AI Weekly #26/2026: When the State Stops the AI Launch

TL;DR

This week in 30 seconds:

• GPT-5.6 Launch: US government intervenes preventively for the first time — only ~20 pre-approved partners receive access instead of everyone
• Anthropic comeback: Mythos-5 export ban lifts after 14 days — 100+ US agencies and companies are unlocked again
• Vacuum effect: Japanese (Sakana AI) and Chinese startups (Zhipu AI) launch frontier models — the export ban accelerates exactly the competition it aims to prevent
• Content deal: Getty stock jumps +145% intraday after ChatGPT partnership — fair image licensing in AI systems becomes a business model

---

Audio Version

13:53 | Download MP3

Chapters

- 0:00 - TL;DR - 0:45 - Story of the Week - 3:45 - More Top Stories - 7:53 - Quick Hits - 8:45 - Tool of the Week - 9:56 - Fail of the Week - 11:50 - Reading List - 12:49 - Next Week

Read aloud with edge-tts (en-US-AndrewNeural)

---

Story of the Week

Historic First: US Government Stops GPT-5.6 — Before It Launches

For the first time in US history, a government has actively restricted the release of a commercial AI model — before its launch [1]. The Office of the National Cyber Director and the OSTP asked OpenAI to release GPT-5.6 only to a small group of pre-approved partners [2]. Not after an incident. Not after a disaster. Preemptively.

GPT-5.6 comprises three models: Sol (flagship, $5 / 1M input tokens, $30 / 1M output tokens), Terra (balanced, half the price of Sol), and Luna (fast and affordable, $1 / 1M input, $6 / 1M output) [1]. Only around 20 companies receive access for now — each additional unlock must be individually approved by US agencies [1].

“We don’t believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.” [1]

— OpenAI

The precedent that triggered this process is the complete export ban on Anthropic’s Fable 5, which took effect on June 12 [2]. The White House has since been working on a framework for safety evaluation of new frontier models and requires voluntary reviews 30 days before release [2].

Why is the government intervening? Officially the focus is on national security and dual-use risks: frontier models with strong coding capabilities could be used for cyberattacks, bioweapon development, or by state adversaries — according to the security agencies’ rationale. The framework aims to ensure risk assessments happen before launch rather than after.

Dean Ball, former White House AI advisor and future OpenAI employee, warns from a libertarian policy perspective — and thus as a party with a clear vested interest — that Trump’s Executive Order establishes a de-facto licensing requirement for frontier AI that could lead to systematic delays [1].

For developers and businesses this means: Frontier AI access is no longer a right but a governmentally controllable privilege. Anyone building on OpenAI or Anthropic today must now factor in not only pricing risks — but availability risks from government decisions. Multi-provider strategies are no longer a precaution but a necessity.

Critical voices: Observers question whether preventive controls without a clear legal basis are constitutionally viable and whether the process is transparent enough for the affected companies. OpenAI itself is cooperating — but explicitly emphasizes that this must not become a permanent state of affairs [1].

Bottom line: The US government has fired the starting gun: anyone developing frontier AI will need government approval going forward — and OpenAI played along for the first time without blocking.

---

More Top Stories

Export Ban Reversal: Anthropic’s Mythos 5 Returns — for the Select Few

Fourteen days after the complete export ban on Anthropic’s Fable 5 / Mythos 5, Commerce Secretary Howard Lutnick lifts the restrictions partially via letter [3]. More than 100 US government agencies and critical infrastructure companies regain access — including for non-American employees of those organizations [3].

“I have determined that appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model.” [3]

— Howard Lutnick, Commerce Secretary

The complete export ban for the general public and international users, however, remains in place. Anthropic, with an ARR of $47 billion in May 2026 (according to industry reports, not independently verified), had been negotiating with the US government since June 12 [3]. The model was simply unavailable to 99% of users for fourteen days — a substantial economic damage that brutally exposes the dependency on government decisions.

