AI Daily · 2026-10-14

Model releases were led by Gemini 4 Argon, now available via API with llm-gemini 0.35 support and priced initially in line with GPT-6 Sol before a December price hike, and LiquidAI's LFM2.5-24B-A2B, a 24B-total/2B-active on-device reasoning agent with explicit chain-of-thought and a 131K context window. LangChain shipped a wave of agent-infrastructure updates around MCP and Deep Agents, including a rewritten MCP adapter, Deep Agents deploy, Open SWE, and self-testing documentation. Modal surpassed $1B in annualized run rate and raised a $750M Series D, while also exploring mixed-precision, mixed-accelerator RL samplers to cut the sampling workload that consumes 80-85% of GPUs in large RL runs. On the policy and ecosystem front, Anthropic opened a Singapore office with a government MOU, NVIDIA committed $250M to UK agentic safety research, and OpenAI reported fast but uneven EU agentic adoption; Anthropic also replaced Git with its own Scion system, and Dyna Robotics launched Dyna-2/2.1 physical agents.

North America · First-hand

Anthropic

⭐⭐ [Other] Anthropic opens Singapore office and signs MOU with the Government of Singapore

Anthropic opened a new office in Singapore—its fifth in Asia-Pacific—and signed a Memorandum of Understanding with Singapore's Ministry of Digital Development and Information, with the signing held on the sidelines of Singapore International Cyber Week by Permanent Secretary Kai Fong Chng and Anthropic co-founder Ben Mann. The MOU sets out cooperation in AI safety, security and model evaluation; research on AI adoption and its impact; and responsible adoption across the public and private sectors. Anthropic said it plans technical exchanges with Singapore's AI safety, security and assurance institutions on model evaluation, testing standards and incident response, and intends to share research and data including the Anthropic Economic Index, help SMEs adopt AI responsibly, and support building AI capability among public officers. The post says Anthropic's latest Economic Index ranks Singapore second of 121 countries in Claude.ai usage relative to population, at nearly six times what its size alone would predict, and that more than 1,500 GovTech Singapore engineers and technical staff have adopted Claude Code while the Cyber Security Agency of Singapore is exploring how Claude could strengthen cyber defense. Anthropic's work with Singapore dates to 2023, when it joined the Generative AI Evaluation Sandbox run by the Infocomm Media Development Authority and the AI Verify Foundation, followed by a red-teaming project in English, Mandarin, Malay and Tamil.

OpenAI

⭐⭐ [Research] The EU’s next chapter in AI is agentic

OpenAI Economic Research published a report on the growth of agentic AI use among EU enterprises and small and medium businesses. Among ChatGPT Enterprise customers, agentic use already represents a substantial share of activity: in the EU sample, Codex and ChatGPT Work together account for more than half of output tokens. For small and medium businesses, weekly active users of Codex and ChatGPT Work in the EU grew more than eightfold since May 2026, with growth in every member state. Adoption is far from uniform, however: the share of output tokens generated through Codex and ChatGPT Work ranged from 22% in Slovenia to 85% in Lithuania, and the kinds of work people bring to ChatGPT also vary by country. Among frontier firms (the top 10% by output tokens per active user), EU frontier firms generated more than four times as many output tokens per active user as typical EU firms in September, yet frontier firms outside the EU generated more than twice as many as frontier EU firms, and EU-based firms made up only 10% of frontier firms globally. The report concludes that closing the gap requires helping businesses move beyond experimentation, giving workers the skills and confidence to use AI, and making advanced capabilities more accessible and affordable; OpenAI CFO Sarah Friar said Europe is in a new era of AI but that there is still plenty of room for use to grow.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐⭐⭐ [Model Release] llm-gemini 0.35

Simon Willison released llm-gemini 0.35, adding support for the newly launched Gemini 4 Argon model to his LLM plugin for access to Google's Gemini family; the model is available via the Gemini API and to $100/month subscription customers. At launch Argon is priced at $2 per million input tokens and $10 per million output tokens — the same as GPT-6 Sol, half of Claude Opus 5.5, and one-fifth of Claude Fable 5.1 and GPT-6 Astra — and is set to double in price on December 1st. Artificial Analysis benchmark-cost data shows Argon costs $2,407 to run the benchmark, close to GLM-5.3 and significantly cheaper than the latest OpenAI, Anthropic, and Meta models. Willison also shared example outputs generated with the model, including SVG pelicans.

⭐⭐⭐ [Model Release] Run fast, private agents on-device with LFM2.5-24B-A2B

LiquidAI released LFM2.5-24B-A2B, a 24B-parameter reasoning model with only 2B active parameters, built for fast, private agents on-device. It thinks before answering, calls tools, and works through multi-step tasks inside popular agent harnesses. On agentic and tool-use tasks it is competitive with Qwen3.6-35B-A3B, which has roughly 50% more total and active parameters, scoring 48.09 versus 38.19 on LiveClaw; it reaches 184 tok/s on an Apple M5 Max and up to 21K output tok/s on a single NVIDIA B300 GPU. The model keeps the LFM2-24B-A2B architecture (40 layers, 64 experts per MoE block with top-4 routing, a hybrid of gated short convolutions and GQA) but grows the context window from 32K to 131K tokens, doubles the vocabulary from 65K to 128K via an in-place tokenizer expansion, and is reasoning-only, producing an explicit chain of thought. Pre-training was scaled from about 17T to about 33T tokens, followed by mid-training with 2T tokens at 32K context and 300B tokens at 131K context, and a five-stage post-training pipeline: SFT, teacher specialization, non-agentic and agentic multi-domain on-policy distillation (MOPD), and agentic RL.

⭐⭐ [Model Release] Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics

Dyna Robotics released Dyna-2, a world-action model it says was pre-trained on over one million hours of human video, with the company claiming new scaling laws on human data across four orders of magnitude. The company then followed up with Dyna-2.1, described as the first physical agent to achieve reliable super long-horizon whole-body autonomy, pairing new semi-humanoid hardware with an agentic system built around Dyna-2. In the podcast, Dyna says the model held a 99.4% success rate folding restaurant napkins for 24 hours straight, including recovering after pulling over a whole stack, that robots can learn new tasks from as little as 45 minutes of demonstrations, and that they are deployed at Din Tai Fung folding thousands of napkins daily. Dyna co-founder and Chief Scientist Jason Ma discusses world-action models versus vision-language-action models, diffusion transformers and flow matching, human-to-robot transfer, why Dyna builds its own hardware, and why dexterous hands and missing sensory data remain major bottlenecks.

Product Update

⭐⭐⭐ [Product Update] MCP in LangChain: Stateless Protocol, Elicitation, and More!

LangChain has reworked its MCP (Model Context Protocol) support to match the protocol's largest spec rewrite since launch. Three things changed: MCP support moved from the separate langchain-mcp-adapters into the main package under langchain.mcp, installed via pip install "langchain[mcp]"; it is now built on FastMCP, which handles transports, auth, connection management and protocol negotiation per connection so both old- and new-spec servers work; and elicitation via LangGraph interrupts plus client-side tool-list caching are now supported. The new stateless core removes session pinning, sticky routing and shared session stores, lets a server declare how long its tool list stays fresh, and lets a tool pause mid-call to ask the caller something. The former MultiServerMCPClient collapses into a single MCPAdapter class, with a migration guide provided.

⭐⭐ [Product Update] Deep Agents Deploy: an open alternative to Claude Managed Agents

LangChain launched Deep Agents deploy in beta, describing it as the fastest way to deploy a model-agnostic, open source agent harness in a production-ready way. It is built on Deep Agents, an open source, model-agnostic harness, with the argument that choosing an open harness means owning your memory rather than having it locked into a proprietary harness or a single model. The steps previously needed for production — deploying agent orchestration logic and memory in a multi-tenant, scalable way; spinning up sandboxes per agent session; and standing up MCP, A2A, human-in-the-loop and memory endpoints — are now bundled into one command, deepagents deploy. Deployment parameters include model, AGENTS.md, skills, mcp.json and sandbox; supported model providers include OpenAI, Google, Anthropic, Azure, Bedrock, Fireworks, Baseten, OpenRouter and Ollama, and sandbox integrations include Daytona, Runloop, Modal and LangSmith Sandboxes. Under the hood it bundles a LangSmith Deployment server that is production-ready and horizontally scalable, exposing 30+ endpoints for MCP, A2A, Agent Protocol, human-in-the-loop and memory, while leaning on open standards such as MIT-licensed deepagents (Python and TypeScript), AGENTS.md, Agent Skills and MCP/A2A/Agent Protocol, plus self-hosted LangSmith Deployments for owning your own memory. The post then begins comparing it with the recently launched Claude Managed Agents, noting the same high-level architecture (harness, agent server, sandboxes), before the text cuts off.

⭐⭐ [Product Update] Context Management for Deep Agents

The LangChain team describes how its open source agent harness, the Deep Agents SDK, manages context. The SDK lets agents plan, spawn subagents, and use a filesystem for complex long-running tasks, and because such tasks often exceed model context windows it implements context compression. Compression uses three techniques triggered at different frequencies: tool responses over 20,000 tokens are offloaded to the filesystem and replaced with a file path reference plus a preview of the first 10 lines; once session context crosses 85% of the model's available window, older write/edit tool-call arguments are truncated and replaced with a pointer to the file on disk; when nothing more can be offloaded, the SDK falls back to summarization, where an LLM generates a structured summary of session intent, artifacts created, and next steps that replaces the full conversation history, while the original messages are written to the filesystem as a canonical record. The SDK uses LangChain model profiles to obtain a given model's token threshold and triggers these steps at threshold fractions of the context window; the post also notes that real-world runs on benchmarks such as terminal-bench can trigger context compression.

⭐⭐ [Product Update] Open SWE: An Open-Source Framework for Internal Coding Agents

LangChain released Open SWE, an open-source framework for building internal coding agents, built on Deep Agents and LangGraph. It captures architectural patterns that converged across internal coding agents built independently at Stripe (Minions), Ramp (Inspect), and Coinbase (Cloudbot). Those patterns include isolated cloud sandboxes for execution, curated toolsets, Slack-first invocation, pulling full context from Linear, Slack, or GitHub at startup, and subagent orchestration. Rather than forking an existing agent, Open SWE composes on Deep Agents, which provides an upgrade path and lets organizations keep org-specific tools, prompts, and workflows as configuration.

Research

⭐⭐⭐ [Research] Toward RL on mixed-precision, mixed-accelerator samplers

This Modal blog post asks whether RL samplers for large models can run at different precisions and on different accelerators than the trainer. It notes that sampling already consumes 80–85% of the GPUs in large RL runs, while serving systems are free to use low-precision formats such as FP8, MXFP4 and NVFP4, co-designed kernels, and specialized accelerators. Two problems stand in the way: a learning problem, since each optimization adds training–inference mismatch and makes samples more off-policy, and a system problem, since RL infrastructure is usually built around homogeneous accelerator clusters. The authors combine off-policy correction (including masked importance sampling) with Modal's autoscaling serverless GPUs to train Qwen3.6-35B-A3B fully asynchronously on agentic coding tasks. In the setup, mixed samplers span six kinds of accelerators (B200/B300 with NVFP4, H100/H200 with FP8 and BF16, A100 and RTX PRO 6000 with BF16, all keeping KV cache in FP8), while a Megatron-LM trainer updates the policy in BF16 on B200 using the Miles framework and SGLang samplers, with evaluation on SWE-bench Pro V2 and its HARD-51 subset.

⭐⭐ [Research] Improving Deep Agents with harness engineering

LangChain's coding agent (deepagents-cli) improved from 52.8 to 66.5 on Terminal Bench 2.0, a 13.7-point gain that moved it from outside the Top 30 to the Top 5. They changed only the harness—system prompt, tools, and middleware—while keeping the model fixed at GPT-5.2-Codex. The experiments used Terminal Bench 2.0's 89 tasks (spanning domains like machine learning, debugging, and biology), orchestrated with Harbor for verification and scoring, with every agent action stored in LangSmith. They turned trace analysis into a reusable Agent Skill that fetches traces from LangSmith, spawns parallel error-analysis agents, and has a main agent synthesize findings and propose targeted harness changes. The improvements described focus on build & self-verify and planning & discovery.

Other

⭐⭐⭐ [Other] Modal crosses $1B in Annualized Run Rate and raises $750M Series D

Modal announced it has crossed $1B in annualized run rate and raised a $750M Series D. The company said usage has tripled since its Series C five months ago and grown more than 15x year-over-year. It listed customers building on the platform, including TypeSafe AI's Jev (which served 1T tokens on Modal in its first three days), DoorDash, Cognition, Ramp, Linear, Hebbia, Runway, Physical Intelligence, Generalist AI, 1X, Chai Discovery, and Achira. Modal said it built its own file system, container runtime, scheduler, and networking on a capacity pool spanning 20+ clouds, and that its Sandboxes product is a $300M run rate business with close to 3 billion sandboxes spun up and scaling to a million concurrent. Recent launches cited include multi-node training, Modal servers, Endpoint Candidates, and generally available virtual machines.

⭐⭐ [Other] From Enterprise to Earthshot: UK’s Ecosystem Moves From Ambition to Application With NVIDIA

NVIDIA published a blog post describing how the U.K. ecosystem is moving from ambition to application, saying industries across the country are putting its full stack to work — from enterprises transforming their businesses to startups tackling hard problems — with deployment scaling quickly. The post mentions a five-year, $250 million compute commitment to strengthen the U.K.'s agentic safety. The visible source text is truncated at that point.

⭐⭐ [Other] NVIDIA Commits $250 Million to Advance UK Agentic Safety and Security Research Over the Next Five Years

NVIDIA announced a five-year compute donation valued at $250 million aimed at strengthening the United Kingdom's agentic safety and security research capability.

⭐⭐ [Other] How We Made Our Docs Test Themselves

A LangChain blog post describes how the company made its documentation code samples test themselves. Inline code snippets are static and go stale as dependencies and APIs change, and while making samples testable (running them in CI and failing the build on breakage) solves this, the upfront setup work often discourages teams from doing it. LangChain delegated the migration to the Deep Agents CLI combined with Skills, reusable instructions the agent loads when a task matches: move code into standalone files under src/code-samples/{product}, add setup and teardown, lint, add :snippet-start:/:snippet-end: (and :remove-start:/:remove-end:) markup, run the samples as tests, and include the generated snippets in the docs. No coding is required, and anyone in the docs repo can open the CLI and ask the agent to migrate a file such as streaming.mdx to testable samples. A companion GitHub Action runs the tests regularly and files tickets when a test fails, with the skill stored at .agents/skills/docs-code-samples/SKILL.md.

⭐⭐ [Other] Redesigning source control at Anthropic

This is a link post by Simon Willison discussing Anthropic's decision to build its own alternative to Git for source control. It cites Anthropic's figures that merged code changes grew from roughly 2,500 per day in May 2026 to over 30,000 per day by September. The new source control system is called Scion and the code review system is called Antique; requirements were extreme, with a main branch history of almost 100 GB, a compressed working copy over 25 GB, and about 25 million code checkouts per day for CI. The system was built by a team of seven and the migration from Git was completed in five months; an engineer said offline support wasn't important because Claude is by far the heaviest user and is necessarily online. The post also notes that Anthropic makes extensive use of adversarial review, which can be myopic, requiring engineers to steer Claude away from over-engineering.