Gemini 4 Argon available to everyone, SpaceXSI brings Grok and Cursor together in The X Subscription, NVIDIA releases Run:ai as open source

Two weeks after reserving it for cyber defenders in the Fairwind Program, Google is opening Gemini 4 Argon to everyone: the Gemini API charges 2 dollars per million input tokens until November 30, the Gemini app and AI Mode offer it to Google AI Pro and Ultra subscribers, and Cursor, Devin, GitHub Copilot and Manus are adding it the same day. SpaceXSI brings Grok, Grok Bot, Cursor and X Premium together in The X Subscription, a single subscription ranging from 0 to 200 dollars per month, and NVIDIA is releasing its entire Run:ai GPU orchestration platform as open source. Liquid AI is releasing LFM2.5-24B-A2B, a reasoning model with 2 billion active parameters for local agents, and NVIDIA is detailing the software ready for Friday’s launch of RTX Spark PCs.

Gemini 4 Argon available to everyone: Gemini API, Gemini app, AI Mode and Antigravity

October 13 — Announced on September 30 and previously reserved for cyber defenders in the Fairwind Program, Gemini 4 Argon is rolling out worldwide. Google presents its frontier model as a model for in-depth work (deep work) in programming, finance, law, cybersecurity and science. Here is who can access it, and how.

Target audience	Access to Gemini 4 Argon as of October 13
Developers	Paid Gemini API, in Google AI Studio and Android Studio
Google AI Ultra subscribers	Antigravity and AI Studio
Businesses	Gemini for business, Gemini API, Antigravity
General public	Gemini app and search AI Mode, for Google AI Pro and Ultra subscribers
Scientists	Highest Antigravity usage tier free for 6 months, by application (10 000 places)
Cyber defenders	Fairwind Program expanded to hundreds of partners, Argon without cyber safeguards

For developers, the identifier is gemini-4-argon, in a stable version. The launch price remains the one announced in late September: 2 dollars per million input tokens and 10 dollars for output, with a 95 % discount on cached input, until November 30, 2026; after that, 4 and 20 dollars. The model accepts text, images, video, audio and PDFs, responds in text, and shares a single window of 1 048 576 tokens across input, thinking and output. A request returns at most 64K tokens with streaming and 32K without streaming (65 536 on Gemini 3.x), but long decoding (long decode) can produce up to one million tokens minus the input: recent SDKs (google-genai 2.30.1, @google/genai 2.29.1) chain the requests automatically. Thinking can be set to three levels (low, medium by default, high), and explicit caching is not supported.

As for results, Google ranks Argon first on the Vals Index and claims 87 % on ERQA, a spatial reasoning benchmark for robotics, ahead of all other frontier models. Among early-access customers, Box measures 14 % higher accuracy than Gemini 3.1 Pro. Google also promises the rest of the Gemini 4 family, without a date.

🔗 The Gemini 4 Argon launch post · What’s new with Gemini 4 Argon in the Gemini API

Argon’s security: 12 out of 12 simulated cyberattacks solved, a misalignment monitor

October 13 — Google is publishing a comprehensive safety report (Safety and Security Report) alongside the model, starting with its cyber capabilities.

The same coding and reasoning improvements that enabled Argon to find and patch vulnerabilities also drove better performance on our cybersecurity tests — solving all 12 of the end-to-end simulated cyberattack scenarios, up from five in our previous generation. — Anca Dragan, VP AI Safety & Behavior, Google

Because its versions without safeguards are also significantly better at troubleshooting biology lab protocols, Google is activating three layers of real-time protection in these two domains: request filtering by small models, activation probes (activation probes) that read the model’s internal signals, and monitoring that can stop the agent mid-response. A misalignment monitor also reads its actions and chain of thought, which Google refuses to penalize: in a trial where a penalty had been applied to it by mistake, the proportion of shortcuts taken silently rose from 11,2 % to 36,4 %. Against prompt injection, the attack success rate drops to 0,7 % on Gray Swan IPI.

🔗 Developing Gemini 4 Argon responsibly and safely

Antigravity: Argon for Ultra subscribers, following an internal trial with more than 100 000 users at Alphabet

October 13 — Argon is coming to the desktop app and Antigravity CLI for Google AI Ultra subscribers and eligible Gemini Enterprise subscriptions; Antigravity must be updated to its latest version. The team describes the internal trial: more than 100 000 people at Alphabet use Argon each month in the internal version of the tool, and over four weeks Google measured 35 % more agent-generated code and 30 % fewer post-submission rollbacks.

Examples cited include halving the CPU footprint of the YouTube comments frontend, closing 97 obsolete or duplicate tickets in less than three hours from a queue of 134, rewriting a Fortran codebase in Rust to run 1,8 times faster, and an access control audit that uncovered four critical vulnerabilities missed by two previous leading models. Usage extends beyond engineering: more than half of Cloud sales teams and nearly half of Google’s finance organization use Antigravity.

🔗 Gemini 4 Argon in Google Antigravity

October 13 — The Gemini app’s release notes include two new features alongside the Argon rollout. Antigravity is integrated into the Gemini app for macOS: a button launches a coding agent, users approve its plan and return to the chat while it builds and tests the code in the background, with a notification when the result is ready. The integration is available in English to eligible accounts aged 18 and over.

