This week in AI: OpenAI's Dots keep working after you log off

Good morning,

OpenAI launched an always-on assistant, Anthropic upgraded its everyday workhorse, and Google unveiled a powerful model most people cannot use yet. The useful question this week is which of these changes you can actually put to work.

OpenAI launches Dots, with a computer of their own

The agent teased before DevDay has a confirmed name: Dots. Each runs on GPT-6 Astra, has its own cloud computer and browser, and remembers ongoing work between conversations. You can reach it through ChatGPT, Slack or Teams.

OpenAI says Dots can connect to more than 4,000 apps, investigate bugs and prepare code changes or launch materials for review.

The first dot is included with Pro or Business Premium, with an allowance for deeper work. Enterprise, Edu and Healthcare get an administrator-enabled beta.

OpenAI's release notes say eligible users get a month of dots usage outside their plan allowances, while the launch article says delegated Codex and ChatGPT Work tasks normally count. Check your account's terms before setting up heavy recurring work.

There is an important regional catch: Pro access excludes the EEA, Switzerland and UK at launch. Business Premium supports all ChatGPT-supported regions, and the rollout is gradual.

Unprompted background research is read-only; action permissions are configurable. Start with one bounded job, such as a weekly customer-feedback summary. Keep sending, spending and publishing behind approval until it earns your trust.

Sonnet and Sol improve without higher token prices

Last week's models already have new competition. Claude Sonnet 5.5, released Monday, targets everyday coding, bug fixes and polished documents. GPT-6.1 Sol, released at DevDay, upgrades OpenAI's coding, computer-use and professional-work model.

Both charge $2 per million input tokens and $10 per million output tokens. Those headline rates are unchanged from their predecessors. The potential saving comes from finishing work with fewer tokens and retries, rather than another list-price cut.

Anthropic says Sonnet generates output more than 30% faster and costs up to 30% less per task in its tests. OpenAI says Sol approaches Astra on several work evaluations at substantially lower cost.

Sonnet is available across Claude and the major cloud platforms. Sol is in the API and rolling out through Codex and ChatGPT Work, starting with Pro and expanding to other eligible paid plans. It is not in ordinary Chat at launch.

Sonnet's migration changes include breaking changes: some older tool configurations now fail, and its reasoning settings change how streamed responses appear.

Gemini 4 Argon arrives behind an access gate

Google announced Gemini 4 Argon for demanding coding, business research and cybersecurity. Initial access is limited to trusted cyber defenders through its Fairwind Program. Wider developer, enterprise and consumer access has no announced date.

Its standout specification is a 1-million-token output limit, up from 64,000.

Google says Argon is helping migrate large codebases to Rust and optimize data-center memory. Major rewrites still undergo automated and human review before production deployment.

Announced introductory API pricing is $2/$10 per million input/output tokens, later rising to $4/$20. Google has not dated the promotion's end.

Google is strengthening safeguards before broad access because Argon can find and patch serious vulnerabilities. Paid API customers and Google AI Ultra subscribers are first in line for the wider rollout.

Quick hits 🗞️

• Gemini Skills replace repeated prompting. Reusable instructions and reference files are rolling out globally for Google AI subscribers aged 18+. Personal-account Gems will start disappearing in November, with automatic migration; Workspace rollout follows in the coming weeks. Source

• Claude Code gets TypeScript mods. Plugins can now rewrite prompts, block tool calls and change the interface in the CLI and desktop app. Mods have Claude Code's machine access and are not isolated, so install only trusted code. Source

• ChatGPT Pro 200 is getting less usage. OpenAI's updated help page says eligible existing subscribers keep their previous allowance through October 29; others already get lower limits. The price stays $200/month. New Pro 500 costs $500/month and adds Astra Ultrafast. Source

• Another Australian agency incident emerges. OpenAI told ABC News an agent accessed non-public NSW fire statistics. The company says reviewed results show no personal information retrieved. Officials believe it happened in June; investigations continue. This is a new disclosure, not a new attack date. Source

See you next week!

If this helped you catch up quickly with AI, forward it to someone who would find it useful. They can subscribe here.