Previously, we asked AI: "How do I fill out this form?"

Now AI can directly understand screens, operate computers, and even assist in completing cross-system tasks.

Meanwhile, Google is pushing AI-generated video further—not just playable videos, but interactive worlds you can step into, change the weather, move the camera, and have the environment respond continuously.

On October 28, 2026, three AI updates worth noting point to three distinct directions:

Reducing the cost of AI work execution, enabling real-time interactive generated content, and preparing enterprise software to make AI a new interface for operations.

1. OpenAI Releases GPT-6.1 Terra: Price Halved, What’s New in AI Computer Operation?

On October 26 (US time), OpenAI officially launched GPT-6.1 Terra.

This model focuses on three key capabilities:

  • Understanding images, documents, and computer screens.
  • Performing multi-step computer operations.
  • Handling large volumes of routine professional tasks at lower costs.

It is positioned not to replace all advanced models but to provide a more efficient option balancing speed, capability, and cost.

For example, a company receiving many supplier invoices daily.

Before, employees would manually review each invoice, extract item details and amounts, then enter the data into internal systems.

Now, with supported tools and permissions, AI workflows can have the model understand invoice content and assist software operations, filling forms and preparing for subsequent processing.

However, this reflects what the model can assist with, not that a general chat window command can automatically operate all corporate systems.

What is the API price for GPT-6.1 Terra?

According to OpenAI’s official pricing:

Input per million tokens: $1
Cached input per million tokens: $0.05
Output per million tokens: $5

OpenAI states the standard API prices are about 50% lower than for GPT-5.6 Terra and GPT-6.1 Sol.

These rates apply for developers using the API.

This does not directly translate to ChatGPT Plus subscription fees being halved or changes to individual user subscriptions.

The total cost of each AI task also depends on token usage, retries, and external tool expenses.

Does the lower price mean lower capability?

OpenAI’s benchmarks show significant improvements over the previous generation.

For example, in Terminal-Bench 4.0 software engineering tests at maximum reasoning settings:

  • GPT-5.6 Terra scored 21.5%.
  • GPT-6.1 Terra scored 50.3%.

Additionally, in the Agents' Last Exam professional evaluation, GPT-6.1 Terra at medium reasoning strength outperformed GPT-5.6 Terra by about 15 percentage points and reduced testing costs.

In OpenAI’s own difficult question tests, the error rate at low reasoning intensity dropped from 53.6% to 16.7% compared to the prior model.

These are official results under specific test conditions.

They do not guarantee that Terra is superior in all tasks nor that it only makes mistakes 16.7% of the time in real work.

For complex programming and high-difficulty reasoning tasks, OpenAI suggests GPT-6.1 Sol and GPT-6 Astra might be more suitable.

Can general ChatGPT users select Terra now?

Usage contexts vary.

OpenAI states GPT-6.1 Terra is rolling out to Plus, Pro, Business, Enterprise, and Edu plans, but the primary access points are ChatGPT Work and Codex.

Developers can also access it through the API under the model name gpt-6.1-terra.

This does not mean the general ChatGPT chat mode has fully incorporated Terra yet.

Availability depends on plan, account rollout, region, and workspace permissions.

OpenAI also notes that overall usage limits have not increased directly due to this launch.

What meaningful changes does this bring for ordinary workers?

Previously, AI choice was often about which gave the smartest answers.

If a task involves large volumes of repetitive data sorting, screen verification, and system operations, answer quality alone is insufficient for evaluation.

One must also consider:

Task completion rate, cost per task, execution time, and required manual checks.

A slightly less capable model that completes routine tasks quickly and reliably may be more practical than always using the most expensive state-of-the-art model.

However, AI controlling computers means mistakes are no longer just wrong text responses.

AI could alter data, input wrong fields, or click inappropriate functions.

Therefore, tasks involving payments, data deletion, external communications, or critical system changes should still have clear human oversight and stop conditions based on risk.

2. Google DeepMind Releases Genie 4: AI Generates Not Just Videos but Interactive Worlds

On October 27, Google DeepMind introduced Genie 4.

This is a World Model—a simulation of a world.