A new Customize menu brings together response settings, app discovery and adding skills, and Gemini now connects to more than 40 apps or services, such as Instacart, OpenTable and Adobe. Customize is rolling out across all subscription tiers, on the web, mobile, Windows and macOS, but most connected apps remain available in English and in the United States. Google also discusses a shift from chat to tasks (tasks), allowing Gemini to keep working while the user is away.

🔗 Gemini app release notes · Gemini 4 Argon in the Gemini app

Gemini 4 Argon in third-party tools: Cursor, Devin, GitHub Copilot and Manus

October 13 — Four tools are opening access to Argon on the same day, each with its own terms and measurement.

Third-party tool	Where to use it	Listed price	Claimed result
Cursor	Cursor editor	2 dollars for input, 0,10 dollar for cache reads, 10 dollars for output per million tokens	52,8 % on CursorBench 4.0 at default effort
Devin	Devin Desktop and Devin CLI	75 % below list price until October 27	60,5 % on FrontierCode 1.1
GitHub Copilot	Pro+, Max, Business and Enterprise plans, eight surfaces	2 dollars for input, 0,10 dollar for cache, 10 dollars for output until December 1	More tasks solved than previous Gemini models, no figure
Manus	Manus, with Google Cloud	Not specified	No figure

According to Cursor, Argon is on par with Opus 5.5. The CursorBench 4.0 table specifies that this parity applies at medium effort (Opus 5.5 Medium at 52,5 %, for 2,91 dollars per task versus 4,39 dollars for Argon): Opus 5.5 remains ahead in High (56 %) and Max (57,8 %, leading the ranking), and Argon consumes nearly three times as many tokens per task. At Devin, the discount combines Google’s 50 % launch discount, valid until December 1, with Devin’s own 50 % discount; on FrontierCode 1.1, Cognition’s in-house benchmark, Argon edges ahead of GPT-6.1 Sol (60,4 %) but remains behind Claude Opus 5.5 (65,3 %) on the chart, titled 1.1 Extended.

GitHub is making it generally available for Pro+, Max, Business and Enterprise (Pro and Free are not listed), with a gradual rollout in the model picker across eight surfaces, from VS Code and Copilot CLI to Xcode and Eclipse; in Business and Enterprise, it is enabled by default unless the administrator has disabled the default activation of new models. Manus announces that it is live with Google Cloud, for coding, planning and 3D modeling, without specifying plans, prices or credits.

🔗 Cursor’s announcement · CursorBench 4.0 · Devin’s post · GitHub changelog · Manus’s announcement

The X Subscription: SpaceXSI brings together Grok, Grok Bot, Cursor and X Premium, from 0 to 200 dollars per month

October 12 — SpaceXSI (formerly SpaceXAI) is launching The X Subscription, a single subscription bringing together Grok, Grok Bot, Cursor and X Premium, and Cursor announced it that same evening: the editor, which joined SpaceX in August, is now sold in the same plan as the Grok assistant and the benefits of the X social network. The offer has been available to new customers since October 12 and is due to become available to everyone worldwide in the coming days.

Subscription plan	Monthly price	Included usage	What the plan unlocks in Cursor
Free	0 dollars	Limited	Starter allowance for coding agents
Plus	8 dollars	Base	Grok models only
Premium	30 dollars	4 times Plus	Cloud agents, automations, Origin, AI code reviews, third-party models at API prices
Super	100 dollars	15 times Plus	Same as Premium, with more usage
Ultra	200 dollars	50 times Plus	Same as Premium, with more usage

Premium includes the benefits of SuperGrok, Cursor Pro and X Premium+, which previously cost a combined 90 dollars per month according to SpaceXSI. At the heart of the offer is a shared usage pool for Grok and Cursor, replenished each week; once it is exhausted, users buy extra usage (top-ups) or wait for the next replenishment. SpaceXSI promises at least twice as much coding usage with its own models as in the comparable Cursor plan. Third-party models, from other labs, remain available in Cursor starting with Premium, but are excluded from the pool: they are billed at the API list price through top-ups, with adjustable spending caps.

Cursor is now part of The X Subscription. Current Cursor subscribers can keep their plan or switch to The X Subscription. — @cursor_ai on X

The old plans are no longer offered to new subscribers. Current SuperGrok, Cursor or X Premium subscribers keep their plan, price and features, or switch with a prorated refund; at Cursor, included usage of third-party models continues until the switch. SpaceXSI warns, however, that it will eventually move remaining customers to the new plans, without a timetable and with advance notice. For teams, Premium, Super and Ultra retain the same per-seat price as individual subscriptions, cheaper than the old Cursor Teams plans according to SpaceXSI, with three times as much coding usage, SSO, consolidated billing and SOC 2 compliance.

🔗 Introducing The X Subscription · Thread from @SpaceXSI

NVIDIA open sources the entire Run:ai platform

October 13 — NVIDIA is open sourcing all of NVIDIA Run:ai, its platform that dynamically allocates GPUs and orchestrates AI workloads on Kubernetes, from data processing to training and inference. The code is hosted in the NVIDIA DSX AI Factory GitHub organization, and the platform is also available on NGC as a community version (Community Product), with a Helm chart that installs the integrated stack. This version requires no paid Run:ai subscription but does not include enterprise support, and downloading it from NGC requires accepting NVIDIA’s software license. Commercial offerings continue: the standalone product, NVIDIA Mission Control, NVIDIA AI Enterprise, and cloud marketplaces.

