Earlier studies showed that Opus 4.6, GPT 5.4, Gemini 3.1-pro-preview, Deepseek R1-0528, and Qwen 3.6-plus were unable to find two of the vulnerabilities discussed in the Mythos blog post without extremely revealing hints. This study continues the previous experiments with 26 distinct Claude-4.6/4.7 and GPT-5.4/5.5 combinations and different context window sizes and reasoning efforts. It found that higher reasoning effort, and even later models, are not always better for triaging security results.
An experiment compared Kimi K2.7 Code with Claude Fable 5 to generate 12 landing pages each. Kimi K2.7 Code cut costs by 94%, being 16 times cheaper than Claude Fable 5.
Notion surveyed more than 6,000 professionals across 10 markets [10], and guess what? Most orgs are still early in their AI transformations. It's not just you: trust, governance, and workflow integration are sticking points for leaders everywhere. Dive into the research and see what can push your teams forward at the live webinar on June 24. SAVE
Vercel launched Connect in public beta, replacing long-lived provider tokens with runtime credential exchange that issues short-lived, task-scoped credentials for agents. The approach reduces the risk associated with shared, persistent access tokens.
Vercel introduced eve, an open-source agent framework with built-in capabilities such as durable execution, sandboxed compute, approvals, subagents, and evaluations. It aims to let developers focus on agent behavior while the framework handles production concerns.
Nvidia XR AI provides a reusable foundation for connecting extended reality devices to GPU-accelerated AI services running in the cloud, data center, workstation, or edge. It gives developers access to an open source library for building intelligent agents for AI glasses, AR glasses, and XR headsets. These agents will be able to see what users see, understand token or typed intent, call enterprise tools, and respond within the same XR session. XR AI is now available in public beta.
MolmoMotion, a new motion forecasting model, predicts future 3D point trajectories in videos using language instructions and initial object positions, outperforming existing methods. The model and its dataset, MolmoMotion-1M, feature vast datasets of 3D point trajectories from 1.16M videos, with a benchmark for accuracy testing.
Cursor's new model will push agentic software development far beyond autocomplete and pair programming. The massive model was trained from scratch on over 100,000 GPUs. It has more than 1.5 trillion parameters and delivers intelligence beyond coding tasks. The model will be released in the coming weeks.
Noam Shazeer is leaving Google to join OpenAI. Shazeer was one of the co-authors of the 2017 paper - 'Attention Is All You Need' - that transformed the AI world. He also improved search spelling correction and created the core algorithm for AdSense.
AI agents and analytics platforms need reliable web activity signals. See how HUMAN [17] detects sophisticated automation and bot traffic, helping organizations make better decisions based on trusted human activity
OpenAI introduced LifeSciBench, an expert-judged benchmark that evaluates AI systems on end-to-end life sciences workflows such as evidence analysis, experimental design, scientific reasoning, and research communication rather than isolated biology tasks.
Microsoft Azure's Copilot Migration Agent turns complex migration data into clear answers. Through natural language prompts you can evaluate readiness, risk, and ROI to make confident decisions. Read the playbook [21].
Your bank has the data. But to actually use it — check cash position, reconcile transactions, understand burn — you have to export it, paste it somewhere else, and do the work outside your account.
Z.ai launched GLM-5.2 with a 1 million-token context window, new reasoning controls, and support for long-horizon coding tasks across entire codebases. The company made the model available immediately to Coding Plan users and plans to release API access, chatbot support, technical details, and MIT-licensed open weights the following week. Z.ai positions GLM-5.2 as a coding-first upgrade focused on agentic software engineering, though it did not publish benchmark results at launch.
DeepSeek raised more than $7.4 billion in a funding round that valued the company at more than $50 billion. Its founder, Liang Wenfeng, invested around $3 billion in the fundraise. He previously held nearly 90% of the company before the financing round. A government-backed fund invested around $150 million. DeepSeek plans to use the new capital to advance R&D and expand computing infrastructure.
Android 17 introduced new AI-focused capabilities centered on AppFunctions and Android MCP, enabling apps to expose orchestratable tools that on-device agents can discover and execute. Google positioned the release as part of a broader shift toward an “intelligence system” with deeper AI integration across the platform.
2026.06.17·
android-developers.googleblog.com·8 min read·TLDR AI
The US government spent a long time being complacent about AI regulation. Anthropic's development of advanced models has caused a stir. The administration's response was incomprehensibly strict and risk-averse, especially after its previous attitude towards risk in the industry. It has been unclear about its policies, and disputes about policy and politics between the administration and Anthropic remain unresolved.
OpenAI described Deployment Simulation, a pre-release evaluation method that replays real conversation contexts with candidate models to estimate deployment-time behavior and uncover new safety issues.
OpenAI's Tejal Patwardhan discussed how frontier evaluations are evolving as existing benchmarks become easier for advanced models. The conversation covered benchmark saturation, evaluation design, forecasting model progress, and the challenges of measuring increasingly capable AI systems.
