AI Daily · Oct 14, 2026: Google's Gemini 4 Argon launches globally at $2 per million input tokens
Gemini 4 Argon is now available globally across the Gemini app, API paid tier and AI Studio with 1M context, priced at $2 per million input tokens and $10 output; DeepSeek V4.1 Flash and HeyGen Voice also launch with limited-time credits and discounts.
Showing the latest briefing (Oct 14, 2026). Every day is generated the next morning.
Pick a date
-
Google releases Gemini 4 Argon globally across API and app
Gemini 4 Argon is available globally in the paid tier of the Gemini API and platform, on Google Cloud, AI Studio, OpenRouter, Cursor and OpenCode, in Google Antigravity for Ultra subscribers, and in the Gemini app and Search AI Mode for Ultra and Pro subscribers. It supports 1M context, up to 1M output tokens, vision, tools and caching, targeting coding, enterprise workflows and cyber defense.
-
Gemini 4 Argon launch pricing: $2/M input, $10/M output
LiteLLM now supports Gemini 4 Argon; launch pricing through Nov 30 is $2.00 per MTok input, $10.00 per MTok output, and $0.10 per MTok cache read; OpenCode Go offers $60 of usage for $10 over the same period.
-
DeepSeek V4.1 Flash offers $100 usage for $10 for three days
DeepSeek V4.1 Flash is offering $100 of usage for the next three days, after which the credit drops to $80; Command Code's GOAT plan gives $100 usage for $10 over the same period, with DeepSeek V4.1 Flash on day 1 and GLM 5.3 Flash on day 2.
-
HeyGen Voice text-to-speech model launches on OpenRouter at 50% off
HeyGen's new text-to-speech model, HeyGen Voice, is now available on OpenRouter with two modes (HeyGen Voice and lower-latency HeyGen Voice Turbo) and 10 built-in voices; API pricing is 50% off through October 31, at $15 per million characters for HeyGen Voice (list $30) and $7.50 per million characters for Turbo (list $15).
-
GLM-5.3 cost per task varies 9x across providers
Artificial Analysis now publishes cache hit rates and cost per task by provider; for GLM-5.3 (max) cost per task varies about 9x across providers, driven by cache hit rate and token pricing, since agentic workloads resend full history each turn.
-
Muse Wonder vs Opus 5.5 coding cost and time compared
On two coding tasks via GMI Cloud, Muse Wonder took 16 min/$3.18 and 1h11m/$16.10 versus Opus 5.5 at 45 min/$8.20 and 1h20m/$23.29, at the same per-token price; Opus won on detail.
-
Qwen3.8-Flash-Next hits 685 output tokens/s per user
Running TokenSpeed on 4x NVIDIA GB300 with NVFP4, Alibaba's Qwen3.8-Flash-Next reaches 685 output tokens per second per user.
-
Liquid AI releases LFM2.5-24B-A2B local reasoning model
LFM2.5-24B-A2B has 24B total parameters with 2B active per token, a 131K-token context, a 128K-token vocabulary, and 33T pretraining tokens, targeting multi-step agent tasks on laptops and PCs; it is now on OpenRouter, reasoning before answering and targeting agents and tool use.
-
NVIDIA open-weights model turns chest CT into report in 30 seconds
A model trained on 70,111 CT scans generates a radiology report from a full chest CT in under 30 seconds on one NVIDIA GPU, with open weights on Hugging Face.
-
llama.cpp ships Windows app for local models with RTX Spark
llama.cpp now has a Windows app called Llama Windows that runs models locally with optimizations for NVIDIA RTX Spark, supports browsing and downloading models from Hugging Face, chatting from the system tray, and keeping prompts on the PC.