Gemini 4 Argon Is Generally Available on Google Cloud and in GitHub Copilot
Google made Gemini 4 Argon, its new frontier model, broadly available after a restricted launch to cybersecurity partners at the end of September. Developers can call it as gemini-4-argon through the Gemini API, Google AI Studio and Google Antigravity; it also powers the Gemini agent in Gemini Enterprise. Launch pricing is $2 input / $10 output per million tokens with a 95% discount on cached input through November 30, 2026, after which prices double to $4 / $20.
Google's coding claims are mostly relative: agentic code submissions in Antigravity rose over 35% while rollbacks fell by nearly a third, the Chrome team used it for an architectural refactor spanning hundreds of files in Chromium, and Argon itself optimized the open-source adk-python package, cutting in-memory session preparation time by 40% and cold import time by 21%. On security, Google reports it leads the Gray Swan indirect prompt injection benchmark with a 0.7% attack success rate across 15 adaptive attempts and ties for first on CWE-Bench v1 at 68%. Offensive cyber tasks are blocked by guardrails; vetted defenders can get an unguarded variant through the Fairwind Program.
The same day, GitHub made Gemini 4 Argon generally available in GitHub Copilot for Pro+, Max, Business and Enterprise users, across VS Code, Visual Studio, JetBrains IDEs, Xcode, Eclipse, Copilot CLI, the Copilot app and the coding agent. It is billed at provider list pricing under usage-based billing, and under default model enablement Business/Enterprise admins will see it switched on unless they have disabled new models.
Read more — Google Cloud Blog
Claude Code 2.1.299–2.1.302: Idle Compaction, Fork Subagents Everywhere, Tighter Hook Validation
Four Claude Code releases landed between October 14 and 16. The headline feature in 2.1.299 is idle compaction: a terminal conversation over 200K tokens that sits idle for about 54 minutes is compacted just before its one-hour prompt cache expires, so resuming a long session no longer pays to re-cache the full context. It can be tuned with CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS, disabled with "idleCompaction": false, and 2.1.302 added an "Idle compaction" toggle to /config. The same release changed MCP allow rules to match only the exact named server — write mcp__<server>__* to allow everything on a server — and made UserPromptSubmit hooks run before a skill or custom slash command expands.
2.1.300 validates updatedInput returned from PreToolUse and PermissionRequest hooks the same way as Claude's own tool input, force-kills background commands 1.5 seconds after a stop request, gzip-compresses large API request bodies, and adds cache_creation_5m_tokens / cache_creation_1h_tokens to OpenTelemetry api_request events. --sandbox/--no-sandbox were removed from claude remote-control in favour of the sandbox.enabled setting.
2.1.301 adds classifier_verdict, permission_mode and classifier_category to OpenTelemetry tool_decision events, vim-mode % bracket matching, and LaTeX rendering in VS Code. 2.1.302 adds CLAUDE_CODE_AUTO_MODE_MODEL to choose the auto-mode classifier model, turns fork subagents on by default outside the terminal (opt out with CLAUDE_CODE_FORK_SUBAGENT=0), starts claude -p and Agent SDK sessions in default permission mode on third-party providers, and notes that standalone Claude Design closes on December 14, 2026.
Read more — Claude Code Docs
Codex CLI 0.163 Adds a Searchable Config Panel and Mid-Response Steering; GPT-5.5 Retires From Codex
OpenAI shipped Codex CLI 0.163.0 on October 15. /config now opens a searchable settings panel that applies some changes immediately and flags which ones need a restart. Users can send a follow-up while Codex is responding to steer it, /keymap introduces a configurable leader key (default Ctrl+X), /daemon restarts the managed background server with a warning about interruptible work, and independent read-only tools now run in parallel. Bug fixes restore a resumed thread's permissions and network approvals, and the network proxy now checks permission before DNS lookups and rejects duplicate Host headers.
On October 14, GPT-5.5 was retired from Codex for users signed in with ChatGPT; API-key sessions are unaffected and the suggested replacement is gpt-6.1-sol. Teams pinning gpt-5.5 in Codex config files under ChatGPT sign-in should update them.
The Codex desktop app also gained a beta of composer predictions (October 9): after Codex responds, it suggests a follow-up message that Tab accepts as an editable draft — nothing is sent automatically. The beta is limited to adult users on personal Pro plans with GPT-6 Astra or GPT-6.1 Sol.