With SiMa.ai's Palette Neat [13], developers build physical AI applications in days/hours - not months. Running your favorite AI model? Building custom, multimodal applications for robotics, automotive, drones, smart vision, or another physical AI project? Use
Cursor Origin is an attempt at an agent-native GitHub competitor. The Git-compatible forge is built around the assumption that lots of AI agents will be cloning, branching, committing, rebasing, reviewing, and fixing failures in parallel. Cursor is trying to own the whole AI software factory. GitHub was built for human-scale development, but Origin was built for agent-scale.
OpenAI's Codex now supports Chrome DevTools Protocol, enabling live browser access to profile JavaScript performance and modify websites in real time. This early-stage feature, available via opt-in settings but excluded from the EEA, UK, and Switzerland, faces performance issues and requires careful prompting. Recent OpenAI moves, such as acquiring Ona, aim to create persistent cloud environments, highlighting a push toward an AI-driven web.
NVIDIA's Blackwell platform dominated MLPerf Training 6.0 benchmarks, achieving the fastest training times and the largest-scale training with 8,192 GPUs. Blackwell's NVLink and NVFP4 innovations enhance performance by efficiently routing MoE models and enabling rapid, large-scale training. Reliability features like the Reliability, Availability, and Serviceability Engine and NVIDIA Resiliency Extension ensure minimal downtime and faster recovery in large-scale AI environments.
Qwen-RobotWorld is a language-conditioned video world model that uses natural language as a unified action interface across robotics, navigation, driving, and other embodied domains.
Anthropic has paused the pricing changes it announced last month, just before they were set to take effect. The change would have treated Claude Agent SDK usage separately from standard Claude usage. Outside SDK usage will now be billed at Anthropic's prevailing API rates. Anthropic is working to update its plans to better support how users build with Claude subscriptions.
Qualcomm is working on over 40 different AI wearable devices. It has announced a platform called Snapdragon Reality Elite for mixed-reality glass and the Scalable Turnkey AI-Ready toolkit, a combination of hardware models and a software stack for AI devices. The company is aiming to position itself as the foundational silicon layer for whatever is coming after the smartphone. It is experimenting with a range of potential form factors that can work with AI technologies.
GPT-Bidi-1 is a bidirectional audio model for ChatGPT's voice mode designed to listen and speak at once, absorb interruptions, and adjust mid-sentence.
Winning teams at PMG's first-ever AI & Tech Sandbox hackathon get a portion of $500K in prize funding, a pathway to see their project deployed by global brands, and keep ownership of the IP. Register now
Wispr Flow [5] is voice for AI. Speak naturally into Claude, ChatGPT, Cursor, or any tool. Flow strips filler, fixes grammar, formats automatically. Detailed prompts in the time it takes to type a summary.
Factory has been building software factories with its customers over the last few months. Its software factories are already in production across the world's largest organizations. Organizations that invest in their autonomous software development will see engineering outcomes surge. Engineers in this era are now responsible for building the factories that build the software. This will see engineering responsibilities grow to span across the business itself.
Sakana Marlin is an autonomous research assistant streamlining extensive strategic analysis in hours. Users input a research topic, and Marlin autonomously generates a detailed strategy report and summary slides without further human intervention. The tool serves various professionals, offering flexible pricing plans and refined features based on feedback from a beta test with around 300 industry experts.
Facebook's new AI Mode transforms the standard search bar into a conversational tool that answers questions by mining public Group discussions, Reels, and Marketplace data. The update aims to increase platform engagement and support Meta's expanding subscription tiers. Critics have raised concerns about data privacy and the accuracy of crowd-sourced AI summaries. The feature is currently rolling out to users in the US.
The US government forcing Anthropic to take down all access to Fable and Mythos seems like a stupid decision. However, it is unknown what motivated the government to make the decision, how much they understand the mechanisms of the technology, whether they demanded or are demanding a narrow fix or a global fix, what they intend to do next, and what they are trying to accomplish. This could just be a terrible misunderstanding that can be sorted out quickly.
Detailing the latest generation of speculative decoding with DFlash and SGLang's Spec V2 engine. Benchmarks showed substantial throughput gains over baseline inference and native MTP speculation.
Coding agents moved the hard part of engineering from writing code to deciding whether to trust it, making review the most leveraged skill in software now that the old happy accident no longer holds. The 2026 data converges: Faros AI's 22,000-developer study found code churn up 861%, per-developer defect rate up from 9% to 54%, review duration up 441%, and zero-review merges up 31%, while GitClear shows 4x raw output for ~12% delivered-value gain, the gap being the review problem in one line.
Fireworks and LangChain developed a cost-effective "perceived error" judge using the Qwen-3.5-35B model, capable of detecting user-identified errors in chatbot interactions. Fine-tuning this judge on chat-langchain data resulted in performance meeting or exceeding frontier models at reduced costs.
AI agents are stateless by default — they forget everything between sessions. Build agent memory architecture on AWS with Amazon Bedrock and tools from across the AI landscape available in AWS Marketplace. Save your spot [16] for the workshop on June 30. Can't wait? Get started with the companion guide [16].
Inference engineering is the discipline of running trained AI models in production efficiently. It involves working with low-level GPU code, model serving frameworks, and the cloud infrastructure that ties them together. Engineers need to optimize for a combination of latency, throughput, cost, and quality, depending on the product they're supporting. Inference engineering is now a broad speciality that any company running serious AI workloads invests in.