Six AI releases from Sep. 28–Oct. 4 that let you hand over the whole job

Six releases from this week are worth a look, covering September 28 to October 4, 2026. Two of them give an assistant somewhere to keep working after you stop watching: OpenAI's dots stay on a job between conversations, and Codex Cloud carries a coding task past the moment you close the laptop. A third reaches the desktop software that never had an API to automate. The last three are this week's groundwork — a cheaper model for everyday tasks, a voice model you direct from the script, and the frontier model Google is building for its longest jobs.

What changed. OpenAI began rolling out dots on September 29, 2026: always-on agents inside ChatGPT that take on continuing work and keep making progress between conversations. Each dot runs on its own cloud computer, connects the apps and messaging channels you choose, and forms memories from them. You give it a goal, define what it may do alone, and it brings results back for your review. Access reaches eligible Pro and Business Premium users first in supported markets; at launch, Pro access excludes the European Economic Area, Switzerland, and the United Kingdom, and Enterprise access is a beta that starts switched off. Your first dot is included with Pro or Business Premium, and for the first month its usage does not draw on your plan allowance. 12

Workflow win. A recurring job stops needing you to start it. The dot can run scheduled checks, research in the background, and review connected apps without a new message from you, then hand back what it found. Connecting a computer is optional and begins off; switched on from the desktop app, the dot can also open Work or Codex tasks, use local skills, and drive the local browser when its cloud browser is blocked. 2

Who should try it. People with a repetitive, checkable task — a weekly report, a research watch, a calendar or inbox sweep — and the patience to write the rules before it starts. The fit improves as the output gets easier to verify.

First test. Give one dot a single recurring job you already do by hand, and write its rules before you let it run: what it may do without asking, what it must ask about, and what it hands back to you. After a week, read the activity list under In progress, Scheduled, and Completed, and widen the rules only for the actions that came back clean.

The catch. Rollout is gradual and geographic, so access can take days and may not reach your account at all. The dot's memory and connected-app data belong to it: deleting the dot deletes its conversations, memories, and scheduled tasks, and disconnecting an app leaves what it already saved in place. OpenAI says a dot can make mistakes even while following your rules, so the reviewing stays with you. 2

What changed. The same September 29 release added Codex Cloud: coding tasks you start and continue from desktop, web, or mobile, running in reusable environments built from your project's repositories, tools, and dependencies. Each task gets its own isolated workspace, and OpenAI says a task can continue while your computer is asleep. In managed workspaces, the workspace's Codex cloud-access settings apply. 17

Workflow win. Setup stops repeating for every task, and your hardware stops being the limit. Build the environment once and each new task starts from the same working state rather than a fresh clone and install; start from mobile and the task runs while you are away from the desk. The step this removes is the one where a laptop stays open purely to keep a build alive. 1

Who should try it. Developers with a repository whose setup they trust, and the kind of task — a failing test, a small migration, a review pass — that can be described in a message and checked afterward.

First test. Prepare one cloud environment for a repository you know well, then launch a small, bounded task from your phone and close the laptop. When you open it again, review the change and count how much of it you keep as it stands.

The catch. The environment has to be configured with its repositories, tools, and dependencies before it is useful, so the first setup is real work. In managed workspaces, access runs through administrator settings, and availability follows your plan. 17

What changed. On October 1, 2026, GitHub put computer use into public preview in Copilot CLI and the Copilot app on macOS and Windows. Copilot can now read accessible app content and visual context, click controls, enter and edit text, press keys, scroll, drag, and move through a workflow across applications. The target is software with no API, command line, or MCP integration — the legacy and GUI-only tools where the work still has to be done by hand. Copilot asks for approval before controlling an app, keeps a list of apps you can review or reset, and stands down when organization-managed settings switch it off. On macOS it walks you through the Accessibility and Screen Recording permissions it needs. 3

Workflow win. Automation reaches the last mile it usually stops short of. Renaming a folder or pulling a report is scriptable; the same information trapped inside a desktop-only claims system is where people still click by hand. Describing the outcome, the applications involved, and the constraints lets Copilot move through those screens, and the approval prompt keeps the decision about what it may touch with you. 3

Who should try it. Developers and operations teams who spend real time in software that has no command line — expense, claims, inventory, and other line-of-business tools — and who can check the result once it is done.

First test. Enable computer use for one low-risk GUI workflow and describe the outcome rather than the steps. Watch the first run end to end, and note every approval it asks for and every point where it needed your guidance; that list becomes the policy you write.

