🍭 OpenAI Got Washington's OK 🏛️

Good morning.

OpenAI shipped GPT-5.6.
Grok 4.5 also dropped.

One of them had to check with Washington first.

Let’s dive in 👇

🍭 What’s Cookin’:

• OpenAI got the government's blessing (GPT-5.6 is live)

• SpaceXAI dropped Grok 4.5 same day. Born blessed

• Anthropic brought Cowork to your phone

• EU regulators gave 110 banks until Halloween to fix their AI cyber gaps

OpenAI's
🙏 GPT-5.6 Is Out. Ask Washington.

The Bite:

GPT-5.6 is a model family split in three: Sol, Terra, and Luna. It was launched publicly today across ChatGPT, Codex, and the OpenAI API.

The models were first previewed on June 26 to roughly 20 government-vetted organizations, at the Trump administration's request.

The Department of Commerce's Center for AI Standards and Innovation ran additional testing before clearing the broader release.

OpenAI sent technical experts to Washington to answer questions in real time.

Sol is the flagship tier, pricing at $5/$30 per million tokens;
Terra at $2.50/$15;
Luna at $1/$6.

OpenAI said it doesn't believe government access "should become the long-term default".

Snacks:

• Sol scored 96.7% on OpenAI's internal cyberattack evaluation, with ExploitBench performance comparable to Anthropic's classified model Mythos Preview at roughly one third of the cost.

• The 12 days of gating period was nominally voluntary under Trump's June executive order, which bars mandatory preclearance.

• Grok 4.5 launched the same day with no government review period disclosed.

• The August 1 deadline requires the National Security Agency to finalize a formal framework designating "covered frontier models". Whatever that classification produces will define who gets gated next.

Why it Bites:

OpenAI called this voluntary. The executive order bars mandatory preclearance.

No licenses required, no approvals needed, just a polite ask from Washington, and twelve days of quiet.

Voluntary frameworks with consequences aren't voluntary.

When the alternative is a Commerce Department directive like the one that pulled Fable 5 offline for 19 days, "cooperating" becomes the only way.

The executive order was sold as innovation-friendly, with no gatekeepers, no preclearance theater.

But what it produced, in practice, is an informal review process with no published criteria, no defined timeline, and no public record of what was tested or what passed.

OpenAI got 12 days. Anthropic got a shutdown. SpaceXAI got a blink.

The difference between those still hasn’t been explained.

If the framework's goal is safety, the question it still hasn't answered is why it applies to some frontier models and not others.

And if the answer turns out to be relationships, that's not a safety framework.

Pick your own word for it.

Steal This Prompt
🚞 Vintage Posters From a Design Museum

Turn any photo into a bold, vintage two-color illustration that looks ripped straight out of a retro screen-printed poster.

Use it to:

• Transform ordinary photos into gallery-worthy vintage prints

• Give your car, motorcycle, or travel photos retro-cool vibes

• Create bold branding, merch, wallpapers, and social graphics

Workflow:

1. Hit this link: Vintage Dual-Color Art

2. Paste into your AI model

3. Replace the #s with your image details or color preferences

4. Watch it cook into a vintage masterpiece

   Try the prompt on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🛠️ Timbal AI

One platform for the full AI app lifecycle: agents, workflows, retrieval, observability, and evals without stitching five tools together.

📈 Auriko

Routes each request across providers by token price, latency, and cache behavior, averaging 30% savings.

💬 Toyo

An exec assistant that lives in iMessage and calls your phone, triages your inbox, preps you for meetings, and pulls context from your company tools over text or voice.

🔐 Opper AI

One API key to 300+ EU-hosted models, zero prompt storage by default, and a base URL change away from your existing OpenAI or Anthropic SDK.

🔌 Blocks.ai

PubNub's control plane for connecting AI agents across any framework or provider without opening ports, changing DNS, or touching your firewall.

_{Everything Else}
🧠 You Need to Know

🦾 SpaceXAI Launches Grok 4.5 For Developers And Enterprise
→ The first SpaceXAI model since going public claims Opus-class performance at $2/$6 per million tokens (roughly one-third the cost of Claude Opus 4.8), co-trained with Cursor's coding data.

📱 Anthropic Expands Claude Cowork To Web And Mobile
→ Cowork sessions now run in the cloud and persist across devices, with scheduled tasks completing overnight after the laptop closes.

🏦 ESRB Rates Frontier AI Cyber Risk "Severe" For EU Banks
→ The European Systemic Risk Board warned that frontier AI models can build working exploits in minutes, with the ECB giving 110 banks until October 31 to file AI cyber action plans.

💾 SK Hynix Begins Nasdaq Trading In Largest ADR Listing In History
→ Nvidia's primary memory chip supplier priced its ADRs yesterday and begins when-issued trading today under SKHY, targeting ~$29 billion.

🤖 OpenAI Launches GPT-5.6 Publicly After 12-Day Government Gate
→ Sol, Terra, and Luna are live across ChatGPT, Codex, and the API after Commerce Department testing ended a 12-day window limited to ~20 government-vetted organizations.