The Future Last Week: AI Is Getting Closer to the Work. Show Me the Receipt.

The big AI story this week was not a chatbot writing a nicer answer. A newly released model can work through screens at a lower published price. Oracle and ServiceNow described ways for AI to reach the records where a company actually operates. Microsoft offered the software around the model that helps an agent use tools and finish a job. Meanwhile, research claims that could reshape parts of mathematics made one point impossible to dodge: producing an impressive artifact and proving what it means are separate jobs.

That is an encouraging week for business adoption. It is also a good week to ask for a receipt. Did the right record change? Did a person approve the consequential action? Did the supposed saving survive the cost of checking? The announcements below landed October 26–30 and were checked on Saturday, October 31. Supplier comparisons and customer stories are identified as such; a promised integration is not a live result.

The week in one minute

• OpenAI introduced GPT-6.1 Terra on October 26, emphasizing computer use and lower published application-programming-interface prices than its preceding Terra model. Its benchmark results are OpenAI measurements, not a guarantee for your workflow.
• OpenAI and Oracle announced connections between ChatGPT and Oracle applications on October 27; OpenAI and ServiceNow announced a broader workflow partnership on October 28. Some capabilities are live, while others are scheduled for 2027.
• Microsoft released its Copilot harness through Foundry on October 28. A harness is the software around a model that supplies instructions, tools, an execution environment and controls. This matters more than another model leaderboard.
• OpenAI announced three AI-produced scientific and mathematical results on October 26 and invited outside groups to examine them. Formalization and released manuscripts are substantial artifacts; broad independent understanding is still being developed.
• NVIDIA and Oracle said some Vera Rubin systems are available on Oracle Cloud and outlined a much larger buildout. Available hardware and future multi-gigawatt plans are different things.
• OpenAI said it will honor a shared website AI-training preference signal on October 30. The proposed mechanism is an Internet-Draft, and OpenAI says implementation will come in the following months.

1. A cheaper computer-using model changes the pilot math

GPT-6.1 Terra is designed to read screens and carry out steps in software as well as write and reason. OpenAI says its standard developer price is $1 per million input tokens and $5 per million output tokens, half its prior Terra standard rates; cached input has a separate lower rate. A token is a small unit of text the model processes. The model is rolling out in ChatGPT Work and Codex, and through the OpenAI API, although access depends on plan and settings. This makes a familiar business idea more affordable to test: give an agent a clearly limited task, such as reading an invoice and entering fields into accounting software, then inspect the saved record.

OpenAI reports that Terra improved sharply over GPT-5.6 Terra on its computer-use tests and completed some tasks faster than larger models. Its system-card addendum also says tests took place in research or API settings that may differ from production ChatGPT. The gap is important. A benchmark can tell you where to look; it cannot tell you whether your invoice fields, exception handling and approval rules will work. Even OpenAI says its more capable models remain a better choice for the most complex coding work.

Real, hype, and who should care

Real: a faster, lower-priced option is available for teams that can safely test computer-based work. Hype: “a lower price per token means a cheaper process.” If a model retries more often, changes a wrong field, or takes longer to review, its total cost per accepted case can rise. This applies most to high-volume administrative work with a clear expected output and a reversible action. A small company using AI for an occasional memo may see no meaningful savings. Finance, operations and customer-service leaders should put 30 ordinary cases and five awkward exceptions through both the current model and Terra. Count correct saved records, corrections, elapsed time, reviewer minutes and actual bill. Stop the pilot if record accuracy or review burden worsens.

The signal to watch is not another supplier chart. It is the cost of a completed, approved assignment in your system, including the human check. The AI business-case guide explains why released staff time and money removed from a budget need separate treatment.

2. The assistant is reaching the systems where work happens

OpenAI and Oracle said on October 27 that they are bringing Oracle applications and data into ChatGPT and adding OpenAI capabilities to Oracle products. One of Oracle’s company-reported internal examples is recruiting research that previously took two to four days to compile and now takes 15–20 minutes of preparation. That is a reported workflow result, not a controlled measure of hiring quality or a promise that other organizations will see the same gain. The more consequential question for an executive is whether a person can move from asking a question to updating a governed business record without losing the context, permissions or audit trail.

