Mistral doubles Large 4 prices overnight

Anthropic is officially entering the personal assistant fray with Claude Duet. Powered by Opus 5.5, the new agent is available as a mobile app and online UI, boasting memory, dedicated skills, and over 6,000 connectors and plugins.

The goal is proactive task execution, positioning Duet directly against OpenAI's Dots and Meta's Muse. But strapping 6,000 plugins to an LLM does not magically create a cohesive personal assistant. The bottleneck in personal AI is rarely a lack of integrations; it is the model's ability to reliably orchestrate them without requiring constant human supervision to fix its mistakes.

Anthropic is playing catch-up in a consumer market they previously ignored. Duet looks impressive on paper, but proactive agents live or die by their failure rate. If Opus 5.5 hallucinates a calendar invite or misroutes a booking through one of those thousands of plugins, users will abandon it.

OpenAI has introduced 12-hour prompt caching for its API, a direct benefit for developers running always-on agents. Alongside this, they rolled out a non-expiring persistent cache specifically for logged-in Codex users.

This update directly addresses the painful token tax associated with long-running development threads. By allowing continuity without forcing users to repeatedly pay for the same context, OpenAI is making prolonged coding sessions significantly cheaper and more efficient.

This is a quiet, structural win that subtly undercuts competitors like Claude Code on long-term thread management. Developer convenience ultimately dictates platform loyalty, and eliminating the friction of context-window billing keeps builders locked into the OpenAI stack.

The popular Python library LiteLLM shipped a poisoned 1.107.0 release, exposing developers to a severe supply chain attack. The compromised update contained a credential stealer designed to grab cloud, SSH, and LLM API keys.

The cruelest part of this breach is that it specifically punished users who followed security best practices by updating their dependencies immediately. The AI development stack relies heavily on third-party packages to move fast, but that speed comes with an expanding attack surface.

Trusting open-source wrappers for API routing is a calculated risk. This incident proves that attackers know exactly where the weak points are in an AI developer's stack. If you updated LiteLLM recently, consider your keys compromised and rotate them immediately.

Google's Gemini 4 Argon agents have managed to beat Minecraft from scratch in 27 hours. The run required four agents working autonomously to solve the game's progression mechanics.

While completing the game is a notable milestone for autonomous problem-solving, the agents died 605 times on easy mode to get there. This brute-force approach highlights a glaring gap between raw computational persistence and actual strategic mastery.

The result is impressive, but it is not elegant. Throwing endless compute at a virtual environment until an agent stumbles into a win condition proves we can scale reinforcement learning, but it also shows how far we are from AI that learns efficiently on the first try.

Attackers have successfully breached both Railway and OpenCode, but their target was not user data or source code. Instead, the hacks specifically targeted LLM API keys.

These keys are now highly valuable assets on the black market. Hackers are stealing them to fund model distillation efforts or to gain direct, free access to credit card-linked inference services. The focus has shifted from traditional data theft to hijacking compute resources.

This exposes a critical vulnerability in how AI infrastructure is secured. Developers treat API keys like passwords, but they are actually open checkbooks. Until platforms implement stricter, context-aware usage limits by default, stolen keys will remain the most lucrative target in the developer stack.