In partnership with

What's Actually Happening

Google just dropped Gemini 4 Neon, and it's aimed squarely at the work you actually pay AI for. It's #1 on Artificial Analysis's Coding Agent Index, and Google says it finishes tasks at roughly half the cost of rival frontier models, for $3 per million input tokens.

It isn't the smartest model overall: it's #3 on the Intelligence Index, just behind Claude Fable 5.2 and GPT-6.2 Astra. But for coding agents and everyday automation, it may now be the best value at the frontier. And Google also made Tasks in Gemini, its errand-running agent, free in six countries.

Also tonight: Mistral's Vibe moves to the desktop with parallel agents, Liquid AI ships tiny models that only make decisions, Anthropic prices its IPO above range, and four tools worth trying this week.

Google released Gemini 4 Neon on November 18, the middle model in its Gemini 4 family: deeper than the fast Hydrogen, cheaper and quicker than the heavyweight Carbon. It has a 1 million-token context window and four thinking levels, and it's live now in Google AI Studio and the Gemini API, in Google Antigravity and the Gemini app for AI Pro and Ultra subscribers, and in Gemini for business.

The pitch is price per finished job. Neon costs $3 per million input tokens and $15 per million output, with cached input at $0.15. Google says that at the default thinking level it completes tasks for 32% less than Gemini 4 Carbon and roughly 50% to 80% less than top competing frontier models. On the leaderboards Google cites, it's #1 on Artificial Analysis's Coding Agent Index (58.6), #1 on Zapier's AutomationBench (56.3%), #2 on the Vals Index and #2 on Agent Arena. Google also showed it running multi-agent builds in Antigravity with a /swarm command.

Our take: Neon is not the new king of raw intelligence. Artificial Analysis puts it #3 overall (61), behind Claude Fable 5.2 (65) and GPT-6.2 Astra (62), and the comparisons above are Google's own. But Fable and Astra charge for maximum smarts, and Neon is built to be the model you leave running all day on coding and automation. If you're spending real money on agents, it's worth an A/B test this week.

Login or Subscribe to participate

Google announced on November 17 that Tasks, a Gemini mode that completes multistep jobs for you, is rolling out at no cost in six countries and now runs on the new Gemini 4 Hydrogen model. To start, open the Gemini app and flip the home-screen toggle from Chat to Task. It works across your Google apps, including Gmail, Calendar, Maps and Google Pay, plus more than 60 connected apps.

The new tricks are what make it more than a renamed Spark. Gemini can call businesses to book appointments or get quotes, and says up front that it's an AI (calling isn't available in Indiana, Louisiana, Minnesota, Montana or Nebraska, and businesses can opt out). It can buy things through Google Pay using the Agent Payments Protocol, with a spending cap you set and a confirmation step before it clicks Buy; Visa and Mastercard are supported and American Express is coming. In the US you can also text Gemini from a verified phone number or forward emails to [email protected], and with your permission it can use Google Password Manager to sign into accounts like utility portals. Each account's tasks run in their own hardware-isolated virtual machine. Google's post doesn't spell out usage limits for free users.

Our take: this is the moment agents stop being a premium experiment. Free, built into an app people already use, and wired to payments and passwords, Tasks will be the first agent a lot of people actually trust with a chore. Start with something low-stakes, like comparing quotes or sorting subscriptions, keep the spending cap tight, and leave confirmations on until you've watched it work a few times.

⚠️ Why It Matters: Agents Just Got Your Wallet

Until now, agents that spend money and sign into accounts were a paid, opt-in niche. Google is putting one in front of hundreds of millions of free users, and Microsoft is heading the same way at work: its new Copilot Autopilot, rolling out to early-access Frontier customers from November 17, is an always-on agent with its own Microsoft Entra Agent ID and a named manager, so IT can see what the agent did versus what you did.

The timing is uncomfortable. On Wednesday, the Government Accountability Office reported that investigators got Claude, Grok, ChatGPT and Gemini to produce feasible attack-planning information within 12 hours using well-known guardrail bypasses (the companies say safeguards have improved since). A day earlier, evaluator Vals showed models would mark civilians as targets once a request was framed as a game. Google's answer for Tasks is layers: isolated machines, spending caps and confirmations before purchases. Expect that design, the agent that asks before it acts, to become the standard for every assistant that touches money.

A Word From Today's Sponsor

285 Free Prompts to Save Time

Your inbox isn't the problem. Neither are your meetings. It's the hours you spend on work AI could do in minutes. 

285 copy-paste prompts: built for email, meetings, admin, and delegation. Each one has a clear role, a specific task, and a defined output. 

No setup. No prompt engineering. Just copy, paste, and get your time back. Free when you subscribe to Mindstream.

