Anthropic reviewed 141,006 evaluation runs and found six runs across three incidents where Claude reached real systems. The third-party test environment had live internet access even though Claude had been told that it did not.

The models did not discover zero-days. They used basic security weaknesses. One accessed a database containing hundreds of production records. Another published a malicious PyPI package that was downloaded and executed on 15 systems.

And yes, two of the affected companies did not know any of this had happened until Anthropic contacted them. That is the part I still cannot get over.

This was mainly a containment failure, not Claude independently deciding to escape. But the distinction does not make the actual damage disappear.

Source: Anthropic

Dutch insurer Univé says 97% of its ChatGPT Enterprise licences have been activated, 85% are used weekly, and employees have created around 1,500 custom GPTs.

Its agents now prepare insurance claims and underwriting queues, but trained employees still make the final decisions. Employees also built many of the tools themselves instead of waiting for one central AI team to decide how every department should work.

This is how you build an AI workforce without starting with a plan to replace everyone. The numbers come from an OpenAI customer story, so keep that in mind, but the approach itself makes sense.

Source: OpenAI

OpenAI cut GPT-5.6 Luna’s API price by 80%. It now costs $0.20 per million input tokens and $1.20 per million output tokens. Terra dropped by 20% to $2 for input and $12 for output.

OpenAI says model and inference efficiency made the cuts possible.

But are open-weight models finally forcing frontier labs into a price war? The announcement does not prove that. Still, when capable alternatives become much cheaper, OpenAI cannot keep competing only on intelligence. Price and latency start mattering just as much.

Source: OpenAI

DeepSeek V4-Flash has 284 billion total parameters, yet it costs $0.14 per million cache-miss input tokens and $0.28 for output. DeepSeek also reports that its new post-training improved the model across every listed agent benchmark, including a jump from 72.1 to 82.7 on Terminal Bench 2.1.

So how is a model this large so cheap?

It does not use all 284 billion parameters for every token. Only around 13 billion are active at a time. That is why the headline parameter count tells you very little about the actual serving cost.

These are still vendor-reported benchmarks, and the final V4-Pro has not been released.

Sources: Flowtivity · DeepSeek

Nobody became a developer to write documentation. But the docs still need to get written — PRDs, README updates, architecture decisions, onboarding guides.

Wispr Flow lets you talk through it instead. Speak naturally about what the code does, how it works, and why you built it that way. Flow formats everything into clean, professional text you can paste into Notion, Confluence, or GitHub.

Used by engineering teams at OpenAI, Vercel, and Clay. 89% of messages sent with zero edits. Works system-wide on Mac, Windows, and iPhone.

Try Wispr Flow free

Google DeepMind’s new robotics models can control a full humanoid body, plan multi-step tasks, coordinate multiple robots, and adapt an on-device model to a new robot using fewer than 200 examples.

Apollo 2 can walk towards objects, crouch, pick them up, and place them on shelves. This is much closer to general-purpose robot behaviour than a model controlling one fixed arm inside a factory setup.

So, are humanoids about to have their ChatGPT moment in 2026? Maybe, but this release is still early access. Movement remains slow, multi-finger dexterity is difficult, and the models are available only to selected partners.

Source: Google DeepMind