The release consists of four components that can be adopted separately:

Released component	Role in the platform
KAI Scheduler	Kubernetes-native scheduler (gang scheduling, fair sharing, topology-aware placement), already open source under Apache 2.0 and a CNCF Sandbox project
GPU Fractioning Operator	Native CUDA 13.4 GPU fractions, with a strict memory limit per workload; Run:ai’s proprietary implementation is removed
KAI Resource Management	Projects, departments, quotas, and node groups
Workload Map	Go library that describes each workload type (PyTorchJob, RayCluster, KServe, Dynamo graphs) so tools no longer need to code for each CRD

Eight partners in the DSX ecosystem are already integrating it: Crusoe, Mirantis (k0rdent AI), Nebius (one-click deployment), OpenNebula Systems, Rafay (fully managed service), Saturn Cloud (Token Factory), Spectro Cloud (PaletteAI), and vCluster. NVIDIA also highlights two related open projects: Model Streamer, which loads weights directly into GPU memory to reduce cold starts for inference replicas (integrated with vLLM and SGLang, and 3 to 6 times faster than the default loader on eight Microsoft Azure H100 VMs), and Grove, Dynamo’s Kubernetes API for workloads with multiple components. The stated goal is to make Run:ai an open foundation for the DSX AI Factory platform, so every partner and every AI cloud gets more from its GPUs.

🔗 Run:ai goes open source (NVIDIA technical blog)

LFM2.5-24B-A2B: Liquid AI releases a reasoning model with 2 billion active parameters for local agents

October 13 — Liquid AI is releasing LFM2.5-24B-A2B, a mixture of experts (MoE) with 24 billion parameters, only 2 billion of which are activated per token. According to Liquid AI, it is the largest model in the LFM2.5 family, which now ranges from 230 million to 24 billion parameters, and a pure reasoning model: it always produces a chain of thought before its answer. The goal is a local agent: with each reasoning token costing little, an agent can run continuously on the user’s machine, call tools, and carry out successive tasks.

The architecture follows that of LFM2-24B-A2B (40 layers, 64 experts per block with top-4 routing, a hybrid of gated short convolutions and GQA attention). What changes: the context increases from 32,000 to 131,000 tokens, the vocabulary doubles from 65,000 to 128,000 tokens, and pretraining increases from around 17,000 to 33,000 billion tokens, before five-stage post-training that ends with reinforcement learning conducted in agent harnesses (OpenClaw, Hermes Agent, Pi, OpenCode).

Benchmark evaluated	LFM2.5-24B-A2B	Qwen3.6-35B-A3B	Gemma-4-26B-A4B-it	Nemotron-3.5-Lightning
AutomationBench	27,69	8,72	5,46	2,61
BFCLv4	65,01	65,00	57,09	53,78
ToolSandbox	80,71	75,02	72,09	75,13
LiveClaw	48,09	38,19	41,25	37,72
PinchBench	83,18	89,93	77,12	84,20
WebArena Verified	26,75	37,62	25,62	28,69
IFBench	74,69	64,21	47,44	71,67
GPQA Diamond	74,42	83,69	74,90	72,42
MATH-500	96,04	96,73	95,66	96,40
SWE-Bench Verified (OpenCode)	51,00	66,20	25,80	53,40

Against three larger MoEs, the model holds its own on tool use and instruction following, but falls behind on knowledge and agentic coding; Liquid AI recommends pairing it with search or a knowledge base for tasks that require knowledge. Speed is the other selling point: 184 tokens/s on an Apple M5 Max, 114 on a Ryzen AI Max+ 395, and more than 21,000 output tokens per second at high concurrency on a single B300. The base and post-trained models are released with open weights under the custom lfm1.0 license, with immediate support in llama.cpp, MLX, vLLM, SGLang, and ONNX.

🔗 LFM2.5-24B-A2B (Liquid AI’s Hugging Face blog)

Local AI on Windows PCs: NVIDIA prepares RTX Spark with Nemotron 3.5 Bolt 74B, Hugging Face launches Llama Windows

October 13 — Three days before RTX Spark laptops launch on Friday, October 16, following the opening of preorders on the 7th, NVIDIA details the software ready for the launch of this chip, which combines up to 128 GB of unified memory and 1 petaflop of AI compute. The centerpiece is a new open model, Nemotron 3.5 Bolt 74B-A9B 24GB: a MoE with 74 billion parameters, 9 billion of them active, quantized to 2 bits with quantization-aware training to fit in 24 GB, and designed for local agents. It launches on Thursday, October 15 in llama.cpp, vLLM, Ollama, Unsloth, and LM Studio, and will be callable from Hermes and OpenClaw; training data, weights, and recipes are open.

On the agent side, NVIDIA puts local LLM downloads at 5 times the total for all of last year. GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi, and WorkBuddy recommend and configure a suitable local model; security relies on the OpenShell sandbox and Microsoft’s MXC containers, and the open source Switchyard library routes requests between local and cloud models. NVIDIA PAIR, the router that distributes inference across PCs on a local network, leaves beta with support for llama.cpp and direct model downloads from Hugging Face. For creative work, Adobe announces up to twice the speed in Photoshop and Premiere (optimizations in beta), and NVIDIA Broadcast 2.3 adds video super-resolution, frame generation, and an MCP server so agents can control its effects. For developers, Windows on Arm gets its first PyTorch-CUDA build and native Python 3.15 by default.

