Daily AI NewsWednesday, October 21, 2026

AI NewsOctober 21, 2026

5 curated artificial intelligence stories from Wednesday, October 21, 2026 — AI model updates, product launches, and research breakthroughs of the day.

5 Stories1 Source10 Min Read
GoogleOpen AI

Top AI News Stories on October 21, 2026

5 Stories
Unrot21 Oct 2026, 05:00 amBy Unrot Team

Cloudflare Previews Clairic, an AI That Pulls Facts Out of Documents

Exa 3.0 is an upgraded web search API for AI agents, with a new search-matching model, a $4 per 1,000 requests Lite tier, and generally available Dynamic Highlights.

Cloudflare Previews Clairic, an AI That Pulls Facts Out of Documents

Install the Unrot app to read the full story

Download on the App StoreGet it on Google Play
Unrot21 Oct 2026, 04:58 amBy Unrot Team

Cloudflare Previews Clairic, an AI That Pulls Facts Out of Documents

Clairic reads text or images and returns typed values such as totals, dates and IDs, with a confidence score, instead of writing text. It costs $0.10 per million input tokens, with no output charge.

Cloudflare Previews Clairic, an AI That Pulls Facts Out of Documents

Install the Unrot app to read the full story

Download on the App StoreGet it on Google Play
UnrotGoogle21 Oct 2026, 04:58 amBy Unrot Team

Google Open-Sources TorchTPU, Letting PyTorch Run Natively on Its TPUs

TorchTPU lets developers run ordinary PyTorch code on Google’s Cloud TPU chips by selecting a “tpu” device. Google and Meta built it, and it replaces the older PyTorch/XLA route.

Google Open-Sources TorchTPU, Letting PyTorch Run Natively on Its TPUs

Install the Unrot app to read the full story

Download on the App StoreGet it on Google Play
Unrot21 Oct 2026, 04:57 amBy Unrot Team

Inception’s Mercury 3 AI Writes Over 1,000 Tokens a Second

Mercury 3 is a diffusion language model: it drafts the whole answer at once and refines it in parallel, instead of writing one word after another. Inception reports 1,068 tokens per second and a 1M-token context.

Inception’s Mercury 3 AI Writes Over 1,000 Tokens a Second

Install the Unrot app to read the full story

Download on the App StoreGet it on Google Play
UnrotOpen AI21 Oct 2026, 04:57 amBy Unrot Team

OpenAI Lets Developers Cache Prompts for 12 Hours

OpenAI’s API now offers 12-hour prompt caching, so repeated instructions and context stay cached and cheaper even when requests are hours apart. For logged-in Codex and Work users, OpenAI’s Tibo says the cache is kept forever.

OpenAI Lets Developers Cache Prompts for 12 Hours

Install the Unrot app to read the full story

Download on the App StoreGet it on Google Play
Daily AI News Archive

AI News Archive: Browse Daily AI News by Date

Explore daily AI news digests, model launches, and research breakthroughs from other dates.

View All Dates Archive
AI News Today: October 21, 2026 — Latest AI Updates | Unrot AI