The catch. It is a public preview on macOS and Windows only, and the applications it drives have to be installed on that machine. The macOS permissions are a real setup step, an organization can disable the feature, and an agent working a graphical interface is only as reliable as the app's own accessibility data. The checking stays with you. 3

What changed. Anthropic released Claude Sonnet 5.5 on September 28, 2026, the second model in the Claude 5.5 family. It keeps Sonnet 5's list price — $2 per million input tokens, $10 per million output tokens, and $0.20 for cache reads — while Anthropic says it needs far fewer tokens for the same job, working out up to 30% cheaper per task, and generates output more than 30% faster. On Terminal-Bench 4.0, an agentic coding evaluation, it scores 70.6% against Sonnet 5's 10.3%. Anthropic positions it for well-scoped everyday work: bug fixes and polished documents, slides, and spreadsheets, with Opus 5.5 still the stronger choice for open-ended, long-judgment work. It is the first Sonnet model to ship with the cybersecurity safeguards Anthropic reserves for its most capable models, and Haiku 5.5 is due in the coming weeks. GitHub added it to Copilot the same day. 48

Workflow win. Routine work stops competing with frontier work for the same budget. Where cache reads and output tokens dominate a bill, a model that finishes the same task in fewer tokens at a lower price changes which jobs you can run all day instead of once. Anthropic's early testers describe fewer steps and less re-prompting: Slack reported about 14% fewer output tokens with unchanged prompts, and Zendesk measured ticket processing 20% faster. 4

Who should try it. Teams with a high volume of well-scoped work: support triage, code review, document and slide production, data extraction. It suits tasks you can specify precisely and check quickly.

First test. Take one routine task you run often and run it on Sonnet 5.5 and on your current model. Record the tokens used, the minutes on the clock, and the number of corrections before the result is usable. The cheaper run wins only when the corrections stay flat or fall. 4

The catch. The performance and cost figures come from Anthropic's own harness, and the customer numbers from companies it chose to quote. Effort level moves the trade: Medium is the default inside Claude's apps and High on the platform, so a cost comparison only holds at the same setting. Anthropic itself says Opus 5.5 remains clearly stronger for complex, open-ended work. 48

What changed. ElevenLabs launched Eleven v4 and its low-latency variant, Eleven v4 Turbo, on September 28, 2026. The pair is built on a new architecture that reads tone, pacing, emotion, and context from the text and follows directions written into the script — [laughs], [said angrily in French accent], [light rain], [phone buzzing] — with better support for custom pronunciations through IPA. Multi-speaker dialogue now tracks the conversation rather than assembling separate lines. Both models cover more than 90 languages, and a voice recorded in one language speaks the others while keeping its identity and taking on a native accent. Instant Voice Clones capture a voice from 10 seconds of audio, and Professional Voice Clones are supported for higher fidelity. Eleven v4 Turbo reaches a median ~100 ms inference latency, and ElevenLabs measures ~150 ms median time to first speech. Both ship in ElevenAgents, ElevenCreative, and the API. 5

Workflow win. The recording session comes out of the schedule, and so does the retake. A narrated explainer, an audiobook chapter, a dubbed video, or a character performance can be written and directed inside the script, then adjusted by editing a line instead of booking a session. A brand voice chosen once carries across languages and across a long production without drifting. 5

Who should try it. Content teams producing narration, dubbing, and character voice, and anyone building a conversational agent where a flat, monotone voice loses the caller.

First test. Clone a voice from a 10-second sample or design one from a description, then direct the same three lines two different ways and compare how much of the script you had to change to get the delivery. Carry the better voice through a two-minute read and listen for drift across the passage.

The catch. The expressiveness and preference claims are ElevenLabs' own, drawn from an Artificial Analysis ranking and the company's own blind tests against named competitors, and the latency figures were measured by ElevenLabs. Turn them into your own listening test on your own script before you build a pipeline on them. 5

What changed. Google announced Gemini 4 Argon on September 30, 2026, a frontier model built for long-horizon work. It raises the output limit to one million tokens, up from 64,000, so a single trajectory can run much longer before it stops; Google reports 77.9% on DeepSWE v1.1 and first place on the Vals Index, which weights finance, coding, legal, and tax work by their share of U.S. GDP. Argon is rolling out first to trusted cyber defenders through Google's Fairwind Program, and Google says developers, enterprises, and consumers will follow, starting with paid API customers and Google AI Ultra subscribers. Introductory pricing is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 afterward. 6

Workflow win. A long job can be given in one request. The tasks Google describes — optimizing a quantum subroutine, hunting memory waste across a data center, migrating an 800,000-line C++ codebase to Rust — are the ones usually split across many sessions because the model runs out of room. The larger output window is the mechanism; whether it pays off for you depends on how long your longest task actually is. 6

Who should try it. Paid API customers and Google AI Ultra subscribers with long-horizon coding, research, or review work, once access opens. It is the natural trial for anyone whose current model has to be re-prompted midway through a large task.

First test. When access arrives, take one task you normally split across several prompts and run it end to end in a single trajectory. Compare the result, the corrections, and the time against your current route.

The catch. Argon is not generally available yet: Google is expanding access in phases, and the timeline for a wider release is its own to set. The benchmark figures are Google's own, and several come from internal evaluations. The highest-cyber-capability configuration ships without cyber guardrails only to trusted defenders, which tells you the rollout is being staged around risk rather than shipped complete. 6

Pick one task you repeat every week whose result you can check, and run it both ways on five to ten real instances: once the way you do it now, once through whichever release above fits it.

Decide the boundary before you start. Name what the tool may do on its own and what comes back to you, then verify the first few results by hand before you trust the pattern.

Keep five numbers:

A release earns a permanent place in your week when the result comes back correct and the checking stays short. When the saving arrives as a longer review, keep the route you already have.