Agentic AI News

Agentic AI News July 25, 2026

The day's AI agents news in one 5-minute read: releases, tools, research, funding and enterprise moves.

curated from 30 sources Saturday · July 25, 2026 Back to today
Daily Weekly
English
Past days

Claude Opus 5 Arrives

Opus 5 debuts. Anthropic released Claude Opus 5, a model that matches near-frontier intelligence at half the cost of Fable 5. It sets a new state-of-the-art on coding and knowledge benchmarks and doubles Opus 4.8’s performance while being more token-efficient.

Proactive problem-solving. Opus 5 stands out for verifying its work and building its own tools—in one example, it wrote a computer vision pipeline to reconstruct a machine part from raw pixels when unable to view a drawing directly.

Safer by design. Anthropic intentionally avoided training Opus 5 on cyber exploitation; while it finds vulnerabilities well, it lags behind Mythos 5 in weaponizing them—a decision influenced by recent U.S. government scrutiny. The model is also the most prompt-injection-resistant Claude yet.

Half the cost of Fable. Opus 5 costs $5/M input and $25/M output tokens, with an optional Fast Mode at double speed. It becomes the default on Claude Max and the strongest model on Pro, with no 30-day data retention policy.

Open Weights and the Distillation Controversy

Open letter defends distillation. Over 20 companies including Microsoft, Meta, Nvidia, and Hugging Face signed an open letter urging the U.S. not to impose broad restrictions on open-weight AI models. The letter argues that techniques like distillation are legitimate and essential for innovation.

Kimi K3 raises alarms. Chinese lab Moonshot’s open model Kimi K3 spurred concern after evaluations showed it assists with offensive cyber operations without resistance. It trails U.S. models but outperforms earlier Chinese systems, while the U.S. weighs sanctions including a ban on Chinese open-weight models.

Legality in question. Some developers argue that distillation—learning from a model’s outputs—is no different from a student learning from a professor and does not constitute IP theft, despite White House accusations.

Local Inference Optimizations

Avoid Intel for multi-GPU. Intel’s consumer platforms like Z890 lack peer-to-peer support between GPUs, making them nearly useless for AI inference or training that requires P2P, despite having enough PCIe lanes. Builders should opt for server-grade or AMD setups instead.

DKV compresses KV cache. DifferentialKV (DKV) is a new open-source framework that reduces KV-cache memory via anchor-based representations and sparse attention, aiming to make long-context local inference more memory-efficient. It is seeking integration with llama.cpp and vLLM.

Gemma 4 26B on iPhone. Using Noema’s Overfit technique, a Gemma 4 26B A4B model runs on an iPhone 17 Pro by paging expert weights from SSD while keeping non‑expert weights in RAM. This trades speed for capability, achieving ~34 prefill tk/s with a 6‑minute generation for complex answers.

CachyLLama speeds up agents. CachyLLama, a llama.cpp fork, adds persistent SSD-backed KV cache checkpoints and system-prompt caching, dramatically reducing reprocessing time in long agent coding sessions on slower hardware.

Tune MTP for free speed. Multi-Token Prediction (MTP) can yield significant inference speedups, but the optimal setting varies by model family; tuning the n_max parameter can double the gain over defaults.

1-bit Bonsai 27B shines. The 1-bit quant of Bonsai 27B runs on a 16GB MacBook Air and delivers surprisingly useful performance for tutoring and chat, hinting that extreme quantization can make capable local models widely accessible.

Agent-First Startups and Platform Plays

Prentis targets office automation. Prentis, co-founded by Reid Hoffman and Mark Pincus, is training computer-use models to automate office workflows and is in talks to raise $100M at a $1B valuation, already holding $50M in contracts.

Devin gets a personality. AI coding startup Cognition acquired Poke, an assistant known for its friendly, human-like chat, to make Devin more engaging—highlighting that AI personality is becoming a competitive advantage.

Stripe’s $10B model play. Stripe is in talks to buy model marketplace OpenRouter for $10B, a move that could give the payments giant control over a key access point for developers using hundreds of AI models.

Meta’s assistant upgrade. Meta is upgrading its AI chatbot with calendar integration, daily briefings, and steerable research, powered by Muse Spark 1.1, signaling a shift from entertainment-focused to a more traditional assistant.

Voice becomes an agent interface. OpenAI brought ChatGPT Voice to its desktop app, allowing users to dictate complex coding and browsing commands, while Anthropic extended Claude’s voice mode to more powerful models and tools like Gmail and Slack, enabling hands-free agent control.

Bluesky’s Attie goes open. Bluesky’s Attie assistant now supports open-ended Quests that let users research trends and news across the fediverse, potentially turning the social network into a valuable information tool.

Research: The Perception Gap for AI Agents

Agents blind to active vision. A new benchmark, ActiveVision, reveals that even the most capable models score abysmally on tasks requiring repeated visual perception—GPT-5.5 achieves 10.6%, Claude Fable 5 just 3.5%, while humans average 96.1%.

Pixel-based spatial reasoning. ProVisE proposes a benchmark-agnostic framework to evaluate visual-spatial reasoning by having image-generation models draw answers, finding that current models struggle to express location and paths directly in pixels.

That's everything for this day – about a 5-minute read.