Oct 28, 2026 | A cheaper agent model does not automatically make work cheaper

Imagine handing AI a task: read the material in your email, update a spreadsheet, then deliver the result to a colleague.

Previously, it would explain the steps. Now it is starting to open pages, find files and make changes. You may do less copying and pasting, while answering a different set of questions: Is this the latest version? Can this information be shared? Is it ready to submit?

As of October 2026, agents are beginning to work across browsers, files and applications on users’ behalf. Our judgment is that the need for people to perform every step will weaken first. Software value will become more concentrated in data, permissions, shared state and recovery when something goes wrong.

Interfaces will still matter. Screens that teach people how to operate software may increasingly give way to screens that show progress, surface exceptions and let people review results.

1. Which part of the interface is being bypassed?

Our previous software article examined ChatGPT operating a browser and Copilot editing Office files. The common change was handing the work of finding controls and following steps to an agent.

October’s newer material takes this further: the place where a person requests work can be separate from the place where the agent executes it.

Anthropic’s October 21 Claude Duet announcement describes a virtual machine and browser that keep working while the user is away, returning with results or requests for review. Initial access is a research-preview waitlist for US Pro and Max users. Anthropic also warns that it can misread information and misjudge urgency.

The useful distinction is that a message on your phone can initiate work somewhere else. The phone need not be the machine performing it.

The conversation, the execution environment and the business record can become three separate things.

2. Does giving AI a computer make software disappear?

Several products in the source material implement this separation, with different boundaries.

OpenAI’s Dots have a cloud computer and browser, with conversations continuing across ChatGPT, Slack and Teams. Proactive background research is restricted to read-only tools; actions that change or send information follow separate permission and approval rules. Being online continuously does not mean unrestricted access to every application.

Meta’s Muse uses a dedicated virtual machine, with its app or WhatsApp as a communication channel. Meta describes work continuing after the app closes, with approval requested before sending email or making purchases.

Grok Bot’s documentation emphasizes persistent browser, file and terminal state. Multiple Bots belonging to one account share a cloud computer, including files and browser sessions. They should not be treated as separate, fully isolated machines.

These are first-party descriptions of product capabilities. They do not establish how much time ordinary users reliably save.

They do make one point concrete: agents still depend on software to store files, retrieve records and execute transactions. Users may simply stop entering every application themselves.

Giving an agent a computer need not mean buying physical hardware. The environment might be a cloud virtual machine or another authorized host. Continuity, clear permissions and retrievable results matter more to the work than the shape of the machine.

3. Why are people still busy when there are fewer buttons to press?

Delegating operation can turn the user into a supervisor.

Anthropic’s internal employment study offers an instructive example. By August 2026, interactive agent work hours were roughly six times employee hours. Yet approximately 80% of elapsed interactive Claude session time was spent waiting for human input. Meta-work, including managing and monitoring agents, grew from about 10% of interactive sessions in fall 2025 to about 22% in spring 2026.

Six times the hours does not mean six times the useful output. Nor does the waiting figure mean employees were idle for 80% of their workday. These measures describe different aspects of agent activity inside one unusually AI-intensive company.

The implication for interfaces is that human attention can become scarce as more tasks run together.

People need not watch every click. They do need to know what is happening, where work is stuck, which data was used and what the next approval would do.

Task lists, change comparisons, permission boundaries, failure alerts and undo controls may become more valuable than another feature buried in a menu.

Less manual operation still leaves someone responsible for understanding and accepting the result.

4. Does a cheaper agent model make software work cheaper?

OpenAI announced GPT-6.1 Terra on October 26 and describes it as a model for visual reasoning, computer use and everyday writing. Its published standard API rates are $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. OpenAI says those standard rates are 50% lower than GPT-5.6 Terra and GPT-6.1 Sol. Its model release notes also say API usage is billed separately from ChatGPT subscriptions. OpenAI’s Terra API documentation also says computer-use and other tool-specific calls may incur a per-call fee. Requests with more than 272K input tokens use twice the input/cache rates and 1.5 times the output rate for the entire request. Standard per-million-token text prices are therefore not the complete API bill.

A lower rate answers one question: how much each token costs. OpenAI also reports results closer to completed work. At medium reasoning effort on Agents’ Last Exam, it says Terra scored 15 percentage points above GPT-5.6 Terra at about one-third the task cost. This is the company’s reported evaluation, not an independent rerun of Terra. OpenAI’s announcement gives the comparisons and their stated conditions. The Terra system-card addendum does not provide the ALE cost calculation. It says evaluations were run in a research environment or through the API and may differ from production ChatGPT because prompts, tools and reasoning effort can differ. It also warns that comparison values for earlier models may reflect later versions rather than their launch versions.

Task cost is more informative than a token rate, but the setup matters. Agents’ Last Exam says it evaluates a complete agent configuration: not only the model, but also the harness, tools, memory, context management and action loop. Its public methodology describes hidden references and deterministic graders, and reports pass rate, score, cost and runtime. ALE’s methodology explains why these measures should not be reduced to a model name. It helps design a comparison; it does not independently verify OpenAI’s Terra results. The system card does not provide the ALE task sample, harness or cost-calculation details, which limits what outside readers can independently reproduce.

There is still a gap between API rates and the cost of doing real work. For example, if a token rate is cut in half but a task uses twice as many tokens with the same input/output mix, the API bill could return to its starting point. Per-call tool fees, runtime costs, failed attempts, retries, human review and corrections can change the total further. This is an arithmetic illustration of cost accounting, not a measurement of Terra.

Before delegating work to an agent, a team should compare representative real tasks under the same conditions. Record the full system and reasoning setup, the rate of accepted completions, total cost per successful delivery, runtime, retries, and the time people spend checking and correcting results. Include failed attempts; do not compare only the model bill for successful runs.

Our judgment

GPT-6.1 Terra gives the market a new vendor case study in computer-use capability and price. Its published rates and OpenAI’s reported tests point to possible cost reductions, but they do not tell a company how much cheaper its own software work will become.

Interfaces may do less of the step-by-step operation while continuing to help people track progress, set permissions, inspect deliverables and handle exceptions. Whether software is displaced depends both on whether an agent can operate its interface and on whether the complete workflow can deliver repeatable results at an acceptable cost and error rate.

Teams considering a trial can start with one bounded kind of work whose results are easy to verify. Measure model and tool usage, retries and human time per accepted result. Expand delegation only if the complete workflow keeps passing review and total cost falls. An independent Terra rerun under the same setup, together with real workflow cost data, could change this assessment.

Evidence note: Product rates and evaluation figures above are reported by OpenAI. SurviAGI has not independently run Terra or measured production savings. The token example is arithmetic; the recommendations are editorial analysis.