On June 25, The Information reported that OpenAI CEO Sam Altman told employees the U.S. government wants GPT-5.6 released first to roughly 20 trusted partners via Amazon Bedrock, with government agencies reviewing each partner individually before granting access1. A broader public release could follow weeks later if the preview goes smoothly1.
The arrangement builds on a June 2 executive order that gave the government 30-day pre-release access to frontier models and established a “trusted partner” priority framework2. The order explicitly stated it “does not authorize a mandatory government licensing or approval regime”2. The GPT-5.6 process appears to have moved past voluntary coordination into something more structured.
On June 13, Zhipu AI pushed GLM-5.2 to every Coding Plan subscriber. Lite, Pro, Max, and Team tiers all received access3. Three to four days later, on June 16 and 17, Zhipu published the model weights under an MIT license on HuggingFace and ModelScope. Anyone can download them, modify them, and deploy them commercially3.
One model needs government sign-off for each user. The other needs an internet connection.
GPT-5.6: A Staged Release With Government Review
GPT-5.6 is OpenAI’s next-generation flagship. It has not been officially announced, and the June 25 reporting suggests it will not launch the way previous models did1.
Altman told staff that U.S. government agencies are growing more concerned about the capabilities of the most advanced models1. The Office of the National Cyber Director (ONCD) and the Office of Science and Technology Policy (OSTP) are involved in the partner review process1. Altman emphasized that employees need to work with the government even when the company disagrees with its positions1.
The review operates on a per-customer basis. Government agencies examine each organization that wants GPT-5.6 access, and approval is required before Bedrock provisions the model1. OpenAI’s hope is that a successful preview period leads to wider availability within weeks1.
The framework for this arrangement traces back to the June 2 executive order, which directed the NSA to develop classification standards for “protected frontier models” within 60 days2. On June 12, the Commerce Department imposed export controls on Anthropic’s Fable 5 and Mythos 5, restricting access by nationality4.
Table 1: U.S. Frontier Model Availability Tightening 124
| Event | Restriction | Date |
|---|---|---|
| Trump EO on AI | 30-day government pre-access plus trusted partner priority | June 2, 2026 |
| Fable/Mythos export controls | Nationality-based access restrictions | June 12, 2026 |
| GPT-5.6 staged release | ~20 partners, per-customer government review | June 25, 2026 (reported) |
GLM-5.2: A Frontier Model With No Gatekeeper
Zhipu made GLM-5.2 available to all paying users on June 133. The model carries 744 billion total parameters with 40 billion active, and supports a 1 million token context window3.
On June 16 and 17, the weights went up on HuggingFace and ModelScope under MIT terms3. That means no regional restrictions, no usage tiers, no commercial licensing fees. Companies can fine-tune it, integrate it into products, or run it on private infrastructure without asking anyone for permission3.
On coding benchmarks, GLM-5.2 scored 62.1 on SWE-bench, above GPT-5.5’s 58.6. On FrontierSWE, it reached 74.4%, nearly matching Claude Opus 4.8 at 75.1%5. The South China Morning Post reported that U.S. entrepreneurs and researchers praised its coding performance and cost efficiency, with some calling it another “DeepSeek moment”5.
Zhipu’s announcement included a pointed statement: “Pure openness: MIT open-source license. No regional restrictions. Technology access knows no borders.”3
Table 2: GLM-5.2 Key Specifications 35
| Metric | Value |
|---|---|
| Total parameters | 744 billion (40 billion active) |
| Context window | 1 million tokens |
| License | MIT |
| SWE-bench | 62.1 (GPT-5.5: 58.6) |
| FrontierSWE | 74.4% (Claude Opus 4.8: 75.1%) |
| Architecture | IndexShare, 2.9x lower per-token FLOPs |
| Availability | Coding Plan tiers plus HuggingFace/ModelScope download |
European Companies Are Diversifying Their Model Portfolios
The split in model availability is pushing European enterprises to spread their bets.
At the VivaTech conference in Paris, Siemens Digital Industries CEO Cedrik Neike said: “You need flexibility. Sovereignty is often confused with self-sufficiency, but self-sufficiency is absolutely not the right approach.”6 Siemens currently runs DeepSeek, Alibaba’s Qwen, NVIDIA’s Nemotron, and various U.S. and European models6.
Renault uses Google, Microsoft, Mistral, DeepSeek, and Dataiku6. ChapsVision, which won French and German government contracts as a Palantir alternative, works with Mistral, Anthropic, OpenAI, and Qwen6. SAP and Sopra Steria treat resilience as a function of diversification, not isolation6. Capgemini notes that most AI vendors are adjusting their products to offer more local deployment options because the European market is too large to lose6.
Orange CEO Christel Heydemann framed the issue in operational terms: “Anthropic’s restrictions have made things extremely clear. If it wasn’t clear before, Europe must have AI services it controls and that can never be arbitrarily shut down.”6
OVHcloud CEO Octave Klaba put it more bluntly: “In open source, look at European models. They’re not great. U.S. companies used to open-source, then they went closed-source, and now the only open-source models left are Chinese.”6
Cost pressure adds to the urgency. An Orange executive said enterprises will be “extremely sensitive to per-token costs” by year-end, citing Uber as an example. Uber burned through its 2026 token budget in four months6.
Availability Is Becoming the Decisive Variable
Put the pieces together, and a clear pattern emerges. AI model availability is splitting into two tracks.
U.S. frontier models are getting harder to access. GPT-5.6 starts with 20 partners and government review. The Fable and Mythos bans restrict by nationality. The June 2 executive order created a pre-release review framework. These models are powerful, but whether you can use them depends on who you are, where you are, and whether you made the trusted partner list.
Chinese open-source models are getting easier to access. GLM-5.2 ships under MIT with no regional limits and no approval process. Kimi K2.7 uses a Modified MIT license7. Their performance sits at or near frontier level, and acquiring them requires nothing more than a download command.
Enterprises are responding pragmatically. Siemens, Renault, and Orange are not picking sides. They are making sure their AI services keep running regardless of which country’s policy changes next. They run U.S., Chinese, and European models side by side, and they use whichever one is available.
This is not a story about who wins. It is a story about who can get their hands on what. When the strongest model requires paperwork and the second-strongest requires a click, the industry does not stand still. It adapts.
What to Watch
- When will GPT-5.6 become available outside the initial 20 partners?
- On August 1, the 60-day deadline for “protected frontier model” classification standards arrives. What criteria will the NSA propose?
- How will the open-source GLM-5.2 perform on independent third-party benchmarks?
- Will multi-model portfolios become the default strategy for enterprises?
References
Conclusion
The gap between GPT-5.6 and GLM-5.2 is not just technical. It is about who can use them, when, and under what conditions. OpenAI’s staged release with government review treats frontier models as controlled