Claude Opus 5 is available today. It's a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price. On coding and knowledge work evaluations, Opus 5 is the new state-of-the-art, though it remains behind Mythos 5 on cybersecurity tasks. Opus 5 is designed for daily use and is now the default model on Claude Max and the strongest model on Claude Pro. It offers improved performance at the same cost as its predecessor, with strengths in software engineering, scientific research, and visual outputs. Users praise its judgment, self-verification, and iterative capabilities. The model also demonstrates the best alignment and safety scores to date.

ElevenLabs outlines how voice AI can make election information more accessible while addressing misuse risks. The company partners with Brazil's TSE and U.S. entities to enhance democratic access, and details multi-layered safeguards including No Go Voices, Voice CAPTCHA, content moderation, and watermarking to prevent voice AI abuse.

We are introducing a new Model API tier, GLM-5.2 Fast, optimized for real-time agentic workloads with the same weights as standard GLM-5.2 but on infrastructure tuned for per-user throughput, offering lower latency and stable performance.

This article compares H100 and H200 GPUs for AI inference. H100 is cost-effective for small-to-medium models and low/sporadic traffic. H200, with larger HBM3e memory and higher bandwidth, is better for very large models, long-context inference, and workloads needing more KV cache headroom. Techniques like MIG and async programming are also covered.

Cut LLM inference latency and cost with continuous batching, speculative decoding, quantization, and more. Learn which techniques fit your workload.

Lessons from Notion and Gamma on picking AI models: swap for cost and reliability, match model to task, lean into open-weight models, and fine-tune to grow the market.

Template OCR works perfectly in demos because the layout never changes, but in real-world scenarios, layout variations cause silent errors. The real alternative is agentic OCR (e.g., LlamaParse) that reads documents by structure, not coordinates, eliminating template maintenance, onboarding delays, and drift. It processes new vendor invoices on the first sight, provides confidence scores for targeted human review, and excels in workflows with high format variability like accounts payable, remittance advice, and logistics.

Anthropic launches the Economic Index connector for Claude, allowing users to ask questions about AI's impact on work using real data. The connector is easy to enable and provides answers grounded in the index's data, with transparency about its limitations.

Anthropic commits $200 million to a research fund supporting external studies on AI's economic impacts, focusing on five priority areas to prepare society for AI-driven disruption.

Cursor launched Cursor Router, an intelligent model routing system for teams and enterprises. It automatically routes each request to the most capable model, reducing costs by 30-60% while maintaining frontier performance. Online A/B tests with millions of requests showed 60% cost savings, and early enterprise customers achieved 30-50% reductions.

Anthropic is donating an additional $20 million to Public First Action, a non-partisan organization focused on AI public education and policy. The donation supports efforts to promote sensible AI safeguards. The article explains the urgency due to rapid AI advances, such as Claude Mythos discovering critical vulnerabilities, and outlines policy proposals including transparency, independent evaluations, and export controls to maintain US leadership and mitigate risks.

Researchers from Baseten successfully added vision capabilities to the open-source language model GLM 5.2 by training only a small 50M-parameter MLP projector, achieving performance comparable to Claude 4.5 Haiku on MMMU-Pro (55%). The process involved SFT and RL with grokking observed, and demonstrated generalization to unseen images.

Heidi Health partners with Fireworks AI to surpass proprietary frontier model quality using SFT and RFT, achieving 3.5x lower latency, 7-second inference time, and significant cost savings. Success hinges on aggressive data filtering and large batch sizes (1.5M effective tokens), enabling open models to rival frontier models.

Fireworks AI announces research showing open-source Kimi K3 is competitive with closed-source Fable 5 across ~1,000 agentic tasks. Routing between them achieves 93% accuracy and up to 50x cost savings, suggesting the best AI comes from a mixture of models.

Unstructured introduces Transform MCP, a tool that lets AI agents process documents as a tool call. It handles 60+ file types, provides structured output, and integrates seamlessly with MCP clients. Key benefits include composability, inspectability, reproducibility, and portability. Available now with a free tier of 15,000 pages per month.

