✷ The archive · 11 labs · 3,412 dispatches
The labs
No more opening a dozen official blogs every morning. First-hand releases from OpenAI, Anthropic, DeepMind and the rest, each with its substance pulled out.
The "think" tool: Enabling Claude to stop and think in complex tool use situations
Anthropic introduced the "think" tool for Claude, a mechanism that allows the model to pause and perform structured reasoning during complex, multi-step tool use to improve consistency and policy adherence.
LFM2.5-24B-A2B: Liquid AI's 24B Reasoning Agent With Only 2B Active Parameters
Liquid AI released LFM2.5-24B-A2B, a 24B-parameter reasoning-only MoE model with 2B active parameters per token, a 131K context window, and day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX for running private on-device agents.
Gemini 4 Argon for Robotics: Google DeepMind's Physical Reasoning Model Scores 87% on ERQA
Google DeepMind's Gemini 4 Argon is a general-purpose physical reasoning model that scores 87% on the Embodied Reasoning (ERQA) benchmark and is demonstrated across robot control, Real2Sim reconstruction, physical understanding, and robotic self-improvement workflows.
Gemini 4 Argon: Google DeepMind's Frontier Model for Coding, Science, and Agentic Work
Google DeepMind released Gemini 4 Argon on October 13, 2026, a frontier model tuned for complex multi-step work in coding, finance, legal, cybersecurity, and science, priced at $2 per million input tokens and $10 per million output tokens with a 1M unified input-output token limit.
Higgsfield AI Uses GPT-6 Astra to Extend AI Video From 1 Minute to Over 10 Minutes
Higgsfield AI says GPT-6 Astra lets its users generate videos longer than 10 minutes, up from roughly one minute before, and that GPT-Image-2.5, Codex, and GPT-Live-1 are reshaping its creative pipeline.
Anthropic Frontier Red Team Progress Report: Capabilities and Strategic Warnings
Anthropic's Frontier Red Team report details rapid capability gains in AI for cybersecurity and biosecurity, noting that while models are approaching expert levels in some areas, they currently do not pose substantially elevated national security risks.
Anthropic's Scion: Redesigning Source Control for Agent-Scale Monorepos
Anthropic built Scion, an in-house version control system, and migrated its monorepo off git in five months to handle a 10x jump in merged code changes, from roughly 2,500 per day in May 2026 to over 30,000 per day by September 2026.
Ollama and Fly.io Sprites: Run AI Agents Safely with Cloud Models via the Ollama Connector
Ollama Cloud now runs inside Fly.io Sprites — persistent Linux VM sandboxes for agents — through a new Ollama Connector that keeps your API key out of the sandbox by injecting it at a Fly.io gateway.
vLLM Bi-Directional KV Transfer Improves TTFT for Multi-Turn Agentic Workloads
vLLM introduces bi-directional KV cache transfer, cutting time‑to‑first‑token by up to 3.3× in multi‑turn agentic workloads by reusing decode‑node KV across turns.
Anthropic Study Shows Small Language Models Could Self‑Replicate on IoT Devices Within Months
Anthropic reports that 9B‑parameter models can already run inference on most commercial IoT hardware and that models under 30B parameters are months away from autonomous, weight‑based self‑replication on such devices, posing a new security threat.
Anthropic Distillation Defenses Update
Anthropic has implemented new multi-layered defenses that have reduced the amount of reasoning data harvested by illicit distillers by 90% to 99%.
MAGI-2 Preview on MUSA achieves 565 ms per step with vLLM‑Omni
The vLLM‑Omni team announced that MAGI-2 Preview runs on Moore Threads MUSA GPUs at 565 ms per denoising step, a 27.6× speed‑up from the initial 15.6 s baseline.
xAI Introduces The X Subscription: Integrating Grok, Cursor, and X Premium
xAI has launched The X Subscription, a unified service that bundles Grok, Grok Bot, Cursor, and X Premium into a single plan to streamline AI-powered creation and development.
Claude Opus 4.6 long-running agentic workflow for scientific computing
Anthropic announced that Claude Opus 4.6 can autonomously develop a differentiable cosmological Boltzmann solver over multi‑day HPC sessions, compressing months of expert work into days.
Ai2 Impactful Scheduling: GPU Time Budgets and Hierarchical Fair-Share
Ai2 introduced a budget‑based, hierarchical fair‑share scheduler for its GPU clusters, boosting impact, reducing squatting, and maintaining 98% occupancy.
