GPT-5.6 Now Behind the Government Gate

OpenAI released GPT-5.6 in three variants, DeepSeek open-sourced a framework for speculative decoding, and Mythos 5 returned to US critical infrastructure after 15 days of blackout. In this edition: what the era of government-gated models means.

Intro · Alex

This week will go down in history as the moment when government-gated AI stopped being a theory and became a fact. OpenAI released GPT-5.6 — but the preview went to only about 20 US administration-approved partners. Meanwhile, DeepSeek did the opposite: it gave the community code for speculative decoding that actually accelerates inference by 60–85% without buying new GPUs. Two worlds: one closes, the other opens. In today's edition I've detailed the three GPT-5.6 variants (Sol, Terra, Luna) with prices and benchmarks, DeepSeek's DSpark framework, the story of the partial restoration of Mythos 5 after a government blockade, European officials' warnings about pro-Russian propaganda sown in chatbots, and the EU AI Act deadline — 34 days to full enforcement. Plus a tip of the day on auditing model dependencies in case your model disappears overnight.

What's worth knowing

OpenAI Releases GPT-5.6 — Sol, Terra and Luna Behind the Government Gate

OpenAI has unveiled the GPT-5.6 family in three variants: Sol (flagship, $5/$30 per 1M tokens, 91.9% on Terminal-Bench 2.1), Terra (balanced, $2.50/$15) and Luna (fast/cheaper, $1/$6). This is the first OpenAI model subjected to a 30-day government quarantine under the Executive Order — the preview went to about 20 approved partners. The era in which access to a model depends on Washington's approval, not an API key, has just begun.

DeepSeek Releases DSpark — +85% Inference Speed in Open Source

DeepSeek has released DSpark — a framework for speculative decoding that accelerates V4-Flash by 60–85% without retraining, changing weights, or new hardware. Along with the framework, DeepSpec was published under the MIT license — a complete training stack for applying this technique in Qwen3 and Gemma models. Anyone with their own GPU can cut latency in half today.

US Government Partially Restores Claude Mythos 5 for Critical Infrastructure

After 15 days of complete blockade, the US government allowed Anthropic to restore Mythos 5 for about 100 organizations defending US critical infrastructure. Fable 5 remains offline. This is a precedent — the government first shut down an AI model, then partially restored it. For companies building on Anthropic's API, it's a signal that model access is a political risk, not just a technical one.

Pro-Russian Groups Plant Propaganda in AI Chatbots — European Officials Warn

European officials, including Estonia's Minister for Digital Affairs Liisa-Ly Pakosta, have warned that pro-Russian groups are systematically manipulating AI models' knowledge bases and prompts to generate responses aligned with Kremlin narratives. This is a new form of hybrid information attack — propaganda enters the knowledge layer of the models themselves, not just the content they generate.

EU AI Act Takes Effect August 2 — 34 Days to Deadline

The EU AI Act enters full enforcement phase in less than 5 weeks. Article 13 (transparency obligations for high-risk systems) and provisions on prohibited practices and AI literacy will be legally binding from August 2. Companies using AI in recruitment, credit scoring, or insurance must have risk mapping, governance policies, and transparency documentation ready. Penalties: up to €35 million or 7% of global revenue.

From the tech world

Operation Endgame Dismantles Three Malware Networks — 27 Million Logins and $47 Million Recovered

A global law enforcement operation with Microsoft, Europol, ESET and IBM X-Force simultaneously dismantled the Amadey, Stealc and SocGholish networks. 27 million stolen logins and $47 million in cryptocurrency were recovered. This demonstrates the real effectiveness of international action against the "malware assembly line" — worth checking if your own infrastructure was in the crosshairs of these botnets.

Tip of the day

Model Dependency Audit — A Plan for a Rainy Day

The era of government-gated AI means that access to frontier models can change overnight — politics, regulations, or national security can cut off your API without warning. Before you build a critical system on a single model, map out the entire dependency chain: what happens if that model disappears, gets restricted, or its price triples? Document which capabilities are critical and which are 'nice to have.' Practical step: create a 'model fallback matrix' for every production flow. For each use case, write down: (1) the specific capability you need, (2) the cheapest open-source equivalent that does 80% of the job, (3) the switching cost. Test at least one alternative model this week. When access is cut off, you won't have time for experiments.

Tool of the issue

DSpark from DeepSeek — Speculative Decoding That Accelerates Inference Without GPU Replacement

A framework for speculative decoding that accelerates model response generation by 60–85% without retraining and without changing hardware. Along with DSpark, DeepSeek published DeepSpec under the MIT license — a training stack for applying the same technique in Qwen3 and Gemma models.

Reading list

GPT-5.6 Launch: Sol, Terra, Luna — Benchmarks and Prices

A detailed breakdown of Terminal-Bench 2.1 benchmarks for all three GPT-5.6 variants with comparisons to Claude Mythos 5 and Gemini 3.1 Pro. Concrete numbers, context, and analysis of where the new model actually outperforms the competition.

DeepSeek DSpark — How Speculative Decoding Works and Why It's a Game Changer

An accessible explanation of the speculative decoding technique and how DeepSeek achieved 85% speedup without changing the model. Includes diagrams and before/after comparison — practical reading for anyone considering implementation.

The more models go through government gates, the more obvious it becomes that the biggest risk in AI is not the model — it's who controls access to it.

Want the next issue in your inbox?

Subscribe — every issue delivered directly to you.

← Previous issue Next issue →

← AI Newsletter Digest