Read more — OpenAI
GitHub Copilot: HydraFusion Max and Agentic Autofix for Code Scanning Reach Users
GitHub updated HydraFusion, its research-preview option that orchestrates multiple models into a single response in VS Code, the Copilot app and Copilot CLI. The model pool now includes Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna; reasoning effort is set per stage (medium for drafts, high for review); the Cascade workflow can add a judging step that requests a repair only when needed; and the Critique workflow now inspects the actual workspace with tools and runs parallel critic models. A new HydraFusion Max tier uses higher reasoning effort for demanding tasks. Both are available on Pro, Pro+, Business and Enterprise with preview features enabled.
Separately, agentic autofix for code scanning alerts became generally available for GitHub Code Security and Advanced Security customers on GitHub Enterprise Cloud. Instead of generating a one-shot patch, Copilot explores relevant files, proposes a fix, re-runs CodeQL to confirm the alert is closed (iterating if not), and opens a draft PR explaining the fix and validation steps. Up to 25 alerts can be assigned together from the alert list or a security campaign, or via the REST API by assigning copilot-swe-agent[bot]. It consumes AI Credits from the Code Security budget plus Actions minutes, and replaces the free "Generate Fix" button when enabled.
Read more — GitHub Changelog
Reflection Releases Beam, a 501B Apache-2.0 Open Model for Agentic Coding
New US lab Reflection released Beam, a sparse mixture-of-experts model with 501B total and about 23B active parameters (52 layers, 256 experts per MoE layer, 8 routed per token), under the Apache 2.0 license. It is text-in/text-out with a 1M-token context window; most layers use 512-token sliding-window attention with nine global-attention layers. Beam has five reasoning effort levels (low through max) and can preserve reasoning across turns with preserve_thinking.
Reflection-reported scores include 81.6% on SWE-Bench Verified, 80.1% on Terminal Bench v2.1, 88.3% on LiveCodeBench and 90.6% on GPQA Diamond. Checkpoints come in BF16 (about 1 TB), FP8 and NVFP4 (about 280 GB); NVFP4 needs Blackwell GPUs with four B200s suggested as a starting point. There is day-0 support in Transformers, vLLM and SGLang plus a DFlash2 drafter for speculative decoding.
For developers who want to try it without hardware, Ollama made Beam free for a limited time through Ollama Cloud, and Google Cloud announced day-0 serving support with llm-d and NVIDIA Dynamo.
Read more — Hugging Face
Meta Publishes Muse Spark Open, an 816B Multimodal Open-Weight Model
Meta released Muse Spark Open, an open-weight multimodal MoE with roughly 816B total and 42B active parameters (256 experts per layer, 8 active). It accepts text, images, audio and video and outputs text, with a 1M-token context window shared across modalities. Meta says it was trained for agentic coding — reading codebases, writing and debugging code, and calling tools — and calls it the most capable US open-weights model to date, though the launch post reports no benchmark numbers or license terms.
Reasoning cannot be disabled; effort can be set from low to xhigh. Running it is a serious hardware exercise: BF16 needs four nodes of 8×H200 in Transformers, while the NVFP4 checkpoint fits a single B200 node or four GB300 GPUs in SGLang. For local experimentation, llama.cpp GGUF quants exist — a 2-bit quant ran at about 18 tokens/s on a 512 GB Mac M3 Ultra — and MLX 3-bit quants need more than 256 GB of RAM.
Read more — Hugging Face
Hermes Agent v0.22.0 "The Threshold Release" Adds Guided Setup, App Connectors and Local Models
NousResearch shipped Hermes Agent v0.22.0, a large release spanning about 24,200 commits since v0.21.0. New users get a guided setup chat on first desktop launch running on a free trial model; hosted apps such as Linear and Notion can be connected through the Nous tool gateway; and local models are one click away via llama.cpp. The project now publishes tested SemVer stable releases, signed desktop packages and a Microsoft Store build, alongside a catalog of about 600 reviewed plugins and support for the Claude 5.5 and GPT-6 model families.
Source installs now follow tagged releases instead of every commit on main (switch back with hermes update --set-channel main). Users should skip v0.22.1, which the maintainers say was built from the wrong commit and could leave the Store app stuck on "Connecting"; v0.22.2 replaces it and fixes gateway startup after reboots, with a known issue that sessions created under the 0.22.1 gateway cannot be deleted.
Read more — GitHub
Links & Sources
- Gemini 4 Argon available on Google Cloud
- Gemini 4 Argon in GitHub Copilot
- Claude Code changelog
- OpenAI Codex changelog
- HydraFusion updates and new Max option
- Agentic autofix for code scanning alerts general availability
- Welcome Beam: a 501B open model you can own, customize, and scale
- Reflection's Beam is freely available on Ollama for a limited time
- Welcome Muse Spark Open
- Hermes Agent releases
You May Also Enjoy