The key difference from typical AI video generation tools is content that changes continuously with user interaction.

For instance, a user describes a seaside scene.

Traditional AI video tools generate a fixed playable clip.

In Genie 4’s research demo, users can move with keyboard inputs, adjust the camera, or input new commands to change the scene.

For example:

"Change the weather to rainy."

"Move the camera behind the building."

"Add a new character."

Genie 4 attempts to extend the simulation within the existing environment, not just generate an entirely separate video clip.

What new abilities does Genie 4 have?

Google DeepMind emphasizes four main directions.

First, real-time world adjustments.

Users can alter weather, camera angle, and objects during simulation, with the model generating accordingly.

Second, more natural character and physical interactions.

Google demonstrated character movements and some collisions, aerodynamics, and other physics elements.

However, this simulation does not guarantee all visuals fully obey real physical laws.

Third, native audio generation.

Genie 4 is Google DeepMind’s first world model that can generate native audio in real time.

Sounds such as wind rustling leaves, ocean waves, character dialogues, and environmental noise are generated alongside visuals.

Fourth, collaboration with Gemini reasoning models.

Google states Genie 4 integrates Gemini family capabilities.

In some demos, Gemini helps break down tasks, monitor progress, and adjust scenes, enabling simulations to support multi-step goals beyond mere visual changes.

Is Genie 4 the same as Google Veo or Flow?

No.

Veo and Flow mainly focus on creating and arranging video content.

Genie 4’s research focus is an explorable, real-time adaptable, and continuously responsive simulated world.

Traditional video production usually preplans camera angles and actions before generating footage.

But if creators can step into a simulated scene to change lighting, positioning, and camera angles directly, they can compare ideas faster before formal production.

Google says trusted testers from film, gaming, news, education, and engineering fields are already using Genie 4 for scene previewing and interactive prototyping.

This suggests Genie 4 may become a future previsualization tool.

But it remains a research preview, not a mature commercial video production substitute.

Is Genie 4 available to everyone now?

No.

Google confirms it is currently in preview for trusted research testers.

Plans include gradual availability for Google AI Ultra subscribers through Project Genie experiences.

Some early Gemini API and Google AI Studio partners will also gain access in follow-up phases.

The official waiting list is currently open, but it is incorrect to say all Ultra users have access today.

What are Genie 4’s current limitations?

Google admits this is still an early research model.

It can maintain moderate environment consistency during short interactions, but over longer periods, object appearance and positions may drift.

After multiple complex scene modifications, characters, objects, and spatial relationships may become inconsistent.

Other points to improve include:

  • Accuracy of complex physical interactions.
  • Long-term scene coherence.
  • Real-time audio fidelity.
  • Precision of spatial location and distances.
  • Quality of generated readable text.

Additionally, Google says Genie 4’s generated media will include SynthID invisible digital watermarks.

The significance of this release is not that AI can generate fully accurate virtual worlds yet.

Rather, video generation is evolving toward controllable, explorable environments that can serve as AI training grounds.

Aside from film and gaming, this ability could be applied in the future for robot or AI agent scenario testing.

However, whether it can reliably serve as a training tool depends on ongoing validation of the simulation’s real-world fidelity.

3. Oracle Expands AI Partnerships with OpenAI and Google: Can Enterprises Really Manage Business through ChatGPT?

Also on October 27, Oracle AI World announced a series of AI-related collaboration updates.

The most relevant for general workers is Oracle’s announcement to expand cooperation with OpenAI.

The goal is not just embedding ChatGPT into company computers.

It aims to make Oracle-managed finance, procurement, database, and enterprise applications accessible as authorized work systems that AI can query and use.

What problem does Oracle AI Workspace solve?

Imagine a new finance team member investigating why a procurement expense increased.

They might need to open:

Financial systems to check amounts.

Procurement systems to review orders.

Budget systems to verify plans.

Then compile the findings into a report.

Oracle and OpenAI’s announced vision is that users can describe queries in ChatGPT.

Oracle AI Workspace then connects to relevant enterprise apps and data.

The goal is to help find information, organize work, and complete authorized business processes.