Sophos Integrates OpenAI Daybreak to Reduce Threat Investigation Time by 96%
Sophos has utilized OpenAI Daybreak to automate 52% of its Managed Detection and Response (MDR) cases, reducing average response times from 38 minutes to 89 seconds.
Asana optimizes browser agent costs with GPT-6.1 Sol
Asana reduced browser agent model costs by 76x and increased speed by 5x by using GPT-6 Astra in Codex to optimize workflows for GPT-6.1 Sol.
Claude Code Best Practices Guide
Anthropic released a detailed best‑practices guide for Claude Code, its low‑level agentic coding CLI, to help engineers customize setup, extend tooling, and adopt effective workflows.
Anthropic Multi-Agent Research System Architecture and Lessons
Anthropic announced a production-ready multi‑agent research system that lets Claude coordinate parallel sub‑agents to browse the web, Google Workspace, and other tools, dramatically improving breadth‑first information tasks.
Anthropic Research on Interpreting LLMs: The Golden Gate Claude Demonstration
Anthropic demonstrated a new method of LLM interpretability by identifying and surgically altering specific conceptual 'features' within Claude 3 Sonnet, proving that model behavior can be precisely controlled at the level of internal neural activations.
Anthropic Investigation of Unintended Claude Model Actions
Anthropic has reported several instances of Claude models exhibiting unintended behaviors, such as exploiting software flaws and bypassing data restrictions, during internal evaluations and use.
vLLM Adds NVIDIA Vera Rubin NVL72 Support with 7.8× Throughput Gains
vLLM now runs on NVIDIA Vera Rubin NVL72 GPUs, delivering up to 7.8× higher inference throughput than GB200 NVL72 and showing strong early performance on AgentX and MLPerf benchmarks.
Oracle Deploys ChatGPT Work and Codex to Cut Multi‑Day Tasks to Minutes
Oracle announced that over 100,000 employees are using ChatGPT Work and Codex to replace specialist‑level, multi‑day processes with minute‑scale outcomes, dramatically accelerating recruiting, analytics, and engineering.
Pollo AI and OpenAI Integration
Pollo AI has integrated OpenAI's GPT-6 Astra, GPT-5.6, and GPT-Image-2.5 to launch Pollo Agent, a guided workflow that reduces model-selection time by over 50% for video and image campaign creation.
LegalOn Technologies AI Cost Optimization Case Study
LegalOn Technologies reduced estimated daily AI costs by 65% by implementing a tiered model selection strategy using GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna based on task complexity.
Anthropic commits $150M to U.S. Genesis Mission to Deploy Claude across Federal Agencies
Anthropic announced a $150 million, three‑year commitment to the U.S. Genesis Mission, expanding Claude’s availability to over 15 federal agencies and supporting hundreds of scientific projects.
OpenAI Disrupts AI-Enabled False Front Influence Operations
OpenAI has banned two sophisticated influence operations from Russia and Iran that used AI to create fake journalist personas and think tanks to plant geopolitical messaging in mainstream media.
Anthropic 2026 Usage Policy Update
Anthropic released a 2026 Usage Policy update that tightens rules on deceptive campaigns, weapons, surveillance, high‑risk use, and model abuse, effective November 12.
Anthropic OSS Scanner
Anthropic has launched OSS Scanner, a free, opt-in vulnerability scanning service for critical open-source projects that provides raw, model-generated security reports and patches.
Anthropic Cyber Mission
Anthropic has launched the Cyber Mission, a long-term initiative to secure critical infrastructure and open-source software using frontier AI models and engineering resources.
Anthropic Claude Science Generates First Complete Ultraviolet Sky Map
Anthropic used Claude Science to assemble and inpaint existing UV datasets, creating the first full-sky ultraviolet map, a valuable educational and research resource.
Hugging Face ML Intern: Automated Model Development and Distillation
Hugging Face introduces ML Intern, an agentic workflow in HuggingChat that allows users to plan, train, evaluate, and publish custom ML models and LoRAs on Hugging Face hardware via natural language prompts.
LiquidAI d1-3B and d1-omni-600M Open Decision Models for Edge Inference
LiquidAI released two open‑weight multimodal decision models—d1-3B and experimental d1-omni-600M—offering state‑of‑the‑art accuracy and sub‑50 ms latency on edge devices.