Announced item	Availability date
NVIDIA PAIR out of beta, Broadcast 2.3, Llama Windows	October 13
Nemotron 3.5 Bolt 74B-A9B 24GB	Thursday, October 15
RTX Spark laptops, practical AI guides (playbooks), DLSS 5 model twice as fast	Friday, October 16
Nemotron 3.5 Bolt E4B 4GB for ACE characters, RTX Remix 1.6	Friday, October 23
DLSS Ray Reconstruction in Blender	November

🔗 Software for the RTX Spark launch (NVIDIA blog)

Llama Windows: Hugging Face puts llama.cpp in the Windows 11 system tray

October 13 — Hugging Face launches Llama Windows, a free, open source WinUI 3 application that installs and runs llama.cpp from the Windows 11 system tray, without a terminal, compilation, or administrator privileges. Users choose a quantized model on the Hub, download it with one click, and load it from a compact panel; at startup, the application connects to an already running llama.cpp server or installs llama.cpp using its official installer. It is the Windows counterpart to llama.app on macOS.

For developers, the local server listens on http://localhost:9931 with an OpenAI-compatible API, and downloads go into the standard Hugging Face cache, shared with the company’s other tools. The application runs natively on Windows on Arm and supports NVIDIA RTX Spark from day one, through work carried out with NVIDIA. The code is released under the MIT license in the ggml-org/Llama-Windows repository.

🔗 Llama Windows (Hugging Face blog)

Security: Anthropic’s defenses against distillation, the Chrome exploit built by XBOW, GitHub’s secrets classifier

Three announcements address model and code security: Anthropic reviews its defenses against copying its models, XBOW shows what a public model makes possible in an attack, and GitHub replaces an LLM pipeline with a small classifier.

Anthropic reduces data harvested by distillers by 90 to 99%, according to early evidence

October 12 — Anthropic reviews its defenses against distillation attacks, campaigns that harvest responses from a frontier model, often through networks of fraudulent accounts, to train another model at lower cost. According to early evidence, its new methods reduce the data harvested by around 90 to 99%. According to Anthropic, Zhipu attempted to distill Claude Fable 5 during the development of GLM-5.3, before falling back on an older Claude model and one from another lab.

Three layers are being added: preserved thinking (preserved thinking) in Claude Fable 5.1 and Opus 5.5, which prevents API accounts created since August 31 from modifying Claude’s earlier context, with Sonnet 5.5 reasoning restricted to the organization that issued the request; a two-stage classifier with a false positive rate below 0.1%, active on Sonnet 5.5, Opus 5 and 5.5, Fable 5 and 5.1; and finally, stricter usage limits, bans, and, since September 24, billing for blocked requests. The reported result: 10 to 100 times fewer reasoning tokens extracted and tens of thousands of accounts banned, often in less than two hours.

🔗 Strengthening our defenses against distillation attacks

XBOW builds a complete Chrome exploit with GPT-5.6 Sol, earning a $250,000 reward from Google

October 13 — XBOW, a company conducting penetration tests with AI agents, built a complete Chrome exploit chain (full chain): simply visiting a web page was enough to take remote control of a Windows machine, exposing around one billion users according to XBOW. Google awarded it a $250,000 bounty, and all three bugs have been fixed. The announcement comes from XBOW, not OpenAI, whose @OpenAIDevs account merely reposted it.

We deliberately picked a model that is available to anyone, not a specialized security model or one with restricted access. — Oege de Moor, founder and CEO of XBOW

The selected model is GPT-5.6 Sol, along with its predecessors. The chain cost $40,000 in compute over six days, from August 2 to 7: the AI went through Chrome’s source code, built proofs of concept, and wrote the first versions of the exploit, with four researchers setting the strategy and checking the work. According to XBOW, each of the three bugs could be found by agents alone within a year.

🔗 XBOW’s account

GitHub details its secrets classifier: 86% less analysis time than its former LLM pipeline

October 13 — Six days after announcing its classifier for leaked secrets, GitHub publishes a behind-the-scenes look. The scanner works in two stages: regular expressions cast a wide net to identify anything resembling a credential, then a fine-tuned ModernBERT classifier, exported to ONNX with 8-bit quantization, reads each candidate with its context (file path, secret type, variable names, surrounding code) to discard lookalikes, such as example values or dummy test data.

Metric compared with the LLM pipeline	Reported difference
Median post-push analysis time	−86 %
Serving cost	over 98 % lower
Alert acceptance weighted by repository owner (A/B test)	+11,5 %
Offline precision	+13 %
Offline recall	+19 %

This speed improvement made it possible to extend generic secret detection to push protection, before the code is even sent.

🔗 GitHub’s model for secret detection (GitHub Blog)

Scion: Anthropic replaces git with its own version control system

October 13 — On its Engineering blog, Anthropic explains why it moved away from git for its monorepo: its teams were merging around 2,500 changes per day in May 2026, and more than 30,000 in September. With a history of nearly 100 GB and roughly 25 million checkouts per day for CI, tooling designed for humans was becoming the bottleneck, even with several hundred internal git mirrors.

Scion, built in five months by seven people, assumes that its clients, mostly agents and automated jobs, are always online: the server computes history, line-by-line attribution (blame), and merges, and a CI job often no longer needs a local copy. Anthropic claims a footprint 200 times smaller than its git mirrors, checkouts in seconds instead of minutes, and 25% of CPU capacity freed up for CI; an internal prerelease of Claude Mythos Preview rewrote the ingestion pipeline to use 20 times less memory. Anthropic has no plans for a public release, as Scion diverges too far from git to be useful elsewhere: instead, it is sharing the principles so others can apply them to git, and says it wants to work with GitHub, GitLab, and the open source community.

