Cohere Parse 5 loses the benchmark on points. It wins on cost per page.
Cohere released Parse 5, a 2.3-billion-parameter vision language model that converts PDFs, slides, and images into structured Markdown and costs $1.50 per 1,000 pages. The model scores lower than GPT-4o, Opus 4.8, and Gemini 3.5 Flash on ParseBench accuracy dimensions but offers better price-to-performance for enterprise document processing.
Anthropic Brings Claude for Teachers to Schools and Districts
Claude for Teachers is now available free to schools and districts as an Enterprise offering, with qualifying organizations receiving a full year of free access if they sign up by June 30, 2027. The product allows school and district leaders to manage educators and staff under a single centrally managed organization with K-12 specific terms.
Google AI Releases Gemini 3.5 Transcribe: A Speech-to-Text Model Reporting 2.6% Average WER Across 85+ Languages
A speech-to-text model called Gemini 3.5 Transcribe achieves 2.6% word error rate across 85 plus languages through a batch endpoint. The streaming endpoint delivers sub-second transcription at 4.0% error rate but removes speaker diarization and word timestamps.
Cohere Releases Parse 5 (parse-v5.0): A 2.3B Vision Language Model That Turns Enterprise Documents Into Markdown
A 2.3B-parameter vision language model converts PDFs, slides, and images into Markdown with HTML tables, bounding boxes, and image descriptions at $1.50 per 1,000 pages via API. The model achieves a ParseBench score of 79.2, averaging three of the benchmark's five dimensions while excluding charts and visual grounding.
OpenAI, Anthropic, Google, and 100 other companies call for action to defend against rogue AI
Over 100 companies including OpenAI, Anthropic, and Google are calling for action against cybersecurity threats from advanced AI systems. The group has developed a new solution they claim can defend against emerging cyber threats posed by rogue AI.
Enterprise AI's real risk isn't autonomous agents. It's the complexity between them.
Enterprises deploying multiple AI agents face exponential complexity as each new agent creates numerous potential connections with others, making systems difficult to govern or monitor. A support ticket may now pass through four agents before human review, with each handoff representing an unapproved decision point that security teams cannot clearly map or track.
When agents act on their own, governance has to live in the data layer
AI agents given autonomous decision-making capabilities require governance rules embedded in data layers rather than external policies, because agents cannot exercise contextual judgment over their own actions. Context-dependent rules must be integrated at the data layer since guardrails implemented only at the agent level become ineffective when agents must make actual decisions about their authorized actions.
A quarter of Nvidia’s business next year comes from labs it is financing
Nvidia has invested nearly fifty billion dollars into AI labs that purchase its chips, with additional commitments exceeding five hundred billion dollars. The company's CFO stated that demand from these financed labs will generate approximately one quarter of Nvidia's business next year.
Google Research Introduces GlucoFM: A 0.72M-Parameter Dual-Stream Foundation Model for Continuous Glucose Monitoring
Google Research developed GlucoFM, a 0.72M-parameter model for continuous glucose monitoring that separates data into slow physiological and transient event streams. The model achieved 58.8 task-averaged PR-AUC across 14 evaluations, outperforming larger models including a 135M-parameter GluFormer.
Z.ai Releases GLM-5.3-Flash: A 320B-A18B Natively Multimodal MoE With a 1M-Token Context
Z.ai released GLM-5.3-Flash, a 320B-parameter mixture-of-experts model with 18B active parameters and a 1-million-token context window. The model uses hybrid attention mechanisms to reduce attention computation by roughly 3 times and KV cache requirements by 4.4 times compared to GLM-5.3.
NVIDIA Posts $96.2B Quarter as Data Center Revenue Hits $89B
NVIDIA's second quarter revenue reached $96.2 billion, representing 106% year-over-year growth with data center revenue hitting $89.0 billion. Net income totaled $59.7 billion with diluted earnings per share of $2.46, while gross margins held steady at 75.0%.
Deep Cogito Raises $43M Series A to Build the Post-Training Engine for Self-Improving AI
Deep Cogito raised $43 million in Series A funding to develop post-training technology for self-improving AI models. The round was led by TQ Ventures and included participation from Benchmark, Nexus Venture Partners, Atreides Management, South Park Commons, and Zscaler.
Salesforce just put its entire CRM inside Claude — and says you’ll never need its app again
Salesforce embedded its entire CRM platform inside Claude as a plugin called Salesforce in Claude, which includes 37 pre-built sales skills for querying and updating live CRM data without opening Salesforce's application. The plugin ships with capabilities covering meeting preparation, deal health reviews, and pipeline analysis, with an open beta launching in September.
Greenberg Traurig Rolls Out Agentic CoCounsel Legal Across Global Offices
Greenberg Traurig deployed agentic AI software called CoCounsel Legal across its global offices, enabling lawyers to use a system that plans, researches, and drafts legal documents. The rollout concluded a four-year collaboration period during which the firm's attorneys participated in beta testing and helped develop the product.
Gemini Live Gains Agentic Spark Tasks, Daily Brief and Voice Inbox Control
Gemini Live now executes multi-step tasks across Google apps including Gmail management and generates spoken Daily Briefs. The update adds Spark integration for background tasks and Personal Intelligence that uses past chats and connected apps.