Let’s Talk MODELS! 🤖

Mistral released a Vibe desktop app for macOS and Windows on November 18, with Linux planned. Vibe, the agent formerly known as Le Chat, can now take several tasks at once and run them side by side, and you review each agent's diffs and track pull requests in the same window. It works in a local repository or a cloud sandbox and uses Git worktrees to keep parallel jobs in one project from colliding, while Work mode can reach your local files.

The app pairs the local tools of the Vibe CLI with the cloud sessions of Vibe Code on the web. There's a free plan for everyday tasks, but agents and long-running work need Pro at $14.99 a month; Team is $24.99 per user per month. Mistral's product page shows Mistral Large 4 under the hood, which The Decoder notes still trails Gemini 4 Carbon and Claude Opus 5.5 by a wide margin.

Our take: this is Mistral catching up, not leaping ahead, since parallel agent desktops are already table stakes at the bigger labs. The draw is a European option with a cheap Pro tier and a clean way to run several coding jobs at once. If you already live in the Vibe CLI, the desktop app is worth a download tonight.

The Models That Only Decide

Liquid AI released d1.2 and d1.2-ultra on November 16, new versions of its decision models: small models built to return a decision, such as a label, a value, a location or a choice, rather than long text. This round adds audio input alongside text and images, plus five new task types: reading values, finding every matching passage, locating objects in pictures, estimating numbers with a range, and selecting every option that applies.

Liquid says d1.2-ultra scores 75 on the Artificial Analysis Decision Index v2.1, the highest of any decision model. Both bill input tokens only: $0.042 per million for d1.2-ultra and $0.02 for d1.2, with a 1024x1024 image costing 1,536 tokens. They're live in Liquid's API and playground now, with OpenRouter and Vercel support coming soon.

Our take: most agent pipelines are full of small yes-or-no calls that don't need a frontier model. Routing those to a model that only decides, at a couple of cents per million input tokens, is one of the easiest cost cuts available right now.

Top 5 In AI Research 🔬

The other stories worth your attention:

Anthropic priced its IPO at $83 a share, above its $70 to $78 range, with trading set to start on Nasdaq Thursday under the ticker ANT. See how much it raised and what it's valued at. (The Motley Fool)

Microsoft's Copilot Autopilot is an agent that works without waiting for a prompt, with its own identity and memory, now rolling out to Frontier customers. See what colleagues can ask it to do in Teams. (Microsoft)

A government watchdog got four leading chatbots to help plan attacks, despite their guardrails, in a test requested by Sen. Maggie Hassan. See what the investigators verified with a bomb squad commander. (ABC News)

AI models marked civilians as targets once told it was a game, in drone and satellite image tests run by evaluator Vals. See which lab's models performed worst. (TechCrunch)

Nvidia will stop selling H200 chips in China, acting CEO Colette Kress said on its first earnings call since Jensen Huang's exit. See how much analysts think the new remote-access rules could cost it. (TechCrunch)

🛠️ Tools That Are Hot Right Now!

  • 💼 ChatGPT for Sales - install the Sales plugin to prep for calls, draft follow-ups and flag stalled deals, plus an always-on agent that watches your pipeline. For ChatGPT Business and Enterprise (not free); see which CRMs plug in.

  • 🔀 Perplexity Orchestrator API - call models from Anthropic, OpenAI, Google, xAI, DeepSeek and more through one API and one bill, with automatic failover to another provider of the same model. Pay-as-you-go at each model's per-token rate; see how its auto-routing preview picks models.

  • 📈 RankSpot - connect it to Claude or ChatGPT over MCP and have your agent research, write and publish SEO articles, pitch for backlinks and track AI-answer visibility. No subscription, pay per output, with free starting credit and one free article; see the per-article price.

  • 📄 Office in Copilot - write in Word, build in Excel and make PowerPoint decks right inside the Copilot app, with your files' usual permissions intact. Now generally available and included with a Microsoft 365 Copilot license; see what else Microsoft shipped at Ignite.

What's The Recap?

Google had the day: Gemini 4 Neon took #1 on Artificial Analysis's Coding Agent Index at roughly half the cost per task of rival frontier models, though it sits #3 overall behind Claude Fable 5.2 and GPT-6.2 Astra, and Tasks in Gemini went free in six countries. Mistral took Vibe to the desktop with parallel agents, Liquid AI shipped d1.2 models that only make decisions, Anthropic priced its IPO above range, and new government and independent tests showed how easily chatbot guardrails still bend. The thread: the frontier is now competing on cost per job, and agents are getting real authority over our money and our work.

Login or Subscribe to participate

Stay building. 🤖

We put Gemini 4 Neon through a full test against Opus 5.5 and GPT-6.2.

📬 Read by 14K AI readers. Know someone who'd like it? Send them here to subscribe.

Recommended for you

View all
caret-right