The Neuron
Posts
😺 ChatGPT gets a $100 tier

😺 ChatGPT gets a $100 tier

PLUS: AI hardware may have found its lane

Grant Harvey
April 10, 2026

Welcome, humans.

Button looks like someone found an old iPod Shuffle, whispered “what if this had ChatGPT,” and actually shipped it. Built by ex-Apple Vision Pro engineers, the device is basically a tiny AI talk button: press, speak, get an answer. That simplicity might be its superpower.

Unlike the parade of cursed AI gadgets that tried to live on your face, chest, or soul, Button only listens when you press it. Which is a nice change of pace for anyone who prefers their hardware helpful and not vaguely haunted.

It leans on your phone for connectivity, which makes it feel less like “the future of humanity” and more like something you might actually use.

After the Humane era, an AI gadget that’s small, straightforward, and not trying to become your personality? That alone feels kind of revolutionary. We’re excited to check these out! You can pre-order here.

Here’s what happened in AI today:

Cerebras showed Codex Spark building a Salesforce clone in 29 seconds.
A court let the Pentagon’s Anthropic blacklist stand for now.
OpenAI paused Stargate UK over energy costs.
Poke launched a text-message AI agent for everyday tasks.
Gen Z is still using AI, but likes it a lot less.

P.S: Want to reach 675,000 AI-hungry readers? Click here to advertise with us.

P.P.S: Love robots? We’re starting a new robotics newsletter! Sign up early here.

😺 OpenAI Adds $100 Tier and Gives Codex-Curious Users Room to Build

OpenAI just made Codex a lot more real.

In a series of posts on X, the company announced a new $100 / month ChatGPT Pro tier aimed at heavier Codex usage. The key hook: it offers 5x more Codex usage than Plus, while the existing $200 Pro tier remains the highest-usage option.

On OpenAI’s official pricing page, Pro is listed as a broader ChatGPT plan, not a Codex-only subscription, with perks including GPT-5.4 Pro, unlimited GPT-5.3 and uploads, and “maximum Codex tasks.”

While energy and compute constraint continues to be the topics du jour around AI, the company continues to double-down on rate resets and encouraging more usage. In three months of 2026, Codex shed its “that other tool” status and with 3 million weekly average users is in a tight deadlock with Claude Code.

The ladder now stretches from Free to Go, Plus, Pro, Business, and Enterprise. Free includes limited Codex access, Plus includes expanded Codex usage, Pro unlocks the most individual access, and Business adds the ability to assign standard or usage-based Codex seats across a team. That’s a pretty wide on-ramp for anyone who wants to experiment before going full goblin mode with agentic coding.

And that’s why this matters. A lot of AI tools still feel like they were built either for hardcore engineers or for executives in a demo. Codex sits in a more interesting middle ground. Used well, it can help normal knowledge workers edit files, automate little annoyances, fix lightweight bugs, and build useful internal tools without spinning up a full software project. They’ve yet to produce a Claude Cowork competitor, but many of the same tasks are possible in Codex as well. We’re low-key shocked they haven’t done this, yet.

Our take: This is win for AI accessibility. A pricing ladder that runs from free to $20 to $100 to $200 gives more people a chance to tinker, get value, and then decide whether the upgrade is worth it. That’s a much healthier path than locking the good stuff behind one giant paywall.

🎓 AI Skill of the Day: NotebookLM in Gemini?!

Build a dedicated AI notebook for any project that lasts longer than a day. Google just rolled out notebooks in Gemini, which let you group chats, files, and custom instructions into one place instead of starting from scratch every time.

The big idea is simple: if you’re working on something ongoing, job search, research project, presentation, content series, side hustle, stop treating every AI chat like a one-night stand. Give it a home.

This centers around the idea of creating persistent context. Instead of re-explaining the same project over and over, you drop in your docs, PDFs, notes, and prior chats, then keep building from there. Google says those notebooks also sync with NotebookLM, so you can use features from both apps on the same material, like starting a notebook in Gemini, then using NotebookLM for things like Video Overviews or infographics.

That means less copy-pasting, less “wait, let me explain this again,” and fewer moments where the AI forgets your whole situation like it just took a hard blow to the head.

Total AI beginner? Start here (goes with this video).

Have a specific skill you want to learn? Request it here.

New episodes air every week on: Spotify | Apple Podcasts | YouTube

🍪 Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 675K+ readers here!

HappyHorse-1.0 lets you turn a prompt or image into polished video clips, and it just shot to the top of a major video-model leaderboard under Alibaba’s anonymous launch.
Repo Prompt picks the right code context so your coding agent can make the change you actually want.
Genspark for Microsoft Office lets you make slides, docs, research, and analysis right inside Microsoft 365, so you can stay in Word, Teams, and your other Office apps instead of bouncing between tabs. Genspark’s Microsoft listing includes a free tier with 100 credits per day.
Google Finance Beta lets you ask a market question like “why is Nvidia up?” and get live data, news, and charts in one place.
Brila turns your Google Maps reviews into a website that sells what customers already love about you—free to try.
Grass gives your coding agent a 24/7 VM so it can code in its own box, not on your machine—free to try.

📰 Around the Horn

Cerebras showed Codex Spark building a working Salesforce-style CRM in 29 seconds, arguing ultra-fast coding models could start replacing parts of SaaS prototyping.
A federal appeals court refused to immediately stop the Pentagon from blacklisting Anthropic, extending the startup’s high-stakes fight over military AI use.
A new Gallup-backed Gen Z survey found young people are using AI just as much as last year but feeling far less excited and more angry about it.
OpenAI paused its Stargate UK data center plans, saying it will revisit the project when energy costs and regulation make the economics work.
Poke launched an AI agent that works through iMessage, SMS, Telegram, and WhatsApp, aiming to make automation as easy as sending a text.

FROM OUR PARTNERS

Build profitable, high-quality software with AI – no tech skills required

Not a side project collecting dust. Not a demo you show at meetups. A real product with real users who pay recurring revenue — built with AI, structured by CTO judgment, and 100% owned by you. The AI Founder Accelerator makes that a system, not a dream.

→ Request Early Access

📖 Intelligent Insights:

The OpenAI Foundation said it is finalizing more than $100 million in grants across six institutions to use AI to speed up Alzheimer’s research and drug discovery.
ngrok explained quantization in plain English and showed why LLMs could often be made about 4x smaller and up to 2x faster: most of their internal numbers clustered near zero, so compressing them barely hurt accuracy.
KellyBench tested AI agents across an entire simulated Premier League betting season and found that even frontier models usually lost money, revealing a big gap between sounding smart and making good long-term decisions.
MARS showed that a regular next-token language model could be fine-tuned to predict several tokens at once, boosting throughput by roughly 1.5–1.7x without extra heads, a second draft model, or much accuracy loss.
Latent-Y showed that an autonomous agent could turn plain-English drug-design goals into lab-tested antibody candidates across nine targets, with some strong binding results and a claimed 56x speedup over expert-only workflows.
Claw-Eval argued we should score the whole agent trajectory, not just the final answer, and found that standard opaque grading missed 44% of safety violations and 13% of robustness failures.