🔗 Redesigning source control at Anthropic

NVIDIA in France: planned investments in Doctolib, Entalpic, Genesis AI and Pigment, Mistral models in orbit aboard Altair-1

October 13 — At the start of its European tour ahead of GTC Berlin, where Jensen Huang will deliver his opening address (keynote) on October 21, NVIDIA announces planned investments in four French start-ups, without disclosing amounts: Doctolib (agents for healthcare professionals and a health companion for patients), Entalpic (atomic-scale materials design for semiconductors), Genesis AI (general-purpose robots and foundation models) and Pigment (business planning). NVIDIA says it has committed nearly 10 billion dollars to the European technology ecosystem.

The same post provides an overview of sovereign AI in France. Bull will expand its European server production to include Vera Rubin Ultra NVL72 systems, scheduled for 2028 and assembled at its Angers factory. Through the CANUT framework agreement, won by LINAGORA and Docaposte, Twake.ai, built on Nemotron, becomes the only sovereign digital workspace listed for more than 2 500 French public bodies. The day ended with a meeting between Jensen Huang and Emmanuel Macron at the Élysée Palace.

On the space front, Mistral shared the announcement: Altair-1, the first satellite in a 50-satellite constellation that NVIDIA values at 1 billion dollars, has reached orbit. Integrated by Loft Orbital and built in Abu Dhabi by Orbitworks, a joint venture between Loft Orbital and Marlan Space, it carries NVIDIA Jetson AGX Orin modules running Mistral models for real-time Earth observation and autonomous operations; Loft Orbital names Ministral 3 and aims to deliver AI-generated alerts within minutes. No performance figures or precise launch date are provided: Loft Orbital places the launch a few weeks ago, NVIDIA this month.

🔗 The French ecosystem and NVIDIA (NVIDIA blog) · NVIDIA’s GTC Berlin thread · Mistral AI tweet

Agents and coding tools: Origin’s marketplace, Devin in Pylon and its Dynamic Workflows, Amp’s Puck accessible via MCP

Cursor opens its code-hosting marketplace to third-party developers, Devin integrates with a support tool and orchestrates teams of sessions, and Amp makes its meta-agent accessible from other assistants.

Cursor opens Origin’s marketplace to public submissions

October 13 — Origin, Cursor’s code-hosting platform launched in beta on August 17, is opening up to third-party tools. Its marketplace offers more than 20 applications for deployment (Cloudflare, Netlify, Vercel…), continuous integration (Buildkite, CircleCI…), code review (CodeRabbit, Greptile), security (Semgrep, Socket Security), monitoring (Sentry) and project management (Linear). Since October 13, it also accepts public submissions: a team creates its application in its namespace, builds it on the Origin API (application JWT, installation tokens, webhook events, check runs, pull request comments and reviews), then submits it with a demo video; Cursor reviews every submission before listing it. No pricing is announced.

🔗 Origin App Marketplace (Cursor changelog)

Devin integrates with Pylon, offering 15 000 dollars in credits to new customers

October 13 — Cognition launches Devin’s native integration with Pylon, the customer support tool. When a ticket is flagged for engineering, Devin starts on its own: it searches logs, reproduces the bug, identifies the root cause and opens a PR with a fix; support tracks progress in Pylon, engineers take over the linked session, and updates sync in both directions. Support can also launch code searches from Pylon to answer product questions. Testimonials cited by Cognition report fixes delivered around 30 % faster, sometimes in 15 to 20 minutes instead of several hours. New Devin customers on Pylon receive 15 000 dollars in credits.

🔗 Devin in Pylon

Devin’s Dynamic Workflows orchestrate teams of sessions through a Python script

October 13 — Cognition highlights Devin’s Dynamic Workflows: Devin writes a deterministic Python script that decides which Devin sessions to launch, in what order and with what instructions, using structured results (JSON schema) from previous sessions. By default, each agent is a full session on its own virtual machine, with code shared through git branches, or it shares the orchestrating session’s machine. An interrupted run resumes where it left off, within a seven-day limit, and a successful workflow is saved as a skill. The feature is available in every Devin session, but remains disabled in Enterprise accounts until an administrator enables it; the documentation warns about costs, as each agent consumes resources like a session. The launch date is not confirmed anywhere: the documentation page is undated, and the feature does not appear in recent release notes.

🔗 Devin Dynamic Workflows (documentation)

Amp makes Puck accessible from ChatGPT, Claude and other agents via MCP

October 13 — Amp makes Puck, its meta-agent that coordinates other agents, accessible from ChatGPT, Claude and other MCP clients. Amp’s MCP server (ampcode.com/mcp) exposes two tools: puck, for chatting with your Puck, and orbs, an interactive view of active orbs with links to their portals. One example: develop an idea using ChatGPT’s voice mode during a walk, then send it to Puck, which assigns it to the right orb. Connection is handled through Puck’s settings in Amp or by pasting the server address into the client, then signing in with your Amp account. Until now, Puck used MCP servers; it now becomes one. No pricing is announced.

🔗 Take Puck with You (Amp)

Partnerships: Replit’s Solutions Partner Program, Manus in OPPO’s Agent mode

Two distribution agreements: Replit relies on systems integrators to reach businesses, Manus on a smartphone manufacturer.