Mortgage document automation often fails at handoffs between stages, not within stages. Without data provenance, every stage re-verifies numbers, increasing costs and closing times. Agentic extraction, like LlamaParse, provides page-level citations and confidence scoring, enabling targeted human review and tractable audit trails.

Anthropic is launching a focused call for rare disease research grants under its AI for Science program, offering up to $50,000 in Claude credits over six months. Two tracks are available: basic science and biotech. Applications are due August 2, 2026.

Cursor's experiments show that a hierarchically structured agent swarm with planner and worker roles significantly outperforms single-agent architectures on complex tasks like building SQLite from scratch. The new design leverages tree decomposition, a custom version control system, and multiple coordination mechanisms to address context drift, conflicts, and efficiency, achieving up to 100% test pass rates across various model mixes.

Baseten's optimized runtime for Wan 2.2 achieves 2.75 seconds per video clip, a 53.6x improvement, using timestep distillation, custom kernels, and NVFP4 quantization.

Kimi introduces its most capable model, K3, with 2.8 trillion parameters, built on Kimi Delta Attention and Attention Residuals, featuring native vision and a 1-million-token context window. It is the world's first open 3T-class model, designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning. While still trailing the most powerful proprietary models, it outperforms other tested models in evaluations.

NVIDIA Nemotron 3 Embed models are now available in Baseten. The larger 8B model delivers leading retrieval accuracy, while the smaller, faster 1B model retains 95% of that accuracy. Both are now available on Baseten for AI agents, enterprise search, and code retrieval.

Fireworks AI built a Blackwell (SM100) kernel for MiniMax M3 sparse attention that uses a KV-stationary execution path, loading each KV block once, achieving ~980 TFLOP/s throughput and ~4.1 TB/s HBM bandwidth, with 1.9–2.4× speedup over a query-stationary baseline (FlashInfer) and ~1.6× over open-source MSA kernels. The article details the algorithm, kernel design, optimizations, and I/O cost model.

Fireworks announces a $1.505 billion Series D at a $17.5 billion valuation, surpassing $1 billion in annualized revenue and processing over 40 trillion tokens daily, with 95% from specialized customer models.

LlamaIndex team doubles in size, releases updates: conversational extract, markdown in fast tier, improved tables in Agentic Plus, and 100 users on every plan. AI news includes GPT-5.6 benchmark, Bun rewriting Zig in Rust, Apple suing OpenAI.

Thinking Machines Lab has released Inkling, a 975-billion-parameter open-weight multimodal model that reasons over text, images, and audio. With a mixture-of-experts architecture and 41 billion active parameters, it balances performance and cost. Baseten provides day-zero support for deployment and scaling.

The article highlights the operational and governance challenges of autonomous AI agents, emphasizing that security and human oversight are more critical than model improvements. It covers execution-layer security, supply chain risks from malicious skills, operational hygiene failures, and the need for human review in regulated environments.

As enterprise AI moves from experiment to necessity, understanding the Total Cost of Ownership (TCO) is critical. This article analyzes visible and hidden AI costs, including token pricing, trade-offs between owning and renting infrastructure, and cost reduction through efficient deployment. Research shows most organizations face unexpected cost overruns, while owning hardware at scale can yield significant savings.

Cohere Labs' Tiny Aya is an open-weight, lightweight multilingual model supporting 70+ languages and capable of running locally. The Expedition Tiny Aya research program brought together global builders to explore applications in education, safety, accessibility, and language understanding, yielding insights and tools that advance multilingual AI.

Anthropic launches Claude for Teachers, offering free premium Claude access to verified K-12 educators in the US, a library of teaching skills, and connections to evidence-based curricula mapped to all 50 state standards. The tool helps with lesson planning, differentiation, and data analysis, while prioritizing data privacy.

Anthropic announces a $10 million CAD donation to Canadian research institutions to support beneficial and responsible AI, partnering with top institutes and providing Claude credits. The Canadian economic index shows Canada ranks 8th globally in Claude usage, with per-capita usage far above expectations.