Dots keep working after the user leaves

What happened?

DOCUMENTED: OpenAI launched Dots on 29 September. A Dot runs on GPT-6 Astra, has its own cloud computer and browser, and can connect to more than 4,000 apps. Users can reach it through ChatGPT, Slack and Teams while it handles several projects in the background. OpenAI describes read-only “proactive research”, custom rules, an Activity View and automated review of sensitive actions. Rollout begins on selected plans and markets; Enterprise, Edu and Healthcare require administrator enablement.

Why is it important?

LEARNAI ANALYSIS: Dots shifts the product from question answering towards a persistent operating layer. Once an agent remembers preferences, receives new data and acts across systems, permission design and quality assurance become recurring operational work rather than a one-off setup.

What does it mean for the reader?

PRACTICAL CONSEQUENCE: Start without local-computer access. Give the agent a bounded project document, require approval before external changes, and review its activity log before expanding authority.

Gemini 4 Argon goes to cyber defenders first

What happened?

DOCUMENTED: Google announced Gemini 4 Argon for software engineering, knowledge work and cybersecurity. Google highlights sustained reasoning and output of up to one million tokens in a single run. Access starts through the Fairwind programme for trusted cyber defenders, while broad pricing and availability were not established in the weekly package. Performance and safety results come mainly from Google and its testing partners.

Why is it important?

LEARNAI ANALYSIS: A very large output limit may reduce the need to split complex work, but it also increases the impact of an error repeated through a long run. Phased access also suggests that sensitive frontier capabilities may no longer be released widely on day one.

What does it mean for the reader?

PRACTICAL CONSEQUENCE: Wait for independent evaluations and concrete access terms. Compare error rates, review effort and recoverability of long runs, not only benchmark scores.

Claude Sonnet 5.5 reduces latency and task cost

What happened?

DOCUMENTED: Anthropic launched Claude Sonnet 5.5. The company says it is over 30% faster than Sonnet 5, generally uses fewer tokens and can reduce cost per task by up to 30%. List pricing remains $2 per million input tokens and $10 per million output tokens. These figures are Anthropic’s own measurements.

Why is it important?

LEARNAI ANALYSIS: A faster mid-tier model may matter more operationally than an expensive flagship because it handles a larger volume of everyday work. Savings only materialise if outputs pass internal quality checks without extra retries.

What does it mean for the reader?

PRACTICAL CONSEQUENCE: Run the same representative tasks on Sonnet 5 and 5.5. Measure total tokens, elapsed time, accuracy and human correction rather than relying on list price or a vendor benchmark.