This Week in AI: September 28 – October 4, 2026
OpenAI gave its AI a computer of its own. Google launched its first Gemini 4. And the gap between "robot on the phone" and "sounds like a person" got smaller again.
OpenAI's DevDay last week was the headline event — but the real story of the week was quieter. Almost every big lab shipped something that makes AI cheaper to run, faster to talk to, or more useful when you're not watching. Here's what happened, and what it means if you run a business.
OpenAI's "Dots": an agent that keeps working after you close your laptop
On September 29, at DevDay in San Francisco, OpenAI introduced Dots — always-on AI agents that run on their own cloud computer, with their own browser and connections to your apps. You hand a dot a goal, close your laptop, and it keeps going: digging through documents, comparing suppliers, drafting follow-ups.
The details that matter:
Powered by GPT-6 Astra, working across ChatGPT, Slack, and Microsoft Teams.
Permissions and approval rules keep you in charge — the agent asks before it does anything sensitive, rather than freelancing with your accounts.
First rollout goes to Pro and Business Premium users, with enterprise pilots underway.
DevDay also brought GPT-6.1 Sol, which OpenAI says matches Astra-level coding ability at roughly a fifth of the token cost, plus a new $500/month Pro plan and Codex moving to the cloud.
If you run a small business, the pattern is worth noticing: the AI is shifting from "answer my question" to "finish this job while I'm busy." Think: chasing down quotes, monitoring competitor prices, keeping the follow-up list warm. The guardrails — approvals before actions — are the part to actually read before turning one loose on your business data.
Google's first Gemini 4 goes to the security teams first
On September 30, Google unveiled Gemini 4 Argon, the first model of a new generation built for long-horizon reasoning — software engineering, enterprise knowledge work, and cybersecurity. The headline spec: up to 1 million output tokens, meaning it can produce enormously long, coherent work product in one go. On Google's published benchmarks, it leads outright in legal reasoning tests (Harvey's Legal Agent Benchmark at 19.6%, versus 5.4% for GPT-6 Astra and 3.8% for Claude Opus 5.5 — roughly a fourfold lead).
Interesting tell: Google is rolling Argon out first to trusted testers and cyber defenders, not to everyone. The model can autonomously find, validate, and patch software vulnerabilities — genuinely useful for defence, genuinely risky in the wrong hands. Broader access for developers, enterprises, and AI Ultra subscribers comes later.
For a small business, the immediate takeaway isn't "buy Argon." It's that the tools you already pay for — document review, contract drafting, code and data work — are about to get sharper at long, complex jobs. The frontier moved this week.
Voice AI week: your phone assistant just got harder to tell apart from a person
September 28 — ElevenLabs launched Eleven v4 and v4 Turbo. The Turbo variant is built for live conversation: median inference latency around 100 milliseconds, and about 150 milliseconds from request to audible speech. For comparison, an average pause between two people talking is longer than that. Eleven v4 took the top spot on the Artificial Analysis speech leaderboard, supports 90+ languages (up from 70), and can clone a voice from roughly 10 seconds of audio. Launch pricing is 72% off list until October 12.
October 1 — Microsoft shipped three voice models of its own through Microsoft Foundry: a streaming transcription model covering 60 languages, a text-to-speech model in 23 languages that Microsoft calls its most expressive yet, and a flash variant built for high-volume, low-latency applications like contact centres.
The practical read, if you're considering an AI receptionist: voice quality and speed are becoming commodities. Three serious vendors upgraded their engines in the same week. The question to ask a vendor is no longer "does it sound human?" — it's "which engine are you running, and what does it cost you per call?"
The money angle: AI that sells and serves in one conversation
Two launches pointed at the same problem: businesses are running one AI to sell and a different AI to support, and neither knows what the other is doing.
Decagon (October 1) announced agents built for "the era of personal agents," including a partnership with Plaid so customers can resolve payments, claims, and disputes inside the conversation — no being sent off to log into a bank site mid-chat, which is where conversations usually die.
Gladly launched "agentic commerce": one AI that both sells and serves, replacing the separate shopping assistants and support bots that hand customers off to each other.
The numbers behind this push are hard to ignore. Adobe Analytics found that AI-driven traffic to U.S. retail sites was up 393% year over year in the first quarter of 2026, converted 42% better than non-AI traffic, and generated 37% more revenue per visit. A separate Adobe survey found 39% of U.S. consumers now shop with AI, and 85% of them say it improved the experience.
For retailers and service businesses alike, the message is plain: your next customer may arrive with an AI helper already open, one question left to answer. If your website, chat, or phone bot can't answer it, the sale walks.
The bottom line for your business
Three questions worth asking before year-end planning:
Can your AI keep working when you're not watching? Dots-style agents make long-running follow-up realistic. If your business leaks leads through slow follow-up, that's the gap to close.
What does an AI task cost you now? Cheaper models (GPT-6.1 Sol, Sonnet 5.5) and faster voice (v4 Turbo, Microsoft's trio) mean your vendors' costs are falling. Ask what you're getting for it.
Does your customer-facing AI sell AND serve, or just answer? The two-bot era is ending. The winning setup is one conversation that carries a customer from question to purchase to support.
Nobody needs all of this on Monday morning. But if you're shopping for an AI receptionist or any customer-facing tool this fall, the bar moved this week — and it's worth knowing where.
Sources: NoloWiz · sneos · BitsMinds · EveryDev.ai · PYMNTS · Digital Applied · Gladly PRNewswire · Decagon


Comments