Welcome, humans.

Heads up y’all! 🔴 THIS FRIDAY: We’re going Hands-On with GPT-5.1 Codex Max LIVE with OpenAI's Codex Product Lead.

OpenAI's newest agentic coding model can work independently for hours: refactoring entire projects and running multi-hour agent loops without losing context.

On Friday, we're going hands-on with Alexander Embiricos, the product lead for Codex, to show us how to set it up from scratch. Whether you're a total beginner or already building with AI, you'll walk out knowing how to manage Codex agents for your next project.

Click here to get notified when we go live → (Friday, Dec 5 | 10:30am PST / 12:30pm CST. We’re super pumped for this one!

Here’s what happened in AI today:

1. OpenAI declared “Code Red”, building new “Garlic” model to best Gemini.
2. Anthropic CEO warned some AI companies are “yoloing” cash.
3. The Winter 2025 AI Safety Index found leading AI companies “far short.”
4. A new startup launched to make AI smarter to make better AI chips to make smarter AI to make better chips…

Advertise in The Neuron here

OpenAI Declared a “Code Red”(But We Hear a Secret Model Called “Garlic” Might Save Them)

Remember when Google declared “code red” back in 2022 because ChatGPT was eating their lunch? Well, well, well… how the turntables have… tabled.

Quick refresher: So, a lot was made about Sam Altman’s recent “Code Red” memo in response to Google’s Gemini 3.0. This video probably does the best summary of the situation (and it’s hilarious AF).

It turns out, the memo requested the team freeze all non-essential projects like shopping agents, health tools, advertising agents, even their new personal assistant Pulse, and to focus everything on making ChatGPT better. Whatcha think Code Red means people!!

Was Gemini 3.0 really that big of a deal for OpenAI? Weeell…

• Gemini 3 topped industry benchmarks when it launched last month, and users are loving it.
• Not like he’s a big deal or anything, but Salesforce CEO Marc Benioff literally tweeted “I'm not going back” after trying it.
• ChatGPT still has 800M weekly users compared to Gemini's 650M monthly users, but Google's ecosystem advantage (Search, Workspace, Gmail) is a monster. Though TBH, Google’s AI workspace integrations are probably the company’s #1 weak spot right now. Have you tried Gemini in Docs? It works about as well as “Apple Intelligence.” As in, it’s NOT intelligent.

Now here's the plot twist: OpenAI isn't just playing defense here. They're building a new model codenamed “Garlic” that's reportedly performing well against both Gemini 3 and Anthropic's Opus 4.5 in internal tests. We might see it drop as GPT-5.2 or GPT-5.5 early next year. From what we hear, this version’s still a little spicy, too… fitting for garlic!

Meanwhile, Anthropic is cashing in (or should we say, cashing out?): While OpenAI scrambles, Anthropic just announced a $200M partnership with Snowflake AND started prepping for an IPO that could value them at $300B+. They hired Wilson Sonsini (the law firm that took Google and LinkedIn public) and are targeting 2026 for a potential listing. Not financial advice, but based on our personal Claude usage… that’ll probably go quite well.

Speaking of getting spicy, Anthropic CEO Dario Amodei told Andrew Ross Sorkin that while he's bullish on AI technology, he has concerns about the economic side of AI spending, noting that some industry players are “yoloing“ and taking unwise risks with their capital deployment.

1. He explained that Anthropic faces a “cone of uncertainty“ in revenue forecasting: the company's revenue has grown 10x annually for three years (from zero to $100M in 2023, $100M to $1B in 2024, and is projected to reach $8-10B by the end of 2025). but the 1-2 year lag time required to build data centers means companies must commit capital now for uncertain future demand.
2. He also said Anthropic manages this risk conservatively by planning for lower revenue scenarios and maintaining strong margins through its enterprise-focused business model (and charging more per models 👀), while acknowledging that other players in the industry (gee, wonder who he’s talking about…) may be overextending themselves financially by betting too aggressively on future revenue growth.

Why this matters: Early leads mean nothing when the technology moves this fast. Google went from “dead on arrival” in AI to forcing OpenAI into emergency mode in just two years.

For anyone building workflows around these tools: don't get locked into one platform. Today's leader could be tomorrow's also-ran. And if Anthropic's IPO actually happens? That'll be the moment we find out if these sky-high valuations are based on real business fundamentals or just hype.

Consider Anthropic’s S-1 filing likely to be one of the most scrutinized in market history… or at least since WeWork and their “community-based EBITDA.“ Good times.

FROM OUR PARTNERS

Fast enough to feel natural. Accurate enough to trust.

