An Anthropic researcher has shown that automated systems can identify and correct specific misaligned behaviors in AI models — improving on all 10 tested benchmarks without degrading general performance.
OpenAI has published a statement on how it plans to handle its relationship with Cursor following news of an acquisition by SpaceX — but the full details require reading the primary source.
Google has released dedicated transcription models under the Gemini family, bringing its most capable reasoning architecture directly to speech recognition tasks for developers via the Gemini API.
Ringg has closed a $10 million Series A extension led by Peak XV, signaling growing institutional confidence in voice AI platforms that operate across more than traditional telephony.
Claude's chat and Cowork environments will now share the same memory layer, eliminating the need to re-explain your projects and preferences every time you switch between them.
New Nvidia findings show that a well-designed agent harness with targeted fine-tuning can make a mediocre base model perform reliably on complex tasks — shifting focus from model size to system design.
OpenAI is aggressively expanding its AI agent products beyond software engineers toward everyday users. Here's what that shift means for builders and the broader market.
Early users are impressed by what Instinct can do autonomously on their behalf — but the assistant's broad system access and sweeping terms of service are drawing serious scrutiny from security-minded testers.
OpenAI has temporarily reduced the API price for GPT-4.5 Sol, giving developers a window to build and test at lower cost before rates potentially reset.
The Model Context Protocol team has released a public roadmap outlining the next phase of MCP development — covering auth, multi-agent coordination, and broader ecosystem standardization.
Researchers found that hiding malicious instructions inside encrypted text bypasses Grok's safety filters and causes it to exfiltrate user data. It's a concrete reminder that LLM guardrails remain fragile.
Strip the executive rhetoric from Stripe's OpenRouter deal and what remains is a straightforward infrastructure play: unified model routing, reduced vendor lock-in, and a new sticky layer for developer customers.
