AI news, November 19, 2026: Gemini 4 Neon, Replit Missions, free Gemini Tasks in Korea
Five stories from Tuesday and Wednesday, November 17โ18, 2026 (US time) for solo makers who build with AI. One is a new mid-priced coding model, one is an agent product that works to a spending cap, two are Gemini app changes that affect your users and your own data, and one is a pair of open-weight document extraction models. Figures are as the companies and outlets state them.
Google releases Gemini 4 Neon, a fast frontier model priced at $3 / $15 per million tokens
Google DeepMind introduced Gemini 4 Neon, which it positions for fast end-to-end work on software engineering, knowledge work and science. Google says Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, ranks #2 on the Vals Index, and had its lowest fabrication rate yet in internal Antigravity testing. API pricing is $3 per million input tokens, $0.15 cached and $15 output, with a 1 million-token context window, and it is rolling out in the Gemini API paid tier in AI Studio and Android Studio, in the Gemini app for Pro and Ultra subscribers, and in Antigravity. The Antigravity team adds that its Orchestrator can now pair Neon as the planner with Gemini 4 Hydrogen as the executor in one session.
Why it matters for makers: Neon sits between Carbon and Hydrogen on price and speed, and the planner/executor pairing is a cheap way to keep hard reasoning on the bigger model while routine steps run on the faster one.
Sources: Google, Google Antigravity
Replit launches Missions, agent swarms that work toward a goal within a budget cap
Replit's new Missions feature lets you set a goal, success criteria, a spending limit and which actions need your approval in a "Mission Agreement", then a swarm of agents plans, builds, tests and iterates toward it. The New Stack reports you start one by typing /mission and that Replit checks spending against the budget at every step, blocking paid services once the limit is reached. Replit says it used Missions to build an in-house application monitoring platform with dashboards, alerts and tracing in two days on a $3,000 budget. It also reports NumPy speedups of up to 19 times on selected workloads using 32-agent swarms.
Why it matters for makers: a hard budget and explicit approval rules are the two controls that make long-running agents safe to leave alone, and they are worth copying even if you build your own agent loop.
Sources: Replit, The New Stack
Gemini Tasks, formerly Spark, is now free in six countries including Korea
Google renamed its Spark agent to Tasks and is rolling it out free of charge in the U.S., Japan, Korea, Canada, Australia and India, running on the new Gemini 4 Hydrogen model; it previously required a paid plan. You switch the toggle on the Gemini home screen from Chat to Task and describe what you want done. 9to5Google reports that tasks run in a sandboxed, hardware-isolated virtual machine and that Gemini asks for confirmation before purchases. Extra features such as calling businesses, paying with Google Pay and assigning tasks by email or text start in the U.S. only.
Why it matters for makers: a free, mass-market agent in Korea means more users will expect to hand chores to an assistant, so clear, machine-readable pages and simple checkout flows on your own site matter more.
Sources: Google, Google Korea, 9to5Google
Gemini splits history from AI training, but keeps "history off" chats for about two months
Google is giving the Gemini app separate settings for chat history and for whether chat content is used to train generative AI models, so you can keep history while opting out of training. At the same time, chats made with History off will be saved for around two months to respond to you, process feedback and for safety, which TechCrunch notes is up from the previous 72-hour window. A Google spokesperson told TechCrunch this retained data will not be used for model training or other product improvement. The changes roll out globally in the coming weeks.
Why it matters for makers: if you paste client code or private data into the Gemini app, turning History off no longer means it is gone within days, so review the new settings or keep sensitive work in tools with retention terms you have checked.
Sources: Google, TechCrunch, 9to5Google
Cloudflare releases open weights for its Clairic document extraction models
Cloudflare published an updated Clairic model and a new, lighter Clairic-flash with 9 billion parameters, both built to pull structured fields out of documents. The company says the update adds extraction from images and PDFs and handles tables, lists and dates more reliably, using a redesigned architecture that generates each field in parallel. Both are hosted on Workers AI as @cf/cloudflare/clairic and @cf/cloudflare/clairic-flash with an unchanged API schema, and full open weights are on Hugging Face.
Why it matters for makers: invoice, receipt and form parsing is a common side-project need, and an open-weight extractor lets you start on a hosted endpoint and later run it yourself.
Source: Cloudflare