Back to Blog

AI News Roundup – 18 November 2026: GPT-6.2, Genie 4, Codex Sandbox & More

AI News
AI News Roundup – 18 November 2026: GPT-6.2, Genie 4, Codex Sandbox & More

Today's AI news: GPT-6.2 Astra & Sol launch, Google Genie 4 arrives, OpenAI Codex sandbox detail, chip design breakthroughs, and VLM safety concerns.

••4 min read

Here is what happened in AI today, Wednesday 18 November 2026.

GPT-6.2 Astra and Sol: Cheaper, But Mixed Reviews

OpenAI has released two new models — GPT-6.2 Astra and GPT-6.2 Sol — that lower benchmark costs by 20% to 600% compared to their predecessors. Astra sits as the number-two frontier model overall, but early testers are divided on what it is actually like to work with day-to-day. Sol, meanwhile, is shaping up as the likely main workhorse model for most developers and enterprise users, suggesting OpenAI sees it as the practical default even if Astra draws the headlines. The release also has a concrete hardware angle: OpenAI published details showing GPT-6 Astra was used to optimise circuits in its Jalapeño silicon programme and to improve RISC-V core designs — a notable demonstration that AI models are now being used to design the very chips that will run future AI.

Google DeepMind Launches Genie 4

Google DeepMind has made Genie 4 available, marking a new generation of its interactive world-modelling system. DeepMind's framing for the project centres on learning general intelligence from pixels, motion, and sound — the idea being that visual experience, rather than text alone, is a viable path toward broader AI capability. No detailed benchmarks were included in today's release notes, but the availability announcement confirms the model is now accessible beyond internal research.

OpenAI Opens Up the Codex Sandbox

OpenAI published an in-depth look at how its Codex coding agent operates inside a sandboxed environment across macOS, Linux, and Windows. The post explains how Codex runs commands on a user's machine while respecting filesystem and network limits that the user controls. The transparency here matters: as agentic coding tools become more capable, the security model underpinning them — what they can and cannot touch — becomes a critical concern for developers considering adoption in production environments.

Higher Benchmark Scores Don't Equal More Work Done

A research paper posted to arXiv delivers a pointed challenge to how the industry measures coding model progress. Researchers used a natural experiment — a 19-day US export-control suspension of Claude Fable 5 that forced 297 users to switch to Opus 4.8 (which scores 11 points lower on SWE-Bench Pro) — to isolate the real-world productivity effect of the higher-scoring model. Using a difference-in-differences model with developer and date fixed effects, they found that access to Fable 5 increased the 7-day pull-request merge rate by only 1.4 percentage points, a result that was not statistically significant. Across five separate model upgrades tracked in the same period, the weighted mean effect was actually slightly negative. The finding raises serious questions about whether coding benchmark improvements translate into genuine developer productivity gains.

Scene-State Safety Vulnerability Found in Vision-Language Models

A new arXiv paper introduces what researchers call a scene-state safety vulnerability in vision-language models (VLMs). The core problem: VLMs can provide harmful assistance simply by interpreting what is happening in an image — even when no explicit harmful request is made in text. The researchers created SceneSafetyBench, a dataset of 375 image–question pairs, and tested eight VLMs, finding attack success rates of 45.6% to 72.3%. They also propose a systematic attack method called S3TA, which achieves a model-average attack success rate of 71.2% across eight open-weight and commercial VLMs — exceeding the strongest prior baseline by 6.1 percentage points. Text-side filters and visual moderation provided limited protection, highlighting a gap in current safety approaches.

Cloudflare Repositions as an AI-Native "Neolab"

Cloudflare has announced it is rebranding its strategic identity around AI, describing itself as a neolab focused on building what it calls System One models — AI designed to make decisions and retrieve data in milliseconds. The company, long known for accelerating and securing web traffic, is positioning this infrastructure-level AI capability as the natural evolution of its network edge business. It is an early but significant signal that major internet infrastructure providers see real-time AI inference, not just content delivery, as their core value proposition going forward.

Frequently Asked Questions

What are GPT-6.2 Astra and Sol?

GPT-6.2 Astra and GPT-6.2 Sol are two new OpenAI models released in November 2026. Astra is positioned as the number-two frontier model and is 20% to 600% cheaper than predecessors, though early user reviews are mixed. Sol is expected to be the main workhorse model for most developers and enterprises.

What is Google DeepMind's Genie 4?

Genie 4 is the latest generation of Google DeepMind's interactive world-modelling system, made publicly available on 18 November 2026. DeepMind's approach focuses on learning general intelligence from pixels, motion, and sound rather than text alone.

Do higher coding benchmark scores mean AI models get more real work done?

A new arXiv study suggests not necessarily. Using a natural experiment created by a 19-day suspension of Claude Fable 5, researchers found that having access to the higher-scoring model increased the pull-request merge rate by only 1.4 percentage points — a result that was not statistically significant. Across five model upgrades, the effect was slightly negative.

What is the scene-state safety vulnerability in vision-language models?

Researchers identified that vision-language models can provide harmful assistance by interpreting what is happening in an image, even without an explicit harmful text prompt. Testing across eight VLMs showed attack success rates of 45.6% to 72.3%, and a proposed attack method called S3TA achieved a 71.2% average attack success rate, exceeding prior baselines by 6.1 percentage points.

Explore Topics

#ai-news#roundup