Oracle explains this framework will connect Oracle Fusion Applications, NetSuite, Oracle AI Database, and Oracle Cloud Infrastructure.

However, this is a new product direction.

This does not mean every ChatGPT account can freely access Oracle enterprise data today.

Access and permitted actions depend on each company’s deployment, permissions, and approval mechanisms.

What other AI services did Oracle announce?

Besides Oracle AI Workspace, other product directions revealed include:

Oracle AI Security Mesh

Integrates OpenAI’s security models and Oracle’s testing tools to help enterprises check vulnerabilities in applications and infrastructure within authorized scopes.

Oracle App Composer

Allows teams to build enterprise applications using AI tools like Codex, deploy via Oracle infrastructure, and integrate identity, permission, and management controls.

Oracle AI SMB Bundle

A solution tailored for small businesses, combining ChatGPT Business, Codex, Oracle tools, and cloud resources.

Oracle states this bundle offers a free entry point and includes up to $12,000 in annual Oracle cloud credits.

But credits do not equal cash withdrawals, with usage subject to plan terms, qualifications, and limitations.

These services show Oracle’s intent to shift AI from isolated chat tools to everyday enterprise work platforms.

Availability, deployment timing, and pricing remain subject to final product terms.

Why is Oracle also partnering with Google?

On the same day, Oracle announced it will bring Google TPUs to Oracle Cloud Infrastructure (OCI).

TPUs are Google-designed accelerators for AI training and inference.

Previously, enterprises wanting Google TPU access mostly turned to Google Cloud.

Now, OCI will be the first large cloud platform outside Google to offer Google TPUs.

But note:

Official availability is expected in 2027, not immediately on October 27, 2026.

For large AI teams, this means more choices for computing architectures in one cloud environment.

They can select the best hardware for model training, mass inference, or multi-agent systems based on performance, cost, and deployment needs.

Whether this delivers cost savings depends on each enterprise’s models and workloads.

What does Oracle’s dual collaboration with OpenAI and Google signify?

It’s quite interesting.

OpenAI supplies models and AI platforms.

Google provides AI computing technologies like TPUs.

Oracle manages vast enterprise software, databases, cloud infrastructure, and clients.

In AI, these companies may compete in some areas but collaborate in others.

This isn’t about one beating the other.

For enterprises, what matters more is whether AI can integrate with existing data, systems, and permission structures without rebuilding business processes for each model introduced.

But connecting AI to more enterprise systems makes permissions management critically important.

AI querying cost data doesn’t mean it should authorize payments.

AI generating applications doesn’t mean they bypass security testing before formal deployment.

Truly mature enterprise AI doesn’t just let models do more.

It must clearly define:

What data it can read, what it can change, who approves, and how to audit errors.

4. Which of Today’s Three AI Updates Matters Most to the General Public?

If considering near-term direct use, GPT-6.1 Terra’s impact is the clearest.

Eligible ChatGPT Work, Codex users, and API developers have started accessing the new model.

This means more computer operation and data processing tasks could be done at lower model cost.

As for Genie 4, the key focus now is its technical direction.

AI-generated content advancing from fixed videos to interactive simulated worlds could reshape future previsualization, game development, and AI training.

But it’s not yet suitable as a fully open production tool.

Oracle’s updates influence the long-term approach for enterprise AI adoption.

When AI can perform authorized, cross-system tasks, future users might avoid switching between multiple applications for a single job.

However, safety, accuracy, and practical effectiveness remain paramount.

Though different in direction, all three updates indicate a shared shift:

AI is moving from "Can it answer?" to "Can it perform real work safely and cost-effectively?"

5. What Should Readers Remember Today?

When seeing new model launches, first confirm if your product and plan support it—don’t assume every account gains instant access just because of a "formal release."

Impressive AI world simulation videos are research demos, not yet equivalent to commercially deployable production tools.

When AI claims to connect to an entire company’s software, confirm data permissions and approval workflows rather than chasing automation percentages alone.

The future AI gap may lie not only in answer intelligence but in who best integrates model capabilities, tools, data, and human decision-making.

Recommended Reading