AI SIGNAL ▮

Wednesday, Oct 14, 2026

Top stories

Gemini 4 Argon: Google's new flagship at $2/$10 per million tokens, introductory

Google released Gemini 4 Argon on 13 October, available at launch through the Gemini API (AI Studio, Android Studio; paid tier), the Gemini app and AI Mode for Ultra and Pro subscribers, and Antigravity for Ultra. Pricing is $2 per million input and $10 per million output tokens with cached input 95% off, but only until 30 November, after which it rises to $4/$20. Google claims the top spot on the third-party Vals Index and 87% on the ERQA embodied-reasoning benchmark, and describes a 1M combined input-and-output token limit. Several other scores come from private, unverifiable evaluations, and a guardrail-free variant is limited to vetted security defenders.

Sources & depth

Anthropic replaced git for its monorepo with an in-house system called Scion

Anthropic wrote up how it moved its monorepo off git in about five months to Scion, a server-centric version control system with always-online clients, lightweight changesets, cache tiers and optional sparse checkouts. The driver was agent-written code: daily merges rose from roughly 2,500 to over 30,000, which strained git's model of every clone being a full offline copy. Anthropic reports about 200x lower infrastructure needs and checkouts in seconds rather than minutes, and says it also built a new code-review tool alongside it.

Sources & depth

Ollama 0.40.4 adds MLX scoring for Kev decision models

Ollama's 0.40.4 release, which went from rc1 to final within a day, adds support in its MLX backend for loading Kev models with their FP32 decision head and release metadata, and for scoring with shared-state caching. It is the first sign of a local runtime treating TypeSafe-style decision models as a first-class model type rather than a text generator. Scope is limited to the MLX (Apple Silicon) path.

Sources & depth

Unsloth beta claims up to 5x faster inference and adds browser-agent controls

Unsloth's v0.1.907-beta says its UI and inference are much faster, with some multi-chat workloads up to 10x quicker and inference up to 5x faster, and improved context compaction. It also adds scripted browser-use controls (click, drag, per-site script approval) and new image and video features including LTX-2.5 generation and faster image-LoRA training. The speedups are the project's own figures and have not been independently measured.

Sources & depth
Also today
Frontier releases
Inference & serving
Open weights
Hardware
Research
From the community