The pattern becomes clear: government access first, general public later (maybe) — and international partners at the very back.

---

The Vacuum Paradox: Export Bans Breed the Competition They’re Meant to Fight

While Fable 5 remains blocked for international users, Japanese and Chinese startups are launching competing models [4]. Sakana AI from Tokyo presents “Fugu” — positioned as on par with “Fable 5 and Mythos Preview,” optimized for agent orchestration and Japanese [4]. Sakana, valued at $2.65 billion (according to company statements, Series B, November 2025) with $135 million in funding, is no longer a garage startup [4].

“Access to top models can disappear overnight.” [4]

— David Ha, Sakana CEO

In parallel, Zhipu AI (Tsinghua spinout) from China brings the models “Tulongfeng” and “Yitianzhen” — specialized in cybersecurity and vulnerability detection [4]. The irony is hard to miss: US export bans, designed to secure technological leadership, are accelerating exactly the international development of frontier alternatives [4].

An important distinction: Japanese alternatives like Sakana AI are considered geopolitically unproblematic for Western allies — Chinese cybersecurity models like those from Zhipu AI are a different category from a US security perspective. The export ban treats both equally, but creates very different risks.

Sakana co-founder Ren Ito puts it plainly: “The first priority should be preserving access for America’s closest allies.” [4]

---

Getty + OpenAI: The First Major Precedent for Fair Content Compensation in AI

Getty Images and OpenAI are entering a multi-year display partnership [5]. When ChatGPT can illustrate a response with a real image, a licensed Getty photo appears — with attribution, without training use, with indemnification for Getty customers [5]. Around 600,000 content creators and 360 content partners are affected, with coverage spanning 160,000+ events per year according to Getty [5].

“High-quality, licensed visual content makes AI-powered search more trustworthy.” [5]

— Craig Peters, CEO Getty Images

Caveat: The model applies only to Getty-represented creators. The majority of photographers worldwide — who don’t license through Getty — receive no compensation and no attribution when their images end up in AI training data.

The market responded decisively: an intraday price surge of up to 145%, closing price approximately +90% [5]. This is no coincidence — it’s a signal that investors take fair licensing seriously as a scalable AI business model. For the entire content industry — music, text, video — this is a precedent: not bans or lawsuits, but partnership models could be the way forward.

---

Quick Hits

Briefly noted:

• Claude vs. ChatGPT: ChatGPT’s market share among paying users fell below 50% for the first time (46.4%); Gemini holds 27.7%, Claude 10.3% — among premium users, Claude scores disproportionately through coding strength and quality [6]
• Adobe + Topaz: Adobe acquires Topaz Labs, known for AI-based image and video upscaling and noise reduction — a direct attack on generative AI competition in the creative market via the Firefly ecosystem [7]
• AI Energy Shift: Former Databricks AI chief founds startup with a promise: 1,000x reduction in AI energy consumption through novel hardware architecture — concrete details are still pending, but the political nerve has been struck [8]

---

Tool of the Week

HuggingFace Jobs — Private vLLM Inference Server with a Single Command

HuggingFace enables launching a private, OpenAI-compatible LLM inference endpoint on HF infrastructure — no own GPU, no Kubernetes, no idle costs [9]. A single command is all it takes:

hf jobs run --flavor a10g-large --expose 8000 --timeout 2h \
  vllm/vllm-openai:latest \
  vllm serve Qwen/Qwen3-4B --host 0.0.0.0 --port 8000

Pay-per-second billing, various GPU flavors (A10G, etc.), authentication via HF token out-of-the-box, and support for large models via tensor parallelism (e.g. Qwen3.5-122B) [9]. The feature is a practical counterpoint to the closed-source chip strategies of OpenAI and Google: anyone who needs open-source models with high-performance inference without wanting to operate their own infrastructure gets a ready-to-deploy solution here.

