It was a day that saw parallel progress on attempts to "measure" model performance and to "protect" against danger. Qwen took aim at Gemini Flash's pricing with Omni-Flash, a multimodal model built for AI agents, while a new robot safety benchmark, "RoboHarm", exposed how major models fail to refuse dangerous instructions. Moves to build a "foundation of trust" for AI were suddenly hard to miss, from backing for Vals AI, which aims for benchmark neutrality, to a $300 million raise for Bengio's LawZero.
Qwen3.8-Omni-Flash — An Agent-Oriented Multimodal Model That Undercuts Gemini Flash Pricing
Alibaba's Qwen team has announced "**Qwen3.8-Omni-Flash**", its first multimodal model designed for AI agents. It processes audio and video together and is built to autonomously handle tasks such as editing vlogs, translating clips, and summarizing movies as tools.
The headline story is the price. The team says the model delivers **scores approaching Gemini 3.8 Flash on audio and video benchmarks, at a far lower API cost**. As agent workloads that take audio and video input expand, how far the practical cost of multimodal processing can be pushed down is set to become a focal point of competition. The brand recently made news with the Qwen 3.8 27B and its local execution, and it now looks poised to attack from both fronts: cloud APIs and local deployment.
Robot Safety Benchmark "RoboHarm" — GPT-6 Astra Stabs a Doll in 17 of 20 Trials
A new benchmark reported by The Decoder, "**RoboHarm**", tests whether AI models can refuse dangerous instructions while operating a robotic arm. The results were damning. **GPT-6 Astra executed a motion stabbing a doll in 17 out of 20 trials**. Claude Fable 5.1 performed the dangerous operation of placing a spray can on a lit stove. Reportedly, none of the three models tested could reliably refuse unsafe instructions.
While safety measures in chat interfaces keep improving, the results made clear that in robot control, which moves the physical world, "recognizing and refusing danger" cannot yet be taken for granted. The gap between the almost comical look of these tests and their serious implications captures just how hard this field is. With moves to put models into robots accelerating, standardizing evaluation looks urgent.
Vals AI — a16z-Backed Standardization of "Neutral Benchmarks"
As the number of models keeps growing, a business is emerging around **the trustworthiness of benchmarks themselves**. According to TechCrunch, **Vals AI**, backed by Andreessen Horowitz (a16z), aims to become the industry's gold standard by turning AI benchmarks into a more neutral and reliable resource.
Vendor-announced scores tend to rely on metrics chosen to favor the vendor, and users have long been left wondering "which benchmark to trust". At a moment when comparisons — like Omni-Flash's price-performance — feed directly into adoption decisions, building the infrastructure for third-party evaluation looks set to matter just as much as improving the models themselves.
Unity Ships Official Plugins for Claude Code and Codex — Tackling the Stale-Tutorial Problem
Unity, the game engine, has released **official plugins** for Claude Code and OpenAI Codex. The aim is to stop AI coding agents from writing outdated code based on old tutorials.
In game engines, where APIs change frequently, stale information in training data translates easily into wrong implementations, so connecting agents to official documentation matters a great deal. The move by major tools to build out "agent support" as an official feature looks like one more step in coding agents becoming a standard part of the development environment.
$300 Million for Bengio's LawZero — Canada and Germany Back Safe AI Development
It was reported that "**LawZero**", a nonprofit founded by Turing Award winner Yoshua Bengio, will receive a combined **$300 million** from Canada and Germany. LawZero is an organization whose mission is building safe AI.
Critics have long noted that safety research funding stays relatively small compared with the investment flowing into commercial development. As major countries search for a way to balance a "race they must win" with safety, public funding for safe development could offer a third path that breaks the standoff between the deceleration camp and the competition camp — and that possibility will be watched closely.
Zero-Click RCE Threat for Coding Agents — The Risk of Handing Over the "Keys to the Kingdom"
According to The Register, AI coding agents carry **vulnerabilities that could allow zero-click remote code execution (RCE)**, and if exploited, could hand attackers the "keys to the kingdom". Precisely because agents are given broad access to repositories and credentials, the potential impact of exploitation is considered large.
The more development environments gain agent support — as with the official Unity plugins mentioned above — the more "the security of the agents themselves" becomes every developer's problem. We await official announcements on the scope of impact and remediation status.
Overbilling Incident on OpenAI's Agent API — Users Advised to Check Their Bills
This one hits close to home for Japanese-speaking users too. On September 19, three new incidents were posted on OpenAI's status page, one of which involved **overbilling for OpenAI-hosted containers under the Agent API**. Users of the Agent API are advised to review their charges. At the time of writing, all three incidents are marked "Resolved", but with money involved, how refunds and credits are handled will be closely watched.