Every millisecond matters in voice AI. Gladia’s speech-to-text engine delivers market-leading accuracy and human-like latency, with partials in ~100ms and robust performance across accents, noise, and overlapping speech.

Whether you’re building voice agents, support tools, or conversational platforms, this is transcription designed for production – fast, stable, and global.

Test Gladia’s real-time STT in the playground →

Prompt Tip of the Day

Here's a phrase that'll improve how you use Gemini's Nano Banana image model: “keeping everything else the same.” Sounds obvious, but it works like magic.

How to use it:

1. Start with a base image: either generate something new or upload one to edit.
2. List out everything you want to tweak, but tackle them individually, only asking for one change at a time per prompt.
3. After each request, add the phrase “…keeping everything else the same.” This locks Nano Banana's focus on the detail you're changing while, y’know, preserving everything else.
4. One change → result. Second change → result. Third change → result. You're iteratively sculpting the image into exactly what you want.

Why this matters: Without this phrase, the model often reinterprets your entire image with each prompt. You then have to wrestle it back to your original vision over and over. This simple addition keeps your edits moving forward instead of sideways.

Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 600K readers here!

1. *Want to run your own ChatGPT-style AI on your own computer? Running massive 70B models (like Meta’s Llama models) usually means choosing between $8,000 GPUs or cloud bills that never stop. The Dell Pro Max with GB10 gives you 128GB for serious AI work at your desk. Get local AI horsepower here.
2. Google rolled out Workspace Studio so you can build Gemini agents that automate everyday Gmail and Docs tasks with “if this, then that”-style flows, no coding required.
3. Taskade builds apps, dashboards, and automation workflows from a single prompt, turning your workspace into the backend; just say what you need, it builds and runs it—check out the community apps on the homepage to see what ppl are building with this.
4. Bloom turns your website into a brand system, then generates marketing assets (social posts, Google ads, banners) that automatically match your brand's colors, fonts, and style.
5. Metagloves Pro Haptic tracks your hand and finger movements with millimeter precision and provides haptic feedback, letting you feel what a robot feels when you're training it or operating it remotely.
6. FrodoBots just open-sourced their sidewalk robot hardware and released the Berkeley-FrodoBots-7K dataset (7,000+ hours of navigation data from 40+ cities, curated using MBRA), so you can build your own autonomous sidewalk robot for under $300.
7. This is a fun use of AI: Stickerbox turns your child's spoken ideas into printed stickers they can color—say “a dragon made of clouds“ and it prints out in seconds—paid only rn ($99.99 for the device).

Around the Horn

Robot that plays ping pong! Paper, Code

1. Anthropic's chief scientist Jared Kaplan warned that humanity faces “the biggest decision yet“ between 2027-2030—whether to permit AI systems to autonomously train successor models in what's called recursive self-improvement—and predicted AI will handle most white-collar work within 2-3 years.
2. OpenAI will buy experiment-tracking startup Neptune to beef up its model-training analytics stack.
3. ByteDance Seed released GR-RL, a vision-language-action model combined with reinforcement learning that achieved 83.3% success rate on continuous shoelace threading (up from 45.7%) by filtering flawed human demonstrations through offline RL (paper, video of a robot threading shoelaces).
4. Ricursive Intelligence came out of stealth as a startup that uses AI to design chips faster, then runs those chips to train even smarter AI, creating a recursive improvement loop (collapsing chip design from years to months).
5. The Winter 2025 AI Safety Index found that leading AI companies including Anthropic, OpenAI, xAI, Meta, and Google DeepMind are “far short“ of global safety standards with no credible plans for controlling smarter-than-human systems if they go off the rails.
6. OpenAI's new experimental confession method trains AI models to honestly report misbehavior with 95.6% accuracy by separating honesty rewards from performance rewards—read our write up on it here.

FROM OUR PARTNERS

Ideas move fast; typing slows them down.

Wispr Flow flips the script by turning your speech into clean, final-draft writing across email, Slack, and docs. It matches your tone, handles punctuation and lists, and adapts to how you work on Mac, Windows, and iPhone. No start-stop fixing, no reformatting, just thought-to-text that keeps pace with you. When writing stops being a bottleneck, work flows. Give your hands a break ➜ start flowing for free today.

Thursday Trivia

One is AI, and one is real. Which is which? Vote below!

A.

B.

Which is AI, and which is real?

Which is AI, and which is real? The answer is below, but place your vote to see how your guess everyone else (no cheating now!)

A Cat’s Commentary

Trivia answer: A is AI, and B is real.