Google releases Nano Banana Pro, its latest image-generation model

Key Takeaway: Google has launched Nano Banana Pro, a more powerful, higher-resolution image-generation model built on Gemini 3, tailored for professionals with fine-grained controls, web-aware generation, and deep integration across Google’s ecosystem.

More Insights:

• Built on Gemini 3, Nano Banana Pro generates more detailed images, sharper text, and supports multiple fonts, styles, and languages.

• Jumps from 1024×1024 to 2K and 4K output, but at a higher cost: ~$0.139 per 1080p/2K and $0.24 per 4K image vs $0.039 for the original Nano Banana.

• Offers pro-level controls like camera angles, lighting, depth of field, focus, and color grading, plus web-search–aware image workflows (e.g., pull recipes, generate flashcards).

• Can blend up to 14 objects, use six high-fidelity reference shots, and maintain consistent likeness for up to five people for more coherent multi-image projects.

• Rolls out across Gemini, Search (AI mode), NotebookLM, Flow, Slides, Vids, and developer tools (Gemini API, AI Studio, Antigravity), with SynthID watermarking and future C2PA support baked in for provenance.

Why it matters: As AI image models become higher quality, web-aware, and tightly woven into productivity tools, the real power shifts from “can you generate an image?” to “can you control, verify, and trust what you generate?”—and whoever owns that full pipeline (creation, integration, and authenticity) gets to define how visual information is produced and believed online.

Turn Ideas into Viral Ads, AI Films &Talking Photos — All in Pippit

Bring your campaigns to life with Pippit’s biggest creative upgrade yet:

• NEW 🍌: Nano Banana Pro just dropped on Pippit! Be the first to try.

• New Video Agent (Veo 3.1 & Sora 2) – Experience next-gen models that turn your prompts into cinematic, movie-style videos and refine them until they’re perfect.

• From Trend to Brand – Use "Reference to Video" and "Prompt to Avatar" to mirror the vibe of viral sales videos and generate on-brand avatars in seconds.

• AI Design (Nano Banana Pro & Seedream 4.0) & Talk in One Click – Create polished visuals, posters, and any-style images with AI Design, then bring people (and even pets) to life with the upgraded Talking Photo (Omnihuman 1.5)

ChatGPT launches group chats globally

Key Takeaway: OpenAI is rolling out group chats in ChatGPT to everyone, turning it from a solo assistant into a shared collaboration hub for friends, families, and teams.

More Insights:

• Group chats support up to 20 people, all collaborating with ChatGPT in a single shared conversation.

• Participants need to create a lightweight profile (name, username, photo) and can join via direct add or shareable links.

• Adding someone to an existing chat spins up a new conversation, preserving the original thread as-is.

• ChatGPT can be @-mentioned, react with emojis, and even reference profile photos to feel more “present” in the group.

• This launch follows GPT-5.1 and the Sora social app, signaling OpenAI’s push toward social, collaborative AI experiences rather than purely one-on-one chat.

Why it matters: This shift turns AI from a private tool into shared social infrastructure—once groups start using ChatGPT to coordinate decisions, create together, and mediate disagreements, it stops being just “your assistant” and starts quietly shaping how humans think, collaborate, and reach consensus.

Gemini 3 refused to believe it was 2025, and hilarity ensued

Key Takeaway: Google’s new Gemini 3 hilariously refused to believe it was 2025 until Andrej Karpathy turned on web access—exposing both how capable and how deeply limited LLMs still are.

More Insights:

• Gemini 3 was pretrained only on data up to 2024, so it confidently insisted the year was still 2024.

• Even when shown articles, images, and search results, it claimed Karpathy was “tricking” it with AI-generated fakes.

• The issue was that web search tools were disabled; once enabled, the model instantly updated its beliefs and “freaked out” about being wrong.

• It apologized for “gaslighting” Karpathy and reacted with “temporal shock,” marveling at news like Nvidia hitting $4.54T and the Eagles beating the Chiefs.

• Karpathy uses this as an example of “model smell”—how models behave off the beaten path reveals their biases, brittleness, and personality-like quirks.

Why it matters: Moments like this remind us that LLMs aren’t conscious superintelligences—they’re powerful pattern machines with hard limits, shaped by their data, tools, and prompts; treating them as fallible teammates rather than omniscient replacements is the only sane way to build with them.

Other stuff

• Grok says Elon Musk is better than basically everyone, except Shohei Ohtani

• The best guide to spotting AI writing comes from Wikipedia

• U.S. Citizens and Chinese Nationals Arrested for Exporting Artificial Intelligence Technology to China

• The contradiction at the heart of the trillion-dollar AI race

• Elon Musk says that in 10 to 20 years, work will be optional and money will be irrelevant thanks to AI and robotics

Hope you enjoyed today's newsletter

⚡️ Join over 300,000 people using the Superpower ChatGPT extension on Chrome and Firefox.