Replit launches its Solutions Partner Program with Accenture, JSOL, SMFL and Snowman Labs

October 13 — Replit launches a partner program for systems integrators, consulting firms and agencies that support businesses on its platform: training, guidance, collaboration on customer opportunities and certification paths for Enterprise features, in a model built primarily around referrals (referral-first). Initial partners include Accenture, Japan’s JSOL and Sumitomo Mitsui Finance and Leasing (SMFL), and Brazil’s Snowman Labs. SMFL says it built a production-ready business application in one week, and JSOL cites a customer that built a prototype in one month. The partner directory listed nine partners on October 13; no financial terms are published.

🔗 Introducing Replit Solutions Partners

Manus partners with OPPO for AI Mind Pilot’s new Agent mode

October 13 — Manus integrates its agentic capabilities into OPPO AI Mind Pilot’s new Agent mode, which will debut on the upcoming Find X10 series running ColorOS 17. The highlighted example: organize a trip using flight tickets, bookings and attraction tickets scattered across the phone’s applications, combined with online information to produce a day-by-day itinerary. On privacy, ColorOS 17 uses on-device models to process tasks such as searching for or renaming files locally, without sending their contents to Manus, and AI Mind Pilot requests explicit confirmation before any external processing. No date, markets or terms are specified: they will be announced at the phone’s launch.

🔗 Manus Partners with OPPO

Generative video and voice: Reactor’s FastH3-Omni, Runway in Replit, ElevenLabs agents in support and a recreated voice in the Lords

Real-time video gains sound, Runway enters web design, and ElevenLabs showcases two uses of its platform, in customer support and the UK Parliament.

FastH3-Omni: Reactor launches a real-time version of MiniMax H3 with synchronized sound

October 13 — The Reactor platform launches FastH3-Omni, a real-time version of MiniMax H3, MiniMax’s video model, which generates video with synchronized sound from a single prompt. Post-trained by Hao AI Lab’s FastVideo team with Nuva Lab, which had already developed FastH3 in late August, it is presented as the fastest and highest-quality open source version of real-time H3. It adds first- and last-frame conditioning, audio input with dialogue, and up to three reference images. FastH3-Omni is available as an exclusive preview through the Reactor API until October 20; the announcement provides no benchmark, pricing or link to weights. MiniMax shared it that same evening.

🔗 Reactor announcement

Runway connects to Replit for web design

October 13 — Runway and Replit introduce Runway for web design on the same day: once Runway is connected to a Replit project through Runway MCP, the development agent generates images, videos and visual assets directly in the project, from the initial idea to the finished site. The Runway MCP page lists Replit among supported environments, alongside ChatGPT, Claude, Grok Bot, Muse and Cursor; Runway MCP was already compatible with Replit at its launch on May 27, so the announcement is primarily a joint promotion. Following ChatGPT’s app directory on October 7 and Claude Motion on October 8, Runway now turns its attention to web design. No pricing or credit costs are specified.

🔗 Runway announcement · Replit announcement

ElevenLabs resolves more than 22 000 support tickets per month with its own agents

October 13 — ElevenLabs shows how it uses its own platform: its customer support runs on ElevenAgents, which resolves more than 22 000 tickets per month without human intervention, escalating to the team when needed. Last month, the agent handled 65 % of tickets from start to finish, the average first response time fell to 60 seconds from several hours previously, and 70 % of conversations received positive ratings. The agent relies on the support team’s processes, converted into Procedures (in a structured version for refunds), and on webhook tools that check each user’s plan, credits or voices; it was validated through simulation tests across around fifty scenarios, then observed in shadow mode on real traffic before going into production. Architect, launched on October 6, is then used to understand escalations and test fixes.

🔗 ElevenLabs support on ElevenAgents

Baroness Campbell addresses the Lords with a voice recreated by ElevenLabs

October 13 — Baroness Jane Campbell of Surbiton, a disability rights campaigner with spinal muscular atrophy and a life peer since 2007, had relied on an assistant to read her speeches since 2012. This time, she addressed the House of Lords with her own AI-recreated voice during the second reading of the Health Bill, on personal health budgets. According to ElevenLabs, this is the first time a member of the UK Parliament has used an AI-generated voice in either chamber. The Baroness recreated her voice herself using ElevenLabs’ cloning tool and old recordings; the text, rendered through text-to-speech, was played through an assistive device with approval from the Lords’ procedure committee. ElevenLabs says it has restored the voices of more than 11 000 people.

🔗 The first speech delivered with an AI voice in the UK Parliament (ElevenLabs)

Research and open data: Prompteina designs proteins, Basis publishes Global Emotions, a critical post on decision models

Three open publications of the day: a text-driven protein design model, an expressive speech dataset in 83 languages and a quantitative critique of decision models.

Prompteina designs all-atom proteins from a natural-language instruction

October 13 — Lead author Joseph Lee presents Prompteina, a study coauthored by, among others, three researchers affiliated with Ai2, which shared it, and David Baker of the University of Washington’s Institute for Protein Design. The arXiv paper describes it as the first text-driven model for generating all-atom protein structures. Whereas existing models require an input representation specific to each property, here the desired property is written in natural language: across 18 structural properties evaluated, Prompteina achieves better control than the baselines on 15. It was trained on 1.8 million structures paired with 52.4 million captions, adapting NVIDIA’s Proteina-Complexa. In agent-driven binder design, it achieves 14 % higher predicted binding confidence than the best baseline across 20 targets. On GM2A, a target for which no de novo binder had been reported, 90 designs tested in the lab yielded nine binders, the strongest of which has an apparent dissociation constant of 107 nM. The code and weights are released.

