Skip to content
Source · Daily Brief

AI Daily Brief — 25 March 2025

One of the most consequential collision days of the AI year. OpenAI made GPT-4o the native image generator in ChatGPT and Sora — ending the DALL-E era — and Google DeepMind shipped Gemini 2.5 Pro Experimental hours later, leaping to #1 on LMArena. Within an hour the Studio Ghibli style trend was global, and OpenAI’s GPUs started overheating.

Top stories

  • OpenAI ships native 4o image generation in ChatGPT and Sora. Rolls out as the default image generator for Plus, Pro, Team and Free users (Enterprise/Edu and API coming soon). GPT-4o handles image generation natively with world knowledge, conversational understanding, vastly improved text rendering, and editing far beyond DALL-E 3. Sora gains a dedicated image section. DALL-E 3 remains only via a dedicated DALL-E GPT. via OpenAI
  • Google releases Gemini 2.5 Pro Experimental — ‘most intelligent model yet.’ Co-credited to Demis Hassabis and Koray Kavukcuoglu, the 2.5 series is framed as the generation where all Gemini models reason before answering. Available immediately at no cost in Google AI Studio and to Gemini Advanced subscribers. 1M token context (2M coming soon). via Google
  • Gemini 2.5 Pro debuts #1 on LMArena at 1443 Elo. Beats Grok 3 and GPT-4.5 by close to 40 points. Tops Humanity’s Last Exam at 18.8% (no tools) vs OpenAI o3-mini 14% and Claude 3.7 Sonnet 8.9%. 86.7% AIME 2025, 84.0% GPQA Diamond, 63.8% SWE-Bench Verified. via 9to5Google
  • Studio Ghibli viral trend ignites within hours. Users immediately began generating Ghibli-style images with GPT-4o. Sam Altman swapped his X avatar for a Ghibli version. ChatGPT added one million users in a single hour during the peak. OpenAI added refusals for living-artist style requests within days. via Simon Willison
  • Nexthop AI exits stealth with $110M Series A. Led by Lightspeed Venture Partners; Kleiner Perkins, WestBridge, Battery and Emergent participated. Founded 2024 by Anshul Sadana (ex-Arista), builds custom networking hardware, network OS and pre-tested optical/electrical interconnects for hyperscalers’ AI buildouts — targeting a $35B market. via BusinessWire
  • OpenAI publishes 4o image-gen System Card addendum. Explicitly dated Mar 25. Details policy refusals (including living-artist style requests) and safety mitigations for the new image-output modality. via OpenAI

Who shipped

OpenAI made native multimodal image generation a baseline ChatGPT feature — the most consequential consumer-AI launch of Q1. Google DeepMind answered with Gemini 2.5 Pro, where reasoning is the default not the opt-in. Both labs ran the day end-to-end.

Open-source pulse

No major open-weights drop dated Mar 25 — frontier closed labs absorbed the news budget. DeepSeek-V3-0324 from Mar 24 was still circulating.

Money, infra & hardware

Beyond Nexthop’s $110M, the day’s other infra story was the CoreWeave IPO build-up: an amended S-1 disclosed a five-year ~$11.9B OpenAI cloud-computing contract with OpenAI receiving a $350M CoreWeave stake tied to the listing. via CoreWeave

Quiet corners

Anthropic, Meta and xAI all quiet on the day — no public product news. No notable EU/FTC regulatory action dated Mar 25.

By the numbers

  • 1443 Elo — Gemini 2.5 Pro LMArena debut score (#1 by ~40)
  • 18.8% / 86.7% / 84.0% / 63.8% — Humanity’s Last Exam / AIME / GPQA Diamond / SWE-Bench Verified
  • 1M+ — ChatGPT users added in a single hour at Ghibli peak
  • $110M — Nexthop AI Series A; targeting $35B market
  • ~$11.9B — CoreWeave’s disclosed five-year OpenAI cloud contract
  • Most-mentioned company: OpenAI

Compiled by AI Feed’s editor from verified web sources for 25 March 2025.