OpenAI GPT-5.6 requires federal clearance

**Elon Musk** (0:02)
I'm not giving up.

**SPEAKER_3** (0:03)
I am selling the building.

**SPEAKER_4** (0:06)
The final season of FX is the Bear.

**SPEAKER_5** (0:10)
The restaurant is flooded.

**SPEAKER_6** (0:12)
Everything's either going to be okay.

**SPEAKER_4** (0:15)
No, stop.

**SPEAKER_5** (0:16)
Or not.

**SPEAKER_8** (0:18)
We are outgunned and we are outmanned, but we have each other.

**SPEAKER_4** (0:24)
FX is the Bear, the final season.

**SPEAKER_9** (0:26)
All episodes now streaming on Disney Plus.

**SPEAKER_5** (0:29)
Welcome to Sephora.

**SPEAKER_10** (0:31)
I'm looking for a perfume that's not too perfumey.

**SPEAKER_7** (0:35)
I got you.

**Elon Musk** (0:35)
Serum moisturizer or moisturizer syrup.

**SPEAKER_7** (0:39)
Let's get into layering.

**SPEAKER_4** (0:40)
My concealer is making me look worse.

**SPEAKER_7** (0:44)
Sounds like the wrong shade.

**SPEAKER_11** (0:46)
Let's get you matched.

**SPEAKER_5** (0:47)
There's only one store that really gets what you're going for.

**SPEAKER_6** (0:51)
Get beauty from people who get beauty. Only at Sephora.
Hi.

**SPEAKER_8** (0:56)
Let's get you a basket.

**SPEAKER_11** (1:01)
This episode is brought to you by Subaru. Go further in a long-range Subaru Hybrid with up to 581 miles per tank in the Forester Hybrid. Longer range, better fuel efficiency, and legendary symmetrical all-wheel drive, standard. The Subaru Forester Hybrid. Visit subaru.com/hybrid to learn more. Maximum range based on EPA estimated combined fuel economy to full tank of fuel. Actual mileage and range may vary.

**SPEAKER_12** (1:29)
OpenAI released its GPT 5.6 model family, not to the public, but to a small handpicked list of trusted partners directly vetted by the United States government.

**SPEAKER_7** (1:41)
Yeah, we are watching a commercial software launch being handled exactly like the transfer of restricted military infrastructure. I mean, you have to view this through the lens of what happened just days prior.

**SPEAKER_12** (1:51)
With Anthropic.

**SPEAKER_7** (1:52)
Right, when Anthropic completely disabled its own advanced models worldwide due to a federal export directive. So, how does the tech industry actually adapt when the most capable software tools require federal clearance before reaching a developer's machine?

**SPEAKER_12** (2:07)
Well, the structure of what OpenAI built with the 5.6 family reflects that exact pressure.
They split the release into three distinct tiers. You have Sol, which is the flagship model, then Terra is the balanced tier, operating at half the cost of the previous generation, and Luna is the volume tier, designed for really fast, low cost work.

**SPEAKER_7** (2:26)
Because the economics basically dictate the entire architecture of how you build software now. I mean, Sol costs $5 per million input tokens, and $30 per million output tokens, whereas Luna drops to $1 for input and $6 for output. So enterprise adoption is no longer an exercise in finding the smartest model.

**SPEAKER_12** (2:44)
It's an operations problem.

**SPEAKER_7** (2:45)
Exactly. It's an operations and routing problem. If you send every single request, your company processes to Sol, you will bankrupt your IT department by Tuesday.

**SPEAKER_12** (2:56)
Yeah. Developers had to design applications that actually look at an incoming task, route the easy classification or simple data extraction to Luna, and then reserve Sol strictly for complex high-stakes debugging.

**SPEAKER_7** (3:09)
The caching rules completely complicate that operational math.

**SPEAKER_12** (3:13)
Right. Because OpenAI set the cache writes at 1.25 times the uncached input rate.

**SPEAKER_7** (3:18)
Yes.

**SPEAKER_12** (3:19)
But the cached reads get a 90 percent discount.

**SPEAKER_7** (3:21)
So think about what that actually looks like for someone running a long context workflow. Say you manage a legal department and you want to feed an entire policy document or massive regulatory frameworks into the model or a complete corporate code base.

**SPEAKER_12** (3:34)
The initial read penalizes you.

**SPEAKER_7** (3:35)
It does. You pay a premium to load that context into the cache because you are essentially occupying dedicated memory on their server cluster. But once it sits in the cache, every subsequent query against that stored context costs a fraction of a cent.

**SPEAKER_12** (3:50)
Which forces developers to manage token retention the exact same way they handle memory allocation in traditional programming. You push the data in, pay the upfront penalty, and then query against it continuously for hours.

**SPEAKER_7** (4:03)
Yeah.

**SPEAKER_12** (4:04)
It operates very similarly to cloud storage architectures. Like with cloud infrastructure, data naturally flows between hot, warm, and cold storage based on utility.

**SPEAKER_7** (4:13)
You keep the data you access every second in the expensive, fast tier.

**SPEAKER_12** (4:17)
Exactly. And push the archive data you rarely touch to the cheap, slow tier. The tiered models combined with prompt caching replicate that exact structure just for cognitive processing instead of static data.

**SPEAKER_7** (4:29)
The problem with applying that storage logic to artificial intelligence, though, is the ambiguity of the input. Well, with static data, an engineer knows exactly what a file is, how large it is, and how often a user needs it.