The end of September brought several new AI models, while early October added tools for voice applications and data privacy. Here are ten developments from September 28–October 4, 2026, with an emphasis on what changed, who can use it and which limitations matter. Topics already covered on RozumTech have links to their full explainers.
1. Claude Sonnet 5.5: speed and cost per completed task
On September 28, Anthropic introduced Sonnet 5.5 for everyday work, documents and coding. The company reports faster output and lower costs for some tasks through more economical token use. That is different from cutting the price of each token. When comparing models for a workflow, consider the cost of a completed task, including corrections. Our Sonnet 5.5 explainer covers that distinction.
2. OpenAI Dots: assistants for ongoing work
On September 29, OpenAI announced Dots, agents with their own cloud computer and access to connected applications. The initial rollout covers Pro, Business Premium and Enterprise plans in eligible markets. The practical question is what an assistant may do independently and when approval is needed. Our Dots article explains background work, app permissions and action review.
3. GPT-6.1 Sol: a new capability and price balance
GPT-6.1 Sol arrived the same day. OpenAI positions it for coding, document work and tasks involving multiple steps, with standard prices of $2 per million input tokens and $10 per million output tokens. Claims about approaching Astra should be read in the context of particular evaluations, rather than as a promise of equal results everywhere. See our DevDay 2026 overview for the wider set of announcements.
4. Gemini 4 Argon: cybersecurity partners come first
Google introduced Gemini 4 Argon on September 30. The Fairwind program offers early access to selected cybersecurity partners. For readers, the distinction between a model announcement and being able to select it in an everyday app matters. Our separate article explains Argon access and announced pricing.
5. Windows 11 26H2: a gradual rollout
On September 29, Microsoft began rolling out the Windows 11 2026 Update. Eligible PCs on 24H2 or 25H2 can move through a small enablement package. Availability expands gradually, and known compatibility problems can trigger safeguard holds. A missing installation button therefore does not necessarily indicate a broken system. Our 26H2 availability guide explains where to check and why bypassing a hold deserves caution.
6. Microsoft updates speech recognition and generation
On October 1, Microsoft announced MAI-Transcribe-2-Streaming and two MAI-Voice-2.1 models. The former converts live audio to text; the latter turn text into speech. These are different components of a voice service, with separate billing and language support. Our audio model explainer distinguishes transcription, speech generation and a finished consumer application.
7. Claude Code gets TypeScript mods
On October 1, Anthropic introduced mods for Claude Code. Developers can change the interface and tool behavior, intercept events and distribute customizations through plugins. This is relevant to teams with their own development workflows. Because mods can affect an agent’s actions, the origin of the code matters; managed environments also have administrative controls over the customizations they load.
8. Cloudflare OHTTP Gateway: privacy built into applications
On October 2, Cloudflare opened a closed beta for OHTTP Gateway. The technology separates knowledge of a client’s network address from the contents of its request across independent parties. Developers must integrate this infrastructure into the interaction between an app and its server. It does not automatically anonymize ordinary browser traffic. Separate operators for the relay and gateway are an essential condition of the privacy model.
9. Microsoft Digital Defense Report: access still matters
Microsoft published its Digital Defense Report 2026 on October 1. One finding is that AI accelerates parts of attacks while accounts, permissions and exposed systems remain important risk points. Agents add questions about access to tools and data. A useful step for a small team is to review which services its AI assistants can access and whether they need all the permissions already granted.
10. The AI safety accord: voluntary commitments
At a September 29 White House meeting, technology leaders supported an accord on oversight and review of advanced AI systems. ABC News reported that the commitments are voluntary; the agreement is not a new law. Its significance should be judged through subsequent evaluations and published results. Our article on the accord explains the context and the limits of the agreement.
What to watch this week
Windows users can check the update offered to their particular PC. Anyone choosing AI for work should compare access and the cost of their intended workflow, alongside model capabilities. Voice product builders need to check input and output languages separately. Catch up on earlier developments in our September 21–27 digest.
Join the conversation
Stay on topic and respect other readers. Your first comment may appear after editorial review.