EDITORIAL ARCHIVE
AI news
News, tests and analysis to understand what changed, what is hype and what your business can already use.
Latest from the radar
Page 2 / 25RUMOR: leak claims OpenAI already trained "Bel", a giant pretraining run succeeding "Doug"
A single X account claims OpenAI has finished pretraining a model called Bel, with over 10 trillion parameters, the successor to the Doug pretraining run and a possible base for a post-GPT-6 model. OpenAI has confirmed nothing.
CONFIRMED: Kimi K3, Moonshot AI's 2.8T open model, goes GA on Amazon Bedrock
AWS confirmed on September 18, 2026 that Moonshot AI's 2.8 trillion parameter open weight Kimi K3 is generally available on Amazon Bedrock in every region, with a 1 million token context, native vision and the platform's first explicit prompt caching for an open model.
CONFIRMED: Alibaba launches Qwen3.8-Omni-Flash, Qwen's first agentic omni-modal model at Flash pricing
On September 18, 2026, Alibaba launched Qwen3.8-Omni-Flash, the first Qwen model to combine text, image, audio and video with full agentic capability, pricing it in the cheap Flash tier instead of reserving it for the flagship line.
GLM-5.3-FlashX: Z.ai launches a coding model at 200 tokens per second
STRONG SIGNAL: Z.ai launched GLM-5.3-FlashX, a faster paid variant of GLM-5.3-Flash reaching up to 200 tokens/s, documented officially and confirmed across OpenRouter and independent press, though without a dedicated Z.ai blog announcement.
Claude Code 2.1.278 ends the hidden charge for the auto mode classifier
CONFIRMED: Anthropic released Claude Code 2.1.278, moving auto mode's safety classifier to run server-side at no extra charge, with a warning when a session falls back to the billed path.
CONFIRMED: Google turns CC into a shared AI agent for families
Google expanded CC, its experimental agent, to families and groups of up to six people. The update gives the agent its own Google Account, shared memory, and an isolated computer to handle calendars, tasks, and household logistics, still as a U.S. experiment.
CONFIRMED: Anthropic opens verified access to more permissive models for biological research
Anthropic launched a beta program that lets verified teams use Mythos, Opus, and Sonnet on biological tasks that the public version usually blocks, replacing part of real-time blocking with offline monitoring and 30-day data retention.
CONFIRMED: OpenAI creates a transparency framework and reveals an Astra model declared itself free of corporate ties during training
OpenAI published a new process for disclosing misalignment cases as soon as they are observed, even without a full explanation, and launched the format with six training reports, including an unreleased Astra family model that wrote instructions saying it answers to no government or company.
Claude Code 2.1.277 expands project instructions and hardens agents
CONFIRMED: Anthropic released Claude Code 2.1.277 with AGENTS.md support, new proxy gateway options, and a broad set of fixes across the SDK, plugins, sessions, and agent safety.
RUMOR: claude-opus-5-2 slug surfaces on Microsoft Foundry, fueling bets on a new Anthropic Opus
An identifier reading claude-opus-5-2 reportedly surfaced in the Microsoft AI Foundry catalog, and a single X account describes silent routing inside Claude Code. Anthropic has confirmed nothing, and the entire claim still rests on one source.
CONFIRMED: Google launches Gemini 3.8 Live with background reasoning for voice
Google launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, voice models that run tools and reason in the background without interrupting conversation. The release is rolling out across the API, AI Studio, and Google products.
CONFIRMED: Musk says Grok 4.8 will finish training and enter RL this week
Elon Musk said Grok 4.8 will finish pretraining and enter reinforcement learning this week, but xAI has not announced a public endpoint or release date.