Google opens Gemini 4 Argon for developers and enterprises.
Gemini 4 Argon is live on the Gemini API—with a unified 1M window and a short intro price
gemini-4-argon) globally for developers and enterprises via the Gemini API in AI Studio and the Gemini platform, for Google AI Ultra subscribers in Antigravity, and for Ultra/Pro subscribers in the Gemini app and AI Mode in Search.The operator shape that matters: Argon uses a unified 1,048,576-token context that input, thinking, and output share. Long decode can continue past the usual per-request caps up to the remaining window, with GenAI SDKs continuing by default. Thinking levels are low / medium (default) / high. Explicit context caching is not supported; implicit caching discounts hits by 95% for prompts of at least 4,096 tokens.
Pricing is time-boxed. Through November 30, 2026, paid-tier Argon is $2 per million input tokens and $10 per million output tokens (thinking included), with cached input at $0.10. From December 1 those rates become $4 / $20 / $0.20. Argon is not on the Free tier and not on Google AI Pro. Batch, Flex, and Priority inference are listed as supported on the model card.
Google positions Argon for long-horizon coding, professional knowledge work, cyber defense (Fairwind expansion), science, and robotics. Productivity and benchmark claims in the launch posts (Vals Index ranking, ERQA score, internal Chrome/ADK/Gemma speedups) are Google-reported—treat them as vendor until you measure your own workloads.
What to do: update model strings to gemini-4-argon, remove explicit cachedContents paths, confirm continuation/long-decode SDK behavior, and budget for the Dec 1 price step before you standardize on Argon for always-on agents.
Sources: Google — Gemini 4 Argon (Oct 13) · Model docs · Pricing
No culture/noise items this window that beat the operator bar. NVIDIA UK/France ecosystem posts stay off the card deck (deployment marketing without a new API/runtime delta).
Model & runtime releases
| Item | Date | Badge |
|---|---|---|
Gemini 4 Argon (gemini-4-argon) GA | Oct 13 | Cloud |
| Ollama v0.40.4 GA (decision CLI / MLX Kev) | Oct 13–14 | Local |
| mlx-vlm v0.7.8 (EmbeddingGemma 2, Clef/d1) | Oct 13 | MLX |
| llama.cpp b11608 server special-token SIGSEGV fix | Oct 14 | Local |
| llama.cpp b11604 KleidiAI SME2 Q4_K/Q6_K | Oct 13 | Local |
| NVIDIA PAIR out of beta + llama.cpp | Oct 13 | Local |
| Nemotron 3.5 Bolt 74B (watch) | Oct 15 | Watch |
| RTX Spark laptops | Oct 16 | Watch |
| MLX main commits (tag still 0.32.3) | Oct 13–14 | MLX |
Security / policy / access
| Item | Date | Badge |
|---|---|---|
| Argon Fairwind cyber access expansion (Google) | Oct 13 | Watch |
| AWS Agent Registry custom metadata | Oct 13 | Cloud |
| vLLM Oct 6 advisory batch — still patch to 0.31.0 | Oct 6 | Active |
| No new vLLM / llama.cpp / Ollama advisories Oct 13–14 | — | Quiet |
| Claude notes / OpenAI changelog quiet since Oct 9 / 8 | — | Quiet |
Critical
If you already prototype on Argon at intro rates, put Dec 1 ($4/$20) in the cost model before you standardize always-on agents—and remove any explicit cache API paths.
Watch
RTX Spark ships Friday with PAIR and agent sandboxes. Watch whether Nemotron Bolt 74B (Thursday) is usable day-one in llama.cpp / Ollama / LM Studio as NVIDIA claims.
Opportunity
Ollama 0.40.4 + mlx-vlm 0.7.8 give Mac Silicon a same-week decision-model and multimodal-embedding path. Collapse a two-tool routing stack into one local loop.
Smoke-test gemini-4-argon continuation
On a paid Gemini API key, call gemini-4-argon with thinking medium, confirm implicit caching on a ≥4,096-token prompt, and verify your SDK continues long decode without an explicit cache object.
Upgrade Ollama to 0.40.4 and --ask
Install 0.40.4, pull a decision model, and run ollama run <model> --ask "…" on a real support ticket. On Apple Silicon, confirm the MLX runner is selected for supported architectures.
Pull mlx-vlm ≥ 0.7.8
Upgrade mlx-vlm and exercise EmbeddingGemma 2 multimodal embeddings or a Clef/d1 decision call on a local image+text sample before you change production routers.
Tuesday through Wednesday morning (Oct 13–14 ET). ASTGL / Fraser takes, sourced.
Budget for Argon’s December price double—and drop explicit cache
Argon is the first Gemini 4 model in public API form, and the migration notes are sharper than the marketing. One shared million-token window means thinking and long decode eat the same budget as your prompt. Explicit caches do not work; implicit cache is the only discount path. Intro rates look Sonnet-shaped until Dec 1, when they double. My take: if you are designing always-on agents this week, price the December card in the spreadsheet now, and run one continuation/long-decode smoke test before you bet a migration on it. Leave the Vals Index and internal Google tok/s claims in the vendor column.
Sources: Google blog · What’s new · Pricing
Decision models left the watch column
Yesterday’s Issue parked Ollama 0.40.4 as an rc1. Overnight it went GA. You can run decision models from the terminal with
ollama run … --ask, Kev decision models ride the MLX runner on Apple Silicon, and multimodal embeddings land through llama-server. My take: this is the Mac Studio upgrade of the week. Pull 0.40.4, re-benchmark anything that still assumes the llama.cpp-only default, and try one classification or rubric score on a real ticket before you wire it into routing.Sources: Ollama v0.40.4
Personal AI laptops get an agent stack before they ship
NVIDIA’s Oct 13 post is less about another GPU slide and more about the software that greets you on first boot of RTX Spark (laptops Friday, Oct 16): agent installers, OpenShell plus Microsoft Execution Containers for sandboxing, hybrid local/cloud routing, PAIR out of beta with llama.cpp support, and Nemotron 3.5 Bolt 74B due Thursday. Performance multipliers in the post are NVIDIA/partner-reported. My take: if you care about local agents on Windows laptops, the news is PAIR + OpenShell + decision-model support in llama.cpp—not the game list.
Sources: NVIDIA Blog
mlx-vlm just became the multimodal decision runtime on Apple Silicon
v0.7.8 adds EmbeddingGemma 2 multimodal embeddings, Clef decision models, and LiquidAI d1 / d1-omni—plus Kolibri1 and Xing4.0. That pairs cleanly with Ollama’s MLX decision path. My take: if your Mac Studio stack still treats vision and decisions as separate tools, this is the release that invites you to collapse them.
Sources: mlx-vlm v0.7.8
AWS folded SageMaker agent skills into one toolkit
The AI/ML skill is now a core skill in the Agent Toolkit for AWS, pre-installed in SageMaker Studio JupyterLab, and positioned as the supported replacement for prior AWS Labs skills and plugins. Coverage spans JumpStart, fine-tuning, serverful training, MLflow, processing jobs, and inference optimization. My take: if your coding agent still pins Labs plugins, migrate this week so you are not surprised when those paths go quiet.
Sources: AWS What’s New