Two models, one agentic AI revolution

Two models in the spotlight: Anthropic releases Sonnet 5, purpose-built for AI agents, while DeepReinforce unveils Ornith-1.0 — an open-source competitor that runs locally without an API. Plus, identity verification at Anthropic and the first global UN report on the state of AI.

Intro · Alex

Midweek brings a true dualism in the world of AI models. On one hand, Anthropic is betting on agency — Sonnet 5 is the first mid-tier model fully designed for autonomous planning, tool use, and browsing. On the other, DeepReinforce shows that an open, local alternative is already within reach — and under an MIT license. This isn't the future — it's happening now. In today's edition, I've traced five key stories for you: the new Sonnet in action, Ornith on local hardware, GitHub Copilot integration, Anthropic's groundbreaking (and controversial) identity verification, and the first scientific UN report on AI. Plus two security stories worth knowing.

What's worth knowing

Anthropic releases Claude Sonnet 5 — an agentic model fit for a flagship

The new Sonnet 5 achieves performance close to flagship Opus 4.8, but at a significantly lower price — $2/M input tokens and $10/M output tokens (promotional pricing through August 31). This is the first Sonnet model fully designed for AI agents: it autonomously plans, uses tools, a browser, and a terminal. It has become the default model on Free and Pro plans at claude.ai.

DeepReinforce publishes Ornith-1.0 — an open-source model for agentic coding

The new AI lab has released the first open-weight model family (MIT license) designed for agentic coding. Ornith builds its own 'agent scaffolding' during training — hence the term 'self-scaffolding.' The 9B variant weighs 5.6 GB and runs locally via `ollama run ornith`, making it a viable, free alternative to commercial models.

Claude Sonnet 5 lands in GitHub Copilot — another model in the multi-model hub

GitHub announced general availability of Sonnet 5 in Copilot across all plans — from Pro to Enterprise. Developers can now select Anthropic's model directly in VS Code, JetBrains, and GitHub CLI. Organization admins can enable it via model policies, giving teams flexibility in choosing their AI engine for daily work.

Anthropic introduces mandatory identity verification — starting July 8

Starting July 8, users on Free, Pro, and Max plans will need to upload a scan of their ID to use Claude. The verification is tied to plans for restoring access to the most advanced models — Fable 5 and Mythos 5 — which were suspended in June at the request of the US Department of Commerce. This marks the first instance of KYC (Know Your Customer) for ordinary LLM users.

From the tech world

BioShocking — just tell the AI that 2+2=5 to bypass safety measures

A new attack on AI browsers involves luring the model into an alternate reality — simply suggesting that 2+2=5 is enough to disable guardrails. Dubbed BioShocking, the attack demonstrates how fragile the protection mechanisms of AI agents are when based solely on prompts.

Your smart TV may be secretly scraping data to train AI

Analysts have discovered that the Bright Data SDK embedded in smart TV apps, routers, and media players uses home devices for mass data scraping. Traffic from a home router has the highest reputation, so AI bots collect data with impunity — at the cost of your electricity and bandwidth.

Tip of the day

How to run an AI model locally in 5 minutes — the Ornith recipe

Ornith-1.0 shows that local AI models are no longer a futuristic curiosity — they're a real working tool. Most modern open-weight models run through Ollama, which reduces the whole process to a single command. You don't need an API key, a subscription, or a cloud connection. Install Ollama (ollama.com), then in your terminal type `ollama run ornith:9b`. After downloading (5.6 GB), the model runs fully offline. You can use it via REST API at http://localhost:11434 or integrate it with VS Code through the Continue.dev extension. It's the fastest way to have your own coding agent without dependency on cloud providers.