AI Roundup March 2025: GPT-4o's Ghibli Machine, Gemini 2.5 Pro, and OpenAI's $40B Mega-Raise

TL;DR

March was loud. OpenAI bolted native image generation into GPT-4o and the whole internet turned itself into a Studio Ghibli frame overnight. Google dropped Gemini 2.5 Pro and took the reasoning lead, then OpenAI closed a record $40B round at a $300B valuation on the last day of the month. But the story that actually matters for people who build: open weights had a monster month, with Gemma 3, DeepSeek V3-0324, and Alibaba's QwQ-32B all landing within weeks of each other.

GPT-4o Got Native Image Generation and Melted the Internet

In late March, OpenAI shipped native image generation inside GPT-4o, and it was not the usual incremental bump. This was image generation baked directly into the multimodal model, which means it actually understands context, follows multi-turn instructions, and (finally) renders legible text inside images instead of cursed alphabet soup. You could describe a scene, then say "make it night, add a cat, keep everyone's faces," and it just did it.

Then the Ghibli thing happened. Everyone fed it photos and got back soft, hand-drawn Studio Ghibli-style portraits, and the feature went so viral that OpenAI reported roughly 700 million images generated in a single week, about 1,200 every second. Sam Altman publicly begged people to chill because the GPUs were on fire.

Why it matters: this is the moment image generation stopped being a separate diffusion pipeline you prompt blindly and became a conversational, steerable part of an LLM. For builders, conversational editing with in-image text that actually renders correctly is the unlock that makes this usable in real product flows, not just meme generation.

Gemini 2.5 Pro Took the Reasoning Crown

On March 25, Google released Gemini 2.5 Pro Experimental, a "thinking" model that does chain-of-thought reasoning natively while keeping full multimodality and a 1 million token context window. It debuted at or near the top of several major benchmarks, and unlike a lot of reasoning models, it did not trade away the multimodal and long-context strengths Gemini was already known for.

That long context is the part builders should care about. A million tokens means you can throw an entire codebase, a stack of PDFs, or hours of transcripts at it in one shot without elaborate chunking and retrieval gymnastics. Google also shipped Gemini Robotics earlier in the month (March 12), a vision-language-action model built on Gemini 2.0 aimed at physical-world control.

Why it matters: the frontier race is now a three-horse sprint between OpenAI, Google, and Anthropic, and for once Google looked clearly ahead on raw capability for a moment. Reasoning plus a million-token window plus competitive pricing is a genuinely strong package for agentic and document-heavy work.

OpenAI Closed a Record $40B Round at a $300B Valuation

On March 31, OpenAI announced a $40 billion raise at a $300 billion post-money valuation, the largest private tech funding round on record by a wide margin. SoftBank led with $30 billion, joined by Microsoft, Coatue, Altimeter, and Thrive. The money is earmarked for frontier research and, mostly, compute.

There is a catch worth noting. SoftBank disclosed its total commitment could shrink to as low as $20 billion if OpenAI fails to restructure into a for-profit entity by the end of the year. So the full check is contingent on OpenAI untangling its unusual nonprofit-capped structure, which is its own ongoing saga.

Why it matters: the capital required to stay at the frontier is now openly absurd, and it is increasingly a compute-and-infrastructure game as much as a research one. For everyone building on top, it is a reminder of why the open-weight track below matters so much: not everyone can raise the GDP of a small country to train a model.

Anthropic Raised $3.5B Right After Shipping Claude 3.7 and Claude Code

On March 3, Anthropic closed a $3.5 billion Series E at a $61.5 billion post-money valuation, led by Lightspeed with a long list of co-investors including Salesforce Ventures, General Catalyst, and Jane Street. The raise landed right on the heels of Claude 3.7 Sonnet and Claude Code, with reported revenue run rate around $1 billion and climbing fast.

Why it matters: Claude Code in particular signaled where Anthropic is aiming, which is agentic coding as a first-class product, not a chat feature. If you write software, the terminal-native agent direction set the tone for a category that got very crowded very quickly after this.

Open Weights Had a Monster Month: Gemma 3, DeepSeek V3-0324, and QwQ-32B

While the headlines went to closed frontier models and giant checks, the most useful month-over-month progress for self-hosters came from open weights.

• Gemma 3 (March 12): Google's open family at 1B, 4B, 12B, and 27B, multimodal, with a 128K context window and multilingual support. The 27B punches way above its weight, trading blows with models several times its size while fitting on a single high-end GPU.
• QwQ-32B (early March): Alibaba's 32B reasoning model that claimed performance rivaling DeepSeek R1, a 671B model, by scaling reinforcement learning on a much smaller base. It talks a lot, but it leads on coding, math, and reasoning for its size.
• DeepSeek V3-0324 (late March): a refreshed V3 with an improved post-training pipeline borrowing tricks from R1, sharper reasoning, coding, and tool use, released open-weight for free commercial use.

Why it matters: this is the good stuff if you run AI on your own hardware. A 27B or 32B model that competes with last year's giants means real reasoning and coding capability you can run locally, fine-tune, and ship without renting frontier API access. The gap between "open and on my box" and "closed and behind a meter" kept shrinking, and March was a big step.

Key Takeaways

• Image generation joined the LLM: GPT-4o made image gen conversational and context-aware, which turns it from a novelty into a real building block for product flows.
• Reasoning plus long context is the new bar: Gemini 2.5 Pro paired top-tier reasoning with a million-token window, raising the floor for agentic and document-heavy apps.
• Frontier money got surreal: OpenAI's $40B round shows staying at the bleeding edge is now a compute-infrastructure arms race, not just a research one.
• Agentic coding went mainstream: Anthropic's Claude Code launch and fresh $3.5B set the template for a category that exploded right after.
• Open weights are closing the gap fast: Gemma 3, QwQ-32B, and DeepSeek V3-0324 gave self-hosters frontier-adjacent reasoning on a single GPU, no API meter required.