→ HuggingFace vLLM Jobs Documentation

---

Fail of the Week

“Orbital AI Data Centers by 2028” — when even SoftBank’s Masayoshi Son hits the brakes

SoftBank CEO Masayoshi Son — known for billion-dollar bets on wild ideas like WeWork and Uber — publicly questions the feasibility of Elon Musk’s plans for orbital AI data centers [10]. Sam Altman also expresses skepticism. When the man who invented the “Vision Fund” pumps the brakes, a second look is warranted.

Root cause: The next two to three years are decisive in the AI race — and orbital-based solutions simply don’t help with the acute infrastructure bottleneck. There’s also a structural problem: more orbital data centers means more SpaceX launches, means more revenue for Musk. When the seller promotes the product, in English that’s called “Talking Your Book” [10].

What we learn: Infrastructure hype is dangerous when timelines and feasibility remain unclear. When evaluating any bold infrastructure vision, ask first: who benefits from the hype — and are the assumptions physically and economically sound?

---

Number of the Week

+145%

That’s how much the Getty Images stock rose intraday on June 21 after the display partnership with OpenAI was announced (closing price: approximately +90%) [5]. This is the largest single-day gain by an established media and content company following an AI partnership — and a clear market signal that licensing models are taken seriously as a sustainable bridge between the content industry and AI.

The signal is unambiguous: the market believes not only that AI and content can coexist — it believes that licensing models will revalue the content industry. For everyone who creates and licenses content, this is a strong sign that fair partnership with AI companies is economic value, not capitulation.

---

Reading List

For the weekend:

1. OpenAI limits GPT-5.6 rollout after government request — The full TechCrunch article with direct quotes from OpenAI and Dean Ball explains the new control regime better than any summary. Essential reading for anyone planning API dependencies. (8 min)
2. Run a vLLM Server with a Single Command on Hugging Face — Hands-on guide with example commands for various models and GPU configurations. Anyone looking to build their own inference endpoints will find the most practical entry point of the week here. (12 min)
3. Asian AI startups launch Mythos-like models as export ban drags on — Explains in detail how Sakana AI and Zhipu AI are exploiting the export ban vacuum — including benchmark claims and investor context. Anyone wanting to understand the geopolitical dimension of AI regulation starts here. (10 min)

---

Next Week

What’s coming:

• GPT-5.6 Rollout Watch: The ~20 approved partners go live in the first days — first real benchmark comparisons with Anthropic Mythos 5 and Gemini 2.5 Ultra are expected
• Anthropic Export Ban Week 27: Will the Fable-5 ban for the general public remain in place, or does the next relaxation follow? The Commerce Department has not communicated a timeline
• Getty Partnership Live: The ChatGPT-Getty display rollout begins — first user experiences and creator reactions will show whether the compensation model works in practice
• Sakana Fugu API: Sakana AI has announced API access for Fugu — availability and pricing for non-Japanese developers pending

---

🤖 Behind This Newsletter

Generated in: ~35 minutes (Phase 1–3) + Validation + Correction Sources scanned: 18 verified stories from 6 feeds Stories found: 18 → 10 selected → 6 main categories Validation: 4 agents in parallel ✓ — Status: FIX_NEEDED → corrected Model: Claude Sonnet 4.6 + Haiku (Validation) Images: Pollinations.ai (Phase 3.5 pending)

Full Metrics

Phase	Metric	Value
Source collection	RSS feeds	6
Source collection	WebSearch queries	4
Selection	Stories presented	18
Selection	Stories selected	10
Draft	Words	~1,500
Draft	Sources cited	10
Validation	Fact-check issues	1 BLOCKER, 5 WARN, 1 INFO
Validation	Balance issues	2 BLOCKER, 3 WARN, 1 INFO
Validation	Quality issues	1 BLOCKER, 1 WARN, 3 INFO
Validation	Legal issues	0 BLOCKER, 2 WARN, 4 INFO

---

This newsletter was researched and written AI-assisted. Images generated with Pollinations.ai.

Financial disclaimer: Price data and financial figures in this newsletter are for informational purposes only and do not constitute investment advice.