OpenAI is reportedly in talks to raise at least $30B in a pre-IPO round at a ~$1.25T valuation, with possible UAE funding and a compute-capacity deal tied in. Terms are still being negotiated.
Meta's Muse Wonder is now available via an OpenAI-compatible endpoint at $4/$20 per million tokens — a base URL and model-name switch away for most stacks. A full safety report has not yet been published.
OpenAI gave METR and Redwood Research deep infrastructure access for three weeks to evaluate its internal monitoring system. Findings strengthened safeguards and will inform future third-party collaborations.
When a tool response omits a safety-relevant fact, an agent may carry out a prohibited action even though the user request is benign and every returned statement is true. The benchmark probes exactly this failure mode.
Chain-of-thought rationales appear to explain model predictions, but there is no guarantee they causally drive the final answer. This work proposes context isolation to make CoT steps genuinely causal.
Looped architectures grow linear KV-cache memory per recurrence during inference. BeNice introduces KV-sharing across recurrences to close the gap while keeping the quality benefits of looping.
Instrumental convergence is often invoked as a general reason to expect power-seeking subgoals in capable agents. Checkpoint evidence from OLMo-3 training runs complicates that picture.
Agentic RL rewards a complete episode, yet the policy needs credit for every decision inside it. The paper proposes twin-world counterfactual comparisons for finer-grained credit.
Inception Labs released Mercury 3, claiming 1,068 tokens/sec with a 1M-token context and an OpenAI-compatible API. It is live on Inception, OpenRouter, Baseten and Vercel AI Gateway at $0.10/$0.35 per 1M tokens.
BAND Desktop is a free workspace where multiple coding agents share rooms, hand off context and unblock each other, with developers able to step in from the terminal.
An npm package that lets AI agents observe iOS accessibility trees, tap real UI elements, verify screen state, and compile successful routes into replayable skills.
A developer tool that lets any AI agent generate video, images and audio on demand, bridging agentic workflows with multimedia generation.
MIMIT and the Italian Institute of AI for Industry opened a public consultation to map the national AI supply chain. Companies, startups and research centers can participate until November 20, 2026.
An Assolombarda-InnovUp study finds Italian innovative startups created 70,991 jobs by 2025, with AI, Data & Analytics firms growing headcount 12.2% year-over-year despite representing only 5.2% of the ecosystem.
Ben Lamm's Warhound emerged from stealth with $200M in funding for defense robotics and AI, backed by In-Q-Tel, Sequoia and 8VC.
Enact Intelligence, building general-purpose robot tech with force and touch sensing, raised a $30M pre-seed led by Redalpine, Notion Capital and Firstminute Capital, with angels Arthur Mensch, Yann LeCun and Thomas Wolf.
The French-American startup raised a $3M pre-seed led by Daphni for browser-automation tech (Reduck MCP) that gives AI agents access to sites without APIs.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
