AI Radar for Companies — Monday, August 10, 2026
· CompaniesAutomation
In the first week of August, the major labs have made AI more deployable, not just flashier. Anthropic opened the beta for self-hosted Claude Code: its programming agent runs within your own network, and code, secrets, and artifacts stay home (only the conversation goes to their API for inference). OpenAI made ChatGPT significantly more reliable—68% fewer responses with errors in the Sol model—and is opening unlimited text chats in the free version, with a toggle to choose how much it thinks and how much it costs. Its real-time voice now reads your files and projects, hands-free. And Anthropic scans third-party plugins before they run. The takeaway: the four usual excuses for not using AI—data leaks, hallucinations, cost, and plugin risks—are falling one by one.
Yesterday's radar was about owning your AI; this weekend the question gets grounded: what was still holding you back from truly using it? In the first week of August, the big labs haven't made AI flashier, but rather more deployable: Anthropic now lets you run its coding AI within your own network—code and secrets don't leave your premises—and scans third-party plugins before they run; OpenAI has made its ChatGPT significantly more reliable—68% fewer responses with at least one error—and is opening its free version to unlimited text chats, with a control to choose how much it "thinks" and how much it costs. The weekend takeaway: the four usual excuses for not implementing AI—that your data leaves the building, that it makes things up, that it's expensive, and that plugins are a risk—are falling one by one.
You can now run coding AI within your own network: your code stays home
On August 6, Anthropic opened the public beta of self-hosted environments for Claude Code: your programming agent's sessions run within your organization's network instead of Anthropic's infrastructure. Repository checkouts, build artifacts, secrets, and any files the session creates or modifies stay on your machines; only the conversation—prompts, responses, and tool outputs—goes to the Anthropic API for inference (via outbound HTTPS; they never enter your network). It's available on Team and Enterprise plans and is disabled by default. For your company: if data residency or compliance were stopping you from using AI in development, this is the answer—code, secrets, and artifacts stay home—but it's not total isolation, as conversation text still goes out, so watch what you write in the prompt and don't paste secrets there. It's the counterpoint to Meta's Muse Code from yesterday: where one makes it cheaper, the other lets you own the execution. Source
Cheap AI now makes far fewer mistakes—and you decide how much it thinks (and costs)
OpenAI updated ChatGPT on August 6: in its Sol model, responses with at least one factual error are 68% less frequent than with the previous GPT-5.5 Instant (62% less in the free model, Luna), and it has added a control to choose how much the model "thinks" for each response. Additionally, free users will have Luna by default and unlimited text chats starting the week of August 10, with anti-abuse protections (this "unlimited" is text-only: files and images keep their limits). For your company: the cheap or free tier is now reliable enough for real text tasks—drafts, replies, summaries, translations; use the low setting for routine tasks and turn it up only for difficult ones to control cost and response time. Fewer errors doesn't mean zero errors: keep reviewing anything that goes out to the client. Source
Voice is no longer a toy: hands-free over your own files and processes
On August 7, OpenAI activated its real-time voice (GPT-Live) as the default voice experience in Enterprise, Edu, and Healthcare spaces—admins just need to activate voice, early access is no longer required—and added two useful features: you can upload files within a voice conversation to analyze them and use voice within a Project, with its chats, sources, and instructions at hand. Meanwhile, Anthropic brought its voice mode to Claude Opus, Sonnet, and Haiku, with access to tools like Gmail and Slack. For your company: voice is finally useful for working without a keyboard—sales reps on the road, warehouse, workshop, front desk; set up a Project with your procedures or price list and let the team consult it by speaking. Start with a specific case (checking a procedure or order status) and measure it before scaling up. Source
Before installing an AI "skill" or plugin, have someone scan it
On August 6, Anthropic added a security scan for skills and plugins in beta for Enterprise plans: when someone uploads or edits a third-party add-on, Claude reviews its content for malicious behavior and returns a verdict—pass, warn, or fail—before it runs; it works in Claude, Claude Cowork, and enterprise plugin marketplaces. For your company: AI add-ons—skills, plugins, connectors—are third-party code that runs with your permissions, the same vector as this week's scares (Langflow, agent connectors); treat them like any other dependency: whitelist what's allowed, define who can install, and run a scan before activating. If your tool doesn't scan automatically yet, manually review the source and the permissions each add-on requests. Source
What to watch this week?
First, whether OpenAI delivers on unlimited free text chats from the 10th and how the anti-abuse protections hold up; next, more "host AI at home" movements and security in the AI supply chain (skills, plugins, connectors), and how the AI Act lands in practice now that AESIA can already inspect and fine. And a question for Monday: of the four excuses not to use AI—it leaves the building, it's wrong, it's expensive, add-ons are a risk—which one did you just run out of?