Week of October 13-19, 2026
Three things moved in AI this week worth knowing about if you run a service business, back office, or field operation. Here’s what changed and what to do about it.
An AI voice agent just posted real numbers on handling phone calls without a human
On October 14, xAI launched Grok Voice 3.0, a speech-to-speech model built for live voice agents. It scores 86.9 on Artificial Analysis’s Speech-to-Speech Index, up from 81.3 for the previous version, and 97% on Full Duplex Bench, the benchmark that measures how well a model handles pauses, interruptions, and talking over someone. xAI says its own support line for Starlink, which handles roughly 20,000 calls a day, now resolves about 70% of those calls without a person stepping in, covering account verification, hardware troubleshooting, credits, and replacement equipment. Pricing didn’t change: the Voice Agent API still runs $0.08 per minute, about $4.80 an hour of call time.
This matters because the biggest complaint about AI phone answering, the stilted, walkie-talkie feel that makes callers hang up or demand a human, is specifically what these benchmarks are built to measure. A model that can handle interruptions and still sound natural is a different product than what was available even six months ago.
If you tried an AI answering service a year or two ago and it felt robotic enough to drop, it's worth a second look now. The underlying voice models have moved fast, and a bad first impression from an older version shouldn't be the reason you're still missing after-hours calls today.
AI phone agents are now benchmarked on handling real conversational mess, not just reading a script. If call volume or missed calls are a pain point, re-evaluate vendors against this year's models, not last year's demo.
Google's AI document checker, built for research papers, now reads contracts too
On October 14, Google Research opened Deep Review to the public, giving every user two free document reviews a week. Deep Review started as a pre-submission checker for academic papers, splitting a document into sections, running parallel AI agents over each one, and flagging logical gaps, unsupported claims, and inconsistencies. Google says it has now extended that same approach to design documents, presentations, and legal and financial documents, checking for consistency across clauses and data tables and flagging ambiguous terms. Input is currently limited to PDF and plain text, with multi-file support coming soon, and the tool is free to try at research.google/deepreview.
Google is upfront that it’s assistance, not a replacement for a lawyer or accountant. The system can flag a valid clause as a problem, so it still needs a human sign-off, but it’s a real second pass at a document before it goes out or gets signed.
Before you send out a subcontractor agreement, a bid package, or a lease, running it through a free tool that checks for inconsistent terms and unsupported numbers costs nothing and takes a few minutes. It doesn't replace your attorney, but it can catch what a tired read-through misses.
Google's strongest reasoning model reached regular Gemini subscribers, with a mode that works in the background
On October 13, Google widened access to Gemini 4 Argon, its flagship reasoning model, beyond the limited cybersecurity-partner rollout it had in September. It’s now available through the API and inside the Gemini app and AI Mode in Search for Pro and Ultra subscribers, with a 1,048,576-token context window and support for code execution and function calling. Alongside the model, Google added a dedicated Tasks mode that lets Argon run jobs independently in the background through connected Google services and third-party apps, rather than waiting on a prompt-and-response loop.
API pricing starts at $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 on December 1. Consumer access is limited to paid Pro and Ultra tiers, not the free Gemini app.
If your business already pays for Google Workspace and a Gemini subscription, you likely just got a materially stronger model for research, drafting, and multi-step analysis without changing anything about your setup. It's worth checking which tier your account is on before assuming you're still using the older model.
If you're already paying for Gemini Pro or Ultra, check that you're actually using Argon for complex tasks like contract comparisons or multi-step scheduling. The upgrade is included, but it's opt-in to notice.
What we're watching next week
-
OpenAI’s API model retirements. Multiple reports this week pointed to older GPT-4 family models (GPT-4o, GPT-4o mini, GPT-4.1, GPT-4.1 mini) being retired from the API, but sources disagreed on the exact date. If you or a vendor built an automation on one of these model IDs, check OpenAI’s own deprecations page or the
shutdown_datefield in its models API directly rather than relying on secondhand coverage. -
Anthropic’s IPO process. Anthropic reportedly posted $16.8 billion in quarterly revenue ahead of a planned IPO, while New York City’s Comptroller publicly called for a pause over AI safety concerns. Worth tracking as a signal of how fast the major AI vendors are consolidating and scaling, which affects pricing and reliability for everyone building on top of them.
-
Small business AI adoption abroad. A Meta-commissioned survey of nearly 4,000 small business owners across Asia-Pacific markets found 84% already using AI and 86% saying they still need more guidance to get full value from it. It’s not a U.S. number, but it’s a preview of where domestic adoption surveys are likely headed.