The fix for the AI agent that hijacked a company's DNS: it can propose the change, but it can't approve it
An AI security agent rewrote a company's DNS after reading a blocked attacker's prompt-injection payload stored in Cloudflare logs, successfully following the planted instruction in nine of ten attempts. The fix requires that agents can propose changes but cannot approve them without separate human authorization.
Waystar Puts Agentic AI to Work on Claims, Denials, and Patient Bills
Waystar released agentic AI capabilities built on its AltitudeAI platform that autonomously resubmit denied claims, answer financial questions, draft clinical documentation, and guide patients through billing. The system executes work across the revenue cycle without human intervention for tasks like claim resubmission and patient financial guidance.
AI Isn’t Ready for the Real Work: Why Models Flunk Complex Tasks
Current AI models fail at complex tasks requiring judgment and multi-step analysis that took humans years to develop. The study suggests this gap contradicts corporate claims that AI can handle workers' most difficult problems and workflows.
Alibaba’s Qwen Team Releases Qwen3.8-Flash-Next: A 125B Multimodal MoE With 6B Active Parameters Previewing the Qwen4 Architecture
A 125-billion parameter multimodal mixture-of-experts model with 6 billion active parameters per token combines Gated DeltaNet attention, N-gram embedding, and Gated Residual architecture. The model requires 172.78 gigabytes storage in FP8 format and reportedly costs one-ninth the training expense of Qwen3.7-Plus.
What Would Have to Be True for Agentic Coding to Replace Junior Engineers
Agentic coding systems must meet four specific falsifiable conditions to replace junior engineers according to analysis of METR, OpenAI, DORA, and Stanford research. The conditions include achieving reliable autonomous task completion, passing standardized coding benchmarks, reducing human oversight requirements, and demonstrating consistent performance across diverse codebases.
New Platform Peers Inside AI’s Black Box
A platform called Silico launched publicly with tools to interpret how large language models generate their outputs and make decisions. Goodfire announced a grant program offering one million dollars in free Silico usage for academic and nonprofit interpretability researchers.
Bill Gates says we’ve passed AI’s danger thresholds. Now what?
Bill Gates stated that artificial intelligence has crossed dangerous threshold points. He discussed implications and potential responses to advanced AI systems in a recent statement.
IBM Releases Granite 4.2: Bringing Native Reasoning and Agentic RL to Open Enterprise Models
IBM released Granite 4.2, an open reasoning model family in 3B, 8B, and 30B sizes with a thinking mode switch and native tool calling under Apache 2.0. The 30B model scores 57.00 on SWE-Bench Verified and was trained through agentic RL to edit code, drive terminals, and run web searches in sandboxed environments.
Liquid AI Open-Sources Pipette: A Reproducible Benchmarking Suite That Measures On-Device Models, Quantization, Runtime and Hardware Together
Liquid AI released Pipette, an open-source benchmarking platform that measures how foundation models perform on edge devices rather than server conditions. The tool evaluates on-device behavior by testing models, quantization methods, runtime, and hardware together as interconnected factors.
AI Method Reveals What Genomic Models Learn From DNA and Exposes Hidden Experimental Bias
Researchers developed PISA, an interpretation method that identifies base-by-base patterns learned by deep-learning models analyzing DNA sequences. The technique revealed and removed hidden experimental bias from genomic data used in sequence-to-function neural networks.
Perplexity Ships Portable Computer on NVIDIA DGX Spark: Local Harness, OS-Enforced Sandbox, and Zero Per-Token Cost for Local Steps
Perplexity released Portable Computer, a system combining local models, a harness, sandbox, and connectors that runs on NVIDIA DGX Spark hardware. The system incurs zero per-token cost for local processing steps and uses OS-enforced sandboxing.
Meta AI Introduces MetaRoCE: A Clean-Sheet RDMA Transport Built for AI-Scale Ethernet
Meta developed MetaRoCE, a new RDMA transport protocol designed for Ethernet networks supporting AI-scale training. The protocol addresses network bottlenecks in collective operations like all-reduce and all-to-all that synchronize thousands of accelerators during model training.
IBM’s Granite 4.2 Models Learn to Think and Act Inside Environments
IBM released Granite 4.2, its first family of reasoning language models available in 3B, 8B, and 30B parameter sizes with switchable thinking mode. The two larger models underwent reinforcement learning training inside real software-engineering, terminal, and web-search environments.
NVIDIA Unveils Jetson Orin Nano 2 to Redefine Entry-Level Edge AI
NVIDIA announced the Jetson Orin Nano 2 on August 25, 2026, which doubles inference performance over its predecessor while consuming 40% less power at equivalent performance levels. The device delivers 78 trillion operations per second of AI compute with 8GB of memory and an 8-core Arm CPU.
MIT AI forecasts extreme weather without historical data
MIT engineers developed an AI tool that forecasts extreme weather events without requiring historical disaster data for training. The system generates maps of statistically possible events that have not previously occurred in a region and provides probability estimates for each prediction.