Singularity Watch: The Machines Keep Working. Who Checks the Work?

#normallikepeter #churchofnormal #carnivalofnormal #SingularityWatch #AI

From the Singularity Watch desk — AI news for people who live on Main Street, not inside a benchmark chart. Filed September 30, 2026, with sources checked through that date.

The laptop goes to sleep. The work keeps going.

That is a useful picture of where this month’s AI news is heading. The pitch is moving beyond a better answer in a chat window toward systems that can keep a project moving: carry its tools, reach its apps, and bring something back for review.

For the person running a shop in Waseca, that changes the question. You still need to ask whether the answer is right. You also need to ask what happened while you weren’t looking.

The workbench is becoming part of the product

OpenAI’s September 29 release notes describe Codex Cloud environments that reuse a project’s repositories, tools, and dependencies. Tasks receive isolated workspaces and can continue while the user’s computer sleeps. That is a coding-workflow feature, with account and workspace controls; it does not establish that an agent can independently run any business.

The same day’s GPT-6.1 Sol announcement puts the new model in the API, while the release notes describe its rollout in ChatGPT Work and Codex. Ordinary Chat is a separate surface; the announcement says Sol 6.1 is not there yet. Its listed standard API rates are $2 per million input tokens, $0.10 for cached input, and $10 for output. Those are developer rates, not the price of a subscription or a completed job.

The Main Street meaning is continuity. A reusable workbench can reduce the repeated setup between tasks. But the workbench still needs the right files, a bounded assignment, and a person who can recognize whether the returned work meets the assignment.

A cheaper task is a claim worth measuring

Anthropic introduced Claude Sonnet 5.5 on September 28. Its standard input and output prices remain $2 and $10 per million tokens. The company’s claim of up to 30% lower task cost comes from using fewer tokens in its testing, rather than cutting those token prices. It also reports output generation more than 30% faster than Sonnet 5.

Anthropic still describes Opus 5.5 as stronger at complex, open-ended work requiring sustained judgment. That qualification belongs beside the speed claim. These are the company’s findings, not our independent trial of either model.

For a small team, the useful measure is the cost of an accepted result. If a fast first draft needs three rounds of repair, count the repair. If an assistant gets a routine task right and leaves a readable record, count that too. A benchmark can suggest what to try; your actual work decides whether it helped.

More doors into the apps you already use

Google’s September 23 announcement says new Gemini connected apps are beginning to roll out. It names productivity services including Airtable and Linear, and creative services including Adobe, Picsart, Squarespace, and Webflow.

A rollout announcement does not promise every account the same features today. Check the connection and the actions it actually supports.

The direction matters, though. A conversation can become a doorway into the places where work lives. In that setting, access deserves as much attention as eloquence. An assistant that can read a draft has a different responsibility from one that can change a website or send a message.

Orbit still asks for the finished result

A separate engineering story offers a useful reality check. In its September 28 report, updated September 29, NASA says the commercial LINK servicing mission concluded without raising the Swift observatory’s orbit. LINK experienced intermittent communications losses and orientation-control problems. NASA ended its involvement September 3; LINK re-entered the atmosphere September 25.

That is a spacecraft-servicing outcome, not evidence that AI caused a failure. The comparison here is editorial: building and launching a system is an achievement; completing its intended mission is another. Honest reporting can acknowledge both the lessons gained and the objective missed.

The same discipline belongs on the desktop. A file created is not necessarily a document finished. A task that kept running is not necessarily a task that succeeded.

What to ask before handing over the next task

• Name the result. “Return a draft with sources” gives you something concrete to check.
• Set the boundary. Say which files and apps the assistant may use, and which actions need your review.
• Ask for evidence. Open the document, follow the sources, test the link, or inspect the proposed change.
• Count the correction time. Speed matters when the work is useful after review.
• Keep a way back. Preserve the original and a record of what changed.

In August’s Watch, we followed the growing work around safeguards and trustworthy evaluation. September brings that question closer to the ordinary workday. More capability and more continuity make it more useful to ask what the system touched, what it finished, and what remains uncertain.

This is the Watch’s interpretation of the week’s receipts, not proof of a technological singularity or general autonomy. From the Church of NORMAL corner of Main Street, the measure remains human: did the work help someone, and can we account for it?

The wall is vibrating. Keep the receipts. Check the work.

Primary sources are linked beside the reporting above: OpenAI’s September 29 release notes and GPT-6.1 Sol announcement; Anthropic’s September 28 Sonnet 5.5 announcement; Google’s September 23 connected-app announcement; and NASA’s September 28 Swift servicing report, updated September 29. Product, cost, and speed statements are attributed to their publishers. The Main Street implications and the orbit comparison are Singularity Watch commentary.

Nothing is lost. Only recompiled.