🔗 Joseph Lee’s thread · Designing Protein Structures from Natural Language (arXiv)

Basis publishes Global Emotions, 525 hours of expressive speech across 108 locales

October 12 — Basis publishes Global Emotions, a 525.17-hour expressive speech dataset recorded by 10 420 paid native speakers across 108 locales and 83 languages, in 48 kHz WAV format. Its distinguishing feature is its structure: the same actor reads the same line at several intensities, neutral, medium, then full, following an instruction (happy, frustrated, sarcastic, whispered, sung…). Keeping the speaker and text constant isolates what changes in the voice, rhythm, pitch or speaking rate as emotion intensifies. The dataset contains 137 696 recordings and 54 575 matched sets, including 12 141 recordings with a nonverbal cue (sigh, laughter). Basis declares it free for commercial use and research under a custom license with conditional access, but the audio is delivered raw, without denoising, and human-validated transcripts are not included: they are sold for 22.7 dollars per hour of audio.

🔗 Basis Global Emotions (Hugging Face blog)

Decision models: a critical post tests their limits and trains one with GRPO

October 13 — In a community post on the Hugging Face blog, Hajime Yagihara examines public demonstrations of decision models, which generate no tokens and return a probability for each option. His argument: the demonstration code simulates each option and writes the result into the prompt, doing the hard work before the model gets involved. In ViZDoom’s Predict Position scenario, Jev without code assistance succeeds in 0 episodes out of 128, below chance (12 out of 128), while the code alone succeeds in 126; on Snake, it drops to 63.5% as soon as it has to predict a single step.

The author then adapts GRPO from LLMs to these models, distributing the final reward across all decisions in an episode. On NanoJev, an open reproduction with 0.6 billion parameters, the success rate rises from 19.4% to 79.1% without an expert, compared with 68.6% when imitating an expert through SFT, and 81.4% when combining the two in sequence; training remains unstable around step 20.

🔗 Hajime Yagihara’s post (Hugging Face blog)

Briefs

Copilot CLI 1.0.97 — Now stable with 25 entries from its prereleases, it lets each custom agent carry its own declarative hooks, discovers all agents in a repository even beyond 100 and, in Autopilot, rejects an “ask” decision from a preToolUse hook instead of getting stuck; the Copilot SDK moves to 1.0.21. 🔗 source · 🔗 SDK
Gemini CLI v0.65.0-preview.0 and v0.64.0 — The preview adds four fixes from October 13, including a quick interruption with Ctrl+C that no longer corrupts the session, while stable v0.64.0 brings nothing new. Gemini 4 Argon is not yet in the selector: a PR from an external contributor adds it, but has not been merged. 🔗 source
Antigravity 2.23.0 — October 12 release (16 improvements, 8 fixes): an agent permissions selector in the message area, a fullscreen viewer for images and videos, .zip archives unpacked by the agent, and agent-driven repair of an MCP server that fails to start. 🔗 source
Qwen Code v0.25.2-preview.0 — The cycle’s first preview, two days after stable v0.25.1: 115 entries including 33 features and 50 fixes, Mods written in TypeScript that can intercept tool calls, a test failure explorer, and a file browser with Excel previews in the Web Shell. 🔗 source
Jev in Amp — Amp’s Changes panel automatically orders files in a diff, putting those that explain the change first, much faster than before thanks to TypeSafe’s Jev decision model; alphabetical ordering remains available. 🔗 source
Devin’s October 9 notes — Catch-up: an idle session can now be resumed for only 14 days instead of 30, Devin reads the AGENTS.md file when setting up a repository, and Devin Ultra becomes available to enterprises with zero data retention, which previously lacked access. 🔗 source
Anthropic keys blocked by default — GitHub’s secret scanning recognizes Anthropic service keys and user keys and blocks them by default through push protection, including in free public repositories; it adds a detector for Cloudsmith, and Salesforce becomes a partner to revoke Marketing Cloud tokens found in public repositories. 🔗 source
AI Scan metrics — In public preview, GitHub’s security overview measures the impact of AI Scan on pull requests (findings generated and resolved, remediation rate, use of autofix suggestions), for GitHub Code Security or GitHub Advanced Security customers. 🔗 source
Copilot managed settings — With forceRemoteSettingsRefresh, enterprise administrators require Copilot CLI, VS Code and the GitHub Copilot app to retrieve the current policy at startup, with the client failing closed if it cannot confirm it; a minimum client version is also enforced through the sandbox key. 🔗 source
Copilot telemetry — The managed setting telemetry.capture.identity, disabled by default, links Copilot activity to a user in OpenTelemetry (GitHub login, system account, machine name) without collecting their prompts; five content capture classes can be configured separately, all disabled by default. 🔗 source
Code Quality auditing — Manual dismissal and reopening of GitHub Code Quality findings now appear in the audit log, with the actor, time, repository, rule and number of findings, on GitHub Team and GitHub Enterprise Cloud. 🔗 source
Withdrawn Dependabot alerts — The dependabot_alert webhook gains a withdrawn action, and the REST API returns 410 Gone instead of 404 for a withdrawn alert, distinguishing it from an alert that never existed. 🔗 source
Anonymous registries for Dependabot — An Anonymous authentication method lets organizations declare a private registry accessible without credentials, such as an internal mirror open for reading; those who used a dummy token must recreate their registry. 🔗 source
GitHub billing budgets — The Budgets and alerts page brings all budgets together in a single list, filterable by scope, product, SKU or user, including per-user budgets. 🔗 source
Uber’s RSSI and AI agents — In an interview published by claude.com, Philip Martin, Uber’s RSSI, argues for applying an insider threat program to AI agents and describes how Enterprise Frontier Safeguards works with ADR, Uber’s open source tool; no figures or product announcement. 🔗 source
Mistral Large 4 in visual grounding — According to Mistral, its model scores 42% on the Dense 200 visual grounding benchmark, ahead even of leading closed models, without naming a competitor or providing a comparison score; a demonstration shows it working on technical drawings. 🔗 source
JurinAI’s JJ-1 — Four decision models with 0.8 to 9 billion parameters for Japanese business processes: in the publisher’s report (88,813 questions), JJ-1-9B reaches 79.20%, ahead of StartLux 9B (78.02%), Jev 1.13 and Microsoft Decision-1 (76.50%); weights under CC BY-NC 4.0, with no commercial use. 🔗 source
AutoTrust — Two online demonstrations, for short videos and news, rank their recommendations with a single decision model, without training or logs: one choice question for the entire batch outperforms querying each candidate individually (AUC 0.931 in 0.42 s versus 0.905 in 1.54 s across 200 cases); these are measurements outside production, the authors caution. 🔗 source
MolmoAct 2 leads at Datapoint AI — Ai2’s robot policy leads the Datapoint Robot Policy Bench, scored by human preference, with 1,082 Elo points across 21,130 votes, ahead of OpenVLA-OFT (1,049) and π0.5-DROID (1,033); this is a second independent ranking after Reality Check on September 30. 🔗 source
Molmo2Fish — Ai2 describes how Kai Van Brunt (MIT) fine-tuned Molmo 2 on sonar video to track salmon and correct its errors with a simple instruction; the prototype performs about as well as a specialized system, with no published figures, and has not been deployed. 🔗 source
Air traffic control speech recognition — NVIDIA publishes two skills that guide a coding agent through fine-tuning Nemotron 3.5 ASR: on a two-hour ATCO2 benchmark, the word error rate falls from 77.00% to 22.46%, while the rate for general English goes from 3.50% to 3.63%. 🔗 source
Post-training at Together AI — An Apache-2.0 cookbook, released without an announcement, documents Together Post Training, where the customer writes their training loop on their own machine while Together AI runs the trainer and sampling: GRPO, SFT, DPO and multi-LoRA recipes, with no pricing or figures. 🔗 source
Supabase in Grok Bot — Supabase joins SpaceXSI’s partner network: a request made in Grok Bot’s chat is handed to Cursor, which executes it on the backend through Supabase’s MCP server; in Cursor sessions, only about 2% of errors stop the session, according to Supabase. 🔗 source
Higgsfield and GPT-6 Astra — In an OpenAI case study, Higgsfield AI says its videos, capped at around one minute before GPT-6 Astra, now exceed 10 minutes; according to its CEO, the model’s tool calls for 3D modeling help maintain consistency in long videos. No cost figures are provided. 🔗 source