The ServiceNow announcement on October 28 makes that question concrete. A connected user could find similar technology incidents, assess a proposed change and update a ServiceNow record from ChatGPT. ServiceNow says existing business rules and permissions govern the available actions. The current plugin is available in OpenAI’s marketplace, but many enhanced features are planned for late 2026 or 2027; a voice experience, security tools and a software-building integration have separate future dates. Treat a roadmap screenshot as a roadmap, not evidence that your chosen workflow is operating today.

Microsoft’s October 28 Copilot harness release points in the same direction from a different vendor. The harness is the working environment around a model: instructions, connections to tools, context, controls and a place to execute steps. Microsoft says it supports more than 20 models and is available through Foundry. Its authors, Tina Schuchman and Vladimir Fedorov, argue that keeping this layer working across changing models is hard. That is a vendor case for a managed product, but the underlying management problem is real. A better model is of little use if it cannot find the right record, respect a permission, recover from a failed step or leave a readable history.

Real, hype, and who should care

Real: multiple major suppliers now offer or plan connections between AI and ordinary business systems. Hype: treating “connected” as “safe to act.” If your company runs Oracle or ServiceNow, choose one narrow process with an owner. Ask the supplier to show the exact actions available under a normal user account, a restricted account and an administrator account. Ask what happens when the source record is stale, a change is rejected, or an agent has to stop halfway through. Require a before-and-after record and a named human approver for any external message, payment or material change. If the demo cannot show a failed and recovered case, the production case is unfinished.

Watch for production references that report completed work and corrected errors, not just employee access. The previous issue covered Uber’s company-wide AI rollout; this week moves the story from who gets a seat to which systems can be touched.

3. The research claims are extraordinary. So is the verification job.

OpenAI said on October 26 that its internal models produced results in three areas: a mathematics result involving a long-standing question about how wave behavior concentrates, a claimed disproof of the Exponential Time Hypothesis in computer science, and mathematical existence results in theoretical physics. The Exponential Time Hypothesis is an assumption about how hard certain computation problems are in the worst case; disproving it would matter enormously to theory, but does not mean a company can suddenly solve every hard scheduling or logistics problem quickly. OpenAI released writing and computer-checked Lean files for the first two areas and says outside scholars are organizing examination and explanation. For the physics work, its external group asked to hold materials while they prepare their own account.

A proof assistant such as Lean checks whether a formal chain of steps follows its stated rules. That is a meaningful check, but it does not by itself tell a non-specialist exactly how broad the conclusion is, how it connects to decades of literature, or which practical algorithms follow. OpenAI itself calls the manuscripts raw research data and says outside groups are needed to turn them into shared scientific understanding. The computer-science group has asked that raw material be public for further verification. A public computer-science discussion contains both excitement about the possible result and requests for careful scrutiny; those posts are reactions, not a scientific vote.

Real, hype, and who should care

Real: openly available artifacts, some formal checking, and named external experts create a path for careful scrutiny. Hype: announcing that the field has already accepted all three claims, or telling a business buyer that a theoretical proof cuts operating cost next quarter. Most executives should not procure a math model because of these headlines. They should learn the process lesson: when AI produces more candidate answers, the bottleneck moves to review, replication and explanation. In research-intensive companies, reserve expert capacity to check the few leads that would actually change a program. Watch whether outside scholars publish explanations or corrections and whether any practical method emerges from the theoretical work.

4. The computing buildout is real; the schedule is a proposal

NVIDIA and Oracle announced on October 27 that Vera Rubin systems and Vera central processors are generally available through Oracle Cloud Infrastructure. They also outlined plans to expand large AI clusters to multi-gigawatt scale. A gigawatt is one billion watts of electrical capacity. The availability of a system today and a plan to build several gigawatts of future infrastructure should never be combined into one delivered-capacity number. The companies say the extra processors support work between model calls, such as retrieving data and running software. That is a useful reminder that AI cost is not just the accelerator chip or the model price.

The physical limits remain worth watching. Associated Press reporting on October 21 documented neighbors asking about electric bills, water, tax deals and local consultation. A vendor capacity announcement does not prove when a particular facility will deliver usable service at a stable price. Nor does local concern mean the buildout has no value; it means its costs and benefits need a real address and a real plan.

