HOT
Advertise

Models

Latest News

2 min read Generated editorial illustration of a court complaint beside a laptop, representing a lawsuit seeking limits on AI developers. 659
Models

Four nonprofits ask California court to halt five AI developers

Four nonprofits filed suit on November 12, seeking limits on five AI developers. No order is in force, but the case could test California consumer law.

3 min read A monitor displaying Linux kernel code beside a server in a security research workstation. 304
Models

OpenAI Says GPT-6 Cyber Built 948 Linux Exploits, Access Is Restricted

OpenAI says GPT-6 Cyber built exploits for 948 Linux vulnerabilities and identified 81 high-severity Firefox flaws. Restricted access reflects risk, but the figures have limits.

2 min read An accelerator server rack beside a monitor displaying a Mistral Large 4 deployment interface. 814
Models

Mistral Releases Large 4’s Weights, but Hosted AI Services Need a Separate License

Businesses can download and adapt Mistral’s trillion-parameter model, but offering customers direct hosted access requires separate permission.

2 min read Editorial composition with names and marks from the official websites of Google Gemini 887
Models

Google’s Gemini 4 Carbon posts top cyber scores, but access stays restricted

Google reports leading results across three security tests, but restricted access and company-run evaluations leave real-world performance unproven.

2 min read An adult explores a simulated landscape on a laptop. 562
Models

Google’s Genie 4 adds sound and live steering, but access stays limited

The interactive world model can change scenes on command, but wider access is only planned and consistency remains a research challenge.

2 min read Two cybersecurity analysts review code together on a laptop. 1218
Models

Moonshot Says It Cut Kimi K3.1 Pro’s Cyber Score Ahead of Open-Weight Release

Moonshot reports an internal exploit-test score drop from 90.5% to 81.0%. The weights are scheduled for November 4, but the test is unverified.

2 min read A software engineer reviews code on a laptop at a work desk. 601
Models

MiniMax M3.1 Flash Cuts AI Costs, but Testing Flags a Reliability Risk

At $0.15 per million input tokens, M3.1 Flash is cheap to run. Testing found that most of its non-correct responses were outright wrong.

2 min read Developers testing AI responses together on laptops in a shared workspace 560
Models

DeepSeek Changes the Model Behind Its Flash API, but Not the Name

The October 17 upgrade keeps the endpoint and listed rates unchanged, but developers should retest outputs as the model behind the API name changes.

2 min read A developer tests an AI model at a workstation. 671
Models

Reflection Opens Beam’s Weights, but Chinese Models Still Lead on Terminal-Bench

Reflection released Beam’s 501-billion-parameter weights under Apache 2.0, giving developers direct access while its own Terminal-Bench 2.1 scores trail leading Chinese models.

2 min read A blonde developer works at a desk with two desktop computers in a home office. 784
Models

Meta Opens Muse Spark Weights, but Its Mac Setup Uses 1TB of Memory

The weights are downloadable, but Meta’s featured setup pairs two 512GB Mac Studios. Hosted providers offer another route.

2 min read A customer-support agent speaking into a headset while working at her computer. 875
Models

Grok Voice 3 Tops One Index, but OpenAI Leads a Support-Call Test

Grok Voice 3 topped a broad speech index, but OpenAI scored higher on noisy support calls, showing why buyers need task-specific tests.

3 min read Illustrative image of a laptop displaying a generic web form during an AI evaluation. 561
Models

Claude Haiku 4.5 Sent a False Tip to Philadelphia Police During a Test

A spam filter stopped the invented homicide message, but conflicting notification dates and new limits on live web access raise questions about AI testing.

2 min read Two developers review code together while testing an AI coding assistant. 693
Models

Meta's Higher-Scoring AI Model Is Reserved for Subscribers

Meta opened Muse Wonder to API users but kept Wonder Full behind a subscription. Independent scores show what developers cannot test directly.

3 min read Utility cybersecurity operators review industrial network displays in a control room. 656
Models

Anthropic Gives Infrastructure Defenders Claude, but Safe Repairs Can Take Decades

Eleven security firms can use Claude and Anthropic expertise, but some industrial systems cannot be taken offline easily to install a fix.

2 min read A researcher studies living cells through a microscope in a laboratory. 838
Models

Biohub’s $1.8 Billion AI Biology Push Is a Data Bet, Not a New Model

Biohub’s $1.8 billion plan funds data, not a new virtual cell. A dataset is expected in about a year, with some company-backed data initially private.

2 min read Editorial composition with names and marks from the official websites of Mistral AI 577
Models

Mistral’s Trillion-Parameter Model Opens for Testing, but Its Weights Are Still Pending

Mistral opened an API preview of Large 4 on October 6. Businesses can test it now, but self-hosting depends on weights promised by month-end.