Updated September 30, 2026. A selection of developments announced September 23–29.
An assistant that keeps working after you close your laptop needs more than a capable model. It needs somewhere to run, a way to remember the assignment and clear limits on what it can do. The latest announcements from OpenAI, Anthropic and Google are filling in those pieces.
Our reading of the week’s news is that the competition is widening: model quality still matters, but so do the cost of repeated work, continuity between sessions and the ability to inspect an assistant’s actions. Here are four developments that make that shift concrete.
1. OpenAI’s dots turn a conversation into ongoing work
OpenAI’s latest product update introduces dots, persistent agents powered by GPT-6 Astra. Each has a cloud computer and browser and can continue researching, preparing documents or building software while your own device is off. OpenAI describes a system that carries work between conversations and returns to the user when a decision needs their input. Read OpenAI’s dots documentation.
Access is still rolling out. As of September 30, the documentation lists eligible Pro users over 18 outside the European Economic Area, the UK and Switzerland, alongside worldwide rollouts for Business Premium and Enterprise. Enterprise access must be enabled by an administrator. Those distinctions matter: an announcement does not mean every subscriber can use the product today. Check the rollout and access details.
The practical change is the assignment. A hypothetical sales team could ask an agent to keep a proposal aligned with changing customer requirements, then review the revisions each morning. That introduces a continuing responsibility, with questions about which sources count, when to interrupt someone and which changes need approval.
2. Claude Sonnet 5.5 makes efficiency the headline
Anthropic released Claude Sonnet 5.5 on September 28. The company says it generates output more than 30% faster than Sonnet 5 and costs up to 30% less per task in its testing. The distinction is important: the standard token prices remain $2 per million input tokens and $10 per million output tokens. The claimed saving comes from completing work with fewer tokens. Read Anthropic’s launch announcement.
Anthropic positions Sonnet for well-defined everyday work, including bug fixes, documents and slides, while retaining Opus 5.5 for more complex assignments requiring sustained judgment. These are vendor claims and test results; the launch material does not establish the same savings for every workflow.
The release also adds cybersecurity safeguards, including falling back to Sonnet 5 for higher-risk cybersecurity tasks. Anthropic says routine software development remains available and acknowledges that evaluations cannot detect every possible failure. See the release’s safety section.
For a team processing hundreds of similar requests, the useful metric is cost per accepted result. A faster draft helps most when it also reduces the time someone spends correcting it.
3. OpenAI is packaging the infrastructure around the model
OpenAI’s update for the week of September 28 adds GPT-6.1 Sol, which the company describes as offering performance close to Astra at a lower cost. It also introduces reusable Codex Cloud environments: developers can prepare dependencies and tools once, publish that environment, and give each new task an isolated workspace based on the prepared filesystem. Read the product update.
The same update lists Codex Security Cloud in research preview, with repository scanning and monitoring of new commits. Availability depends on account and workspace access, and findings and proposed patches still need review. See the preview details.
Our assessment is that this infrastructure may be as consequential as a model upgrade. If a coding assistant spends less time rebuilding its setup, teams can repeat useful work more easily. An isolated environment is also a clearer place to inspect what happened when a task fails.
4. Google is tackling the memory behind persistent assistants
On September 23, Google DeepMind described an extension to Private AI Compute designed to support persistent memory across devices. The proposed architecture combines encrypted storage with protected cloud processing, while the keys needed to unlock the stored information remain on the user’s devices. Read Google’s technical announcement.
Google also describes a public record of server software and verification before a device sends personal data. This is a technical architecture update about what the platform will enable; it should not be read as confirmation that every Gemini user already has a new memory feature.
Continuity could make an assistant easier to use: people would spend less time restating preferences and reconstructing an unfinished project. It also makes memory controls more important. A useful product needs to let people understand what was retained, correct it and decide which contexts should stay separate. Encryption addresses part of that problem; product design has to address the rest.
What to watch next
Taken together, these announcements point toward assistants that are cheaper to run repeatedly and better equipped to continue an assignment over time. That is our interpretation of the releases, rather than a claim that dependable autonomy has been achieved.
The next useful evidence will come from ordinary use: how often an agent finishes a task correctly, how much review it requires and whether people can understand and stop its work. For anyone testing the new tools, one recurring task with a clear definition of success will reveal more than a long list of impressive demos.