Real, hype, and who should care

Real: new hardware is being offered, and suppliers are investing for larger demand. Hype: equating a multi-gigawatt ambition with power connected, cooling operating and customers served. A company buying a large committed AI workload should ask its cloud provider which region is available now, when reserved capacity becomes usable, what price or service terms can change, and how the provider handles a delayed site. A smaller buyer should avoid long commitments until the workload and accepted-output economics are clear. The next signal is commissioned capacity and a customer service record, not a ribbon-cutting rendering.

5. A small web rule could become a big publishing decision

OpenAI said on October 30 it intends to honor a common way for website operators to state whether they want content used for foundational AI training. The mechanism uses robots.txt, a file websites already use to communicate crawler instructions. The published Internet-Draft describes a per-purpose token in an ordinary robots.txt group. It is a draft, not a completed internet standard; OpenAI says it aims to implement the signal in the coming months. The announcement says its existing crawler controls remain in effect.

For a publisher, this is less about a technical line of text than a business choice. Some pages need broad discoverability; others contain paid research, licensed material or work whose reuse should be limited. The proposed signal expresses a preference about training, not a guarantee of payment, licensing or enforcement. It does not mean every AI provider already follows it. It also should not be confused with hiding pages from search. The ordinary question for a content owner is: which material do we want widely found, which use do we want to limit, and who will keep the rule accurate as our site changes?

Real, hype, and who should care

Real: a common signal could reduce the work of maintaining separate provider rules. Hype: assuming a draft immediately controls every crawler or resolves the debate over training rights. This applies to research firms, publishers, training businesses and companies with valuable public knowledge bases. Ask the web owner and legal or licensing owner to review existing crawler settings and content rights before changing anything. Watch which providers actually adopt the signal and whether the draft changes before becoming a standard. In the meantime, avoid a rushed robots.txt change that accidentally blocks the pages customers need to find.

What people are saying

The outside reactions are more useful than a winner-take-all verdict. Microsoft’s Schuchman and Fedorov argue that the hard part is maintaining the software around changing models. OpenAI’s research announcement acknowledges that external scholars may disagree about how AI-produced work should be released and understood. A computer-science discussion contains both enthusiasm about a possible major result and calls for scrutiny. These public posts show individual reactions, not representative evidence. I checked public X search during reporting but did not find an attributable post that added verifiable information beyond the primary documents; none is presented here as proof.

Five useful moves for Monday

1. Give one process owner a narrow, reversible computer-based task. Have that person collect 30 ordinary cases and five difficult ones, with the expected final record for each. Compare your current model with a lower-cost one on accepted outputs and total reviewer time.
2. If you use Oracle, ServiceNow or a similar system, request a live demonstration under three permission levels. Ask to see a rejected action, a stopped half-complete action and the record of who approved the final change.
3. Ask your AI team to label new research outputs as candidate leads, formal checks, independently reviewed findings or validated business outcomes. Fund the human verification that moves a result from one stage to the next.
4. For a material cloud commitment, ask what computing capacity is serving customers now, what is merely planned, and what happens contractually if a facility or grid connection is delayed.
5. Have the web and content owners inventory your current robots.txt rules and the rights in your high-value pages. Record a desired training-use policy before changing crawler instructions; revisit the shared signal when providers implement it.

My optimistic read is that AI is becoming a practical worker in more kinds of software, at a price that lets more teams test it. The discipline that turns that into value is almost stubbornly ordinary: define the job, see the record it changed, count the corrections, and keep the person who can say yes or stop. The science story reminds us that even a beautiful result needs other people to make it trustworthy.

Source and method notes

Checked October 31, 2026. Model performance and prices, product availability and roadmaps, internal customer results, hardware capacity plans and research claims are attributed to their announcing organizations. The mathematics and science claims are not presented as independently settled conclusions. Associated Press reporting supplies the external infrastructure context. Public forum discussions are firsthand reactions only, not samples of users or experts. The training-preference proposal is identified as a draft and future implementation rather than a live standard. The two images are AI-generated metaphors, not documentary scenes. Links beside consequential claims lead to the underlying release or reporting.