example, a report accepted with fewer revisions. After the trial, each assesses the result separately. If they disagree, examine a specific example before expanding the test.

11. October 21: SAP offers sales and service assistants to selected customers

According to SAP, the Customer Experience update includes assistants for qualifying sales leads, handling service cases and supporting marketing. The new capabilities are available to selected customers. The vendor's feature descriptions do not establish general availability or independently verified returns.

If your company uses SAP, confirm access first. Then pass a test case, including the customer's conditions and conversation history, between stages. Compare what the receiving employee sees with the original. Fix any lost condition before using the workflow on live cases.

12. October 21: Oracle announces an industry agent studio

According to Oracle, AI Agent Studio for Industry Applications is intended to combine agent creation, testing and monitoring. A footnote schedules availability within the next 12 months, initially in the US. The platform is not yet open for general use.

Ask Oracle for an access date, supported applications and initial launch restrictions for your use case. Record what is confirmed and what remains a promise. Until access and the application you need are confirmed, keep the product on a watch list and retain your current method.

13. October 22: Decisions API leaves public beta

According to OpenAI, Decisions API is generally available following the beta launched on October 6. Its documentation describes returning a category, score or estimated probability from text and images. Applications can use these decisions, but accuracy needs testing on your own data.

Use requests with known correct routes, including borderline complaints. Compare Decisions with your current sorting method. Record missed complaints and routine requests incorrectly marked urgent separately. If dangerous misses increase, keep routing with the current process; revise the rules and test again.

14. October 22: Microsoft publishes tools for interactive charts

According to Microsoft Research, Flint-interactive lets developers define how selection and filtering connect interactive charts. Tools and examples are published. A chart's interactivity does not establish that its data, calculations or conclusions are correct.

Ask an analyst to recreate a familiar report. Calculate the expected total for a chosen filter by hand and compare it with the chart; check the label and units. If the calculation or meaning of the measure differs, fix the chart before using it for decisions.

15. October 23: Forter publishes observations of agent orders

According to Forter, agents navigating store pages accounted for 70% of detected agent orders on its network in the week beginning October 12. That share covers only detected agent orders on this network. An agent may have started shopping before the system learned to identify it reliably.

Ask your analyst to show visits, placed orders and fulfilled orders separately, along with the rule used to identify an agent. If the rule changed, ask whether the earlier period can be recalculated using the new rule. If it cannot, mark the data as noncomparable and defer conclusions about channel growth.

What became cheaper, and what became more valuable

New tools may reduce the time needed for a summary, document, shopping cart or chart. In your chosen test, add up preparation, tool use, review and corrections. Compare that total with your previous method on the same task.

Define an acceptable result in advance: a spreadsheet with correct calculations, a cart with the right product and price, or a report that meets every requirement. A fast draft can still be costly to finish.

How to read this

The dates above refer to releases, announcements and publications. Research data may have been collected earlier; product access may change later.

For your chosen update, record three statuses: announced, available to us, and tested on our task. Walmart and Oracle still call for checking dates and conditions. Available tools can be compared with the way you already work.

People: work and accountability

Choose a task that often needs rework and keep its previous result for comparison. Managers should agree with the person doing the work on unacceptable errors and who will accept the result.

After the trial, discuss specific corrections. If a report arrives faster but takes longer to approve, find the cause of the delay first.

Business: decisions, economics and risk

Put tools you can access on a test list. Keep planned launches on a watch list with a date to check again.

For your chosen task, compare tool costs and staff time through acceptance of the result. Continue if the gains hold at the required quality. In commerce, check fulfillment separately: the confirmation screen shows only part of the process.

Trust: boundaries, verifiability and consequences

For the test, record the permitted action and when a person takes over. For example, the assistant prepares a file; an employee checks the calculations and decides whether to send it to the customer. Name the person who will handle an error.

For purchases, check the item, payment and marketing consent separately. Order confirmation does not constitute consent to messages.

Working map for the week

Choose the row that fits your task. One test is enough.

Decision When it applies First move Boundary
Test document review Finding errors takes too long Use text suitable for upload with a known error; check the feedback and review time Continue if the error is found and review is faster; a specialist verifies the feedback
Check a product through an agent You sell through Shopify and have access to a compatible agent Find a product and check its variant, price and availability before paying Send mismatches to support; postpone a purchase test until they are fixed
Agree on a test with an employee A team's process is changing Choose a task and criterion together; assess the trial separately If assessments differ, examine an example before expanding the test

Continue in the Practicum

Use the evaluation set builder to record familiar cases, expected results and an error you cannot afford to miss. This gives your comparison a reference.

In the Agent Contract, record the permitted action, when a person takes over and who is responsible. This is your team's agreement; an outside shopper's authority needs a separate check.

One sensible next move

Choose one update and complete this note: “We will test ___ on ___. We will compare it with ___. ___ will accept the result. We will continue if ___ and stop if ___.” Set the criteria before the test. If access is unavailable, clarify the conditions and set a date to check again.

Confidence and sources

This is a selection of primary sources published since the previous issue. Links sit beside the facts. Releases and documentation establish stated functions and access rules but do not independently demonstrate value. Accenture and Charter report participant responses; these do not establish the causal effects of adoption. Forter's observations concern its own network.

For Charter, we used the public findings page without verifying the full methodology. Ampersand tests include simulated users. OpenAI's access status comes from the newer dated changelog: a search copy of the guide still described beta access. Recheck these conditions before implementation.

The proposed checks and criteria are editorial recommendations. They help readers choose a next step but do not themselves prove a tool's value. Community discussion was not used to establish facts.

What's new in AI since October 5