GPT-6 'Spud' Official: What's Behind OpenAI's Ambitions

OpenAI picked an interesting moment for this announcement.

Yesterday, GPT-6 became official. Internally codenamed ‘Spud’—yes, like the potato—which made me do a double take. You name your trillion-parameter flagship model after a root vegetable?

But reading through the release docs, I get it. This isn’t a potato. It’s a bunker buster.

The Numbers: Parameters and Training

OpenAI didn’t disclose exact figures, but industry consensus puts GPT-6 at 8-10 trillion parameters. Compare that to GPT-5.4’s roughly 1.8 trillion, and you’re looking at a genuine generational leap.

More significant is the training approach. OpenAI used what they call ‘multimodal-native’ architecture—not text first with vision bolted on later, but all modalities trained together from scratch. What does this mean?

It means GPT-6 understands the world more like humans do. We don’t learn text first and images second; we absorb everything simultaneously.

On training costs, The Information reports GPT-6’s single training run cost between $1.5-2 billion. For perspective, that’s equivalent to burning two unicorn startups to the ground.

Technical Breakthroughs: Beyond ‘Bigger’

OpenAI highlighted two capabilities:

First, cross-modal reasoning.

Previous models described images when you asked about them. GPT-6 understands scenes. In the demo video, there’s an example: a messy kitchen photo with the question ‘If I wanted to cook something using the oven, what should I clean first?’

The model pointed out oven debris, then reasoned that ‘the sink piled with dirty dishes will limit counter space.’ That causal chain represents a clear step up.

Second, proactive tool use.

GPT-5.4 could call tools, but basically waited for instructions. GPT-6 starts deciding which tools it needs to achieve goals.

In the official demo, a user says: ‘Plan a 3-day Kyoto trip with temples and food, budget 5000 RMB.’

Instead of immediately writing an itinerary, GPT-6 asks: ‘Do you prefer traditional ryokans or modern hotels? How do you feel about raw food?’

This matters. It’s not blindly executing—it’s actively gathering information to optimize output.

But I Have Questions

After the announcement, one thought kept circling: With a model this capable, how does OpenAI plan to release it?

GPT-5.4 was already expensive. If GPT-6 follows current pricing logic, API costs could be 3-5x higher. Can ordinary developers afford that?

Latency is another concern. Larger parameters mean slower inference. What about real-time scenarios like voice assistants? OpenAI acknowledged this, mentioning ‘inference optimization’ and ‘tiered services’—simple tasks use smaller models, complex ones wake up the big guns.

The logic makes sense, but execution remains to be seen.

The Ambition Behind Spud

I suspect the codename signals something: this is OpenAI’s ‘staple crop.’

Fresh off $122 billion in funding at an $852 billion valuation, where does that money go? GPT-6 is the answer. OpenAI is going all-in on one direction: the path to AGI.

I’m not an AGI optimist, but this release felt different. Not just ‘stronger again’—but ‘stronger differently.’ Multimodal-native training, proactive tool use, contextual reasoning—these aren’t incremental improvements. They’re qualitative shifts.

Of course, the model isn’t open yet. When I can test it hands-on, I’ll tell you what I actually find.