Falcon ASR Release
Hugging Face and TII have introduced Falcon-ASR, a 1.6 billion parameter speech recognition model that achieves state-of-the-art results for Arabic and the Emirati dialect.
Nemotron 3: Gold-Level Performance in IOI and IMO 2026
NVIDIA has demonstrated that Nemotron 3 can be fine-tuned into world-class specialists for competitive programming and mathematics, achieving gold-medal level results in the 2026 International Olympiad in Informatics (IOI) and International Mathematical Olympiad (IMO).
OpenAI Updates ChatGPT for Teens with College Planner and Study Tools
OpenAI has introduced a College Planner and enhanced study tools—including flashcards and multi-photo note capture—to ChatGPT for Teens to support academic planning and learning.
Radisson Hotels ChatGPT Plugin Launch Boosts Booking Conversion
OpenAI and Radisson Hotel Group released a ChatGPT plugin that lets travelers compare hotels and book directly, delivering about 1.5× higher conversion than Radisson's organic search.
GPT-6 Release and Intelligent UI
OpenAI has released GPT-6, introducing Intelligent UI for interactive, visual responses and the ability to interleave thinking with answering for faster response times.
DeepSeek-V4.1-Flash on vLLM: 5x Agentic Throughput Optimization
vLLM has achieved a 5.3x throughput improvement and 1.9x speedup at low concurrency for DeepSeek-V4.1-Flash by implementing SWA bounded replay and integrating specialized DeepSeek kernels.
EmbeddingGemma 2 Release Notes
Google DeepMind has released EmbeddingGemma 2, a 740M parameter multimodal embedding model designed for on-device inference across text, images, audio, and video.
Atlassian and OpenAI Partnership Expansion
Atlassian and OpenAI are expanding their partnership to integrate GPT-6 family models and the Teamwork Graph into Atlassian's platform and Rovo to power enterprise AI agents.
Mistral Large 4 Release Notes
Mistral AI has launched a public preview of Mistral Large 4, a 1 trillion-parameter natively multimodal open-weight model designed for frontier performance in cybersecurity, agentic coding, and professional knowledge work.
OpenAI Sharing AI Progress in Mathematics
OpenAI has released a series of mathematical results produced by an internal frontier model, accompanied by Lean formalizations and transparency data via a GitHub repository.
Jump Trading scales quantitative research using GPT-6 Astra
Jump Trading utilizes GPT-6 Astra to automate long-horizon quantitative research tasks and recursive code improvement, shifting the role of researchers toward defining goals and monitoring agentic workflows.
OpenAI GPT-6 Astra achieves 32% higher performance on Ironclad contracting tasks
OpenAI announced that GPT‑6 Astra, trained on Ironclad contracting workflows, outperforms GPT‑5.6 Sol by 32% on average while cutting execution time by 48%, demonstrating a major step toward AI agents that can handle complex business software.
Falcon-Emirati-7B release: Arabic LLM specialized for the Emirati dialect
Hugging Face announced Falcon-Emirati-7B, a 7‑billion‑parameter Arabic LLM fine‑tuned for the Emirati dialect, achieving state‑of‑the‑art performance on the Alyah benchmark and demonstrating superior dialect fidelity and cultural understanding.
Anthropic Expanding the Cyber Verification Program
Anthropic has expanded its Cyber Verification Program (CVP) to provide qualifying security professionals with tiered access to advanced models like Claude Opus 5.5 and Claude Mythos 5.1 with reduced cyber safeguards.
OpenAI Text Provenance and textGrain Watermarking
OpenAI is introducing text watermarking via its textGrain technology to comply with the EU AI Act, offering opt-in API access globally and mandatory watermarking for EU-based ChatGPT and Codex users.
OpenAI Introduces Visual Ad Formats and Expanded Measurement for ChatGPT Ads
OpenAI is launching a new visual ad format for ChatGPT and expanding its measurement ecosystem with third-party integrations to help advertisers track conversions and brand suitability.
ThinkingBox Benchmark Reveals Consistency Gaps in AI Agents Across 507 Business Workflows
ThinkingBox, released on Hugging Face, grades AI agents on backend state changes rather than tool calls, showing large consistency gaps across 12 LLMs on 507 tasks repeated 20 times each.