What this means

Argon’s launch shows how a frontier model now spreads: on the day it becomes generally available, Cursor, Devin, GitHub Copilot and Manus add it, each with its own metric, and those figures cannot be compared with one another. On CursorBench 4.0, Argon matches Opus 5.5 at medium effort but costs more per task and falls behind as soon as effort is increased; on FrontierCode 1.1, Cognition’s benchmark, it leads GPT-6.1 Sol by a tenth of a point and remains nearly five points behind Claude Opus 5.5; GitHub publishes no figures. Price matters as much as the score: 2 and 10 dollars per million tokens through November 30, double that afterward. For a team choosing its model, the calculation will need to be done again in December.

The X Subscription changes Cursor’s position. The editor is now sold in the same plan as Grok and X Premium, with a shared usage pool and at least twice as much coding usage on SpaceXSI’s models. Models from other labs, such as Gemini 4 Argon, made available in Cursor the same day, remain accessible, but only starting at 30 dollars per month, and billed at API prices outside the pool. For a new subscriber, the offer therefore steers them toward in-house models; for existing customers, nothing changes until SpaceXSI sets the timetable for its announced migration.

Cybersecurity runs through the day’s news. Google justifies its next-generation safeguards with a result: Argon solves all 12 simulated cyberattack scenarios, compared with 5 for the previous generation. XBOW, meanwhile, shows what a model available to everyone already makes possible: a complete Chrome exploit chain for 40,000 dollars in compute. Defenses focus on real-time monitoring: three layers of protection around Argon, a two-stage classifier at Anthropic, which says it reduces the data harvested by distillers by 90 to 99%, and, at GitHub, a small ModernBERT model that replaces an LLM pipeline, with 86% less analysis time and a serving cost more than 98% lower.

Local AI, finally, gets practical tools. Nemotron 3.5 Bolt 74B fits in 24 GB thanks to 2-bit quantization, LFM2.5-24B-A2B activates only 2 billion parameters per token and claims 184 tokens/s on an Apple M5 Max, and Llama Windows reduces llama.cpp installation to a few clicks. On the server side, NVIDIA opens Run:ai’s code so its partners can orchestrate their own GPUs, and Anthropic shows with Scion what the pace of agents does to tooling: more than 30,000 merges per day in September, compared with 2,500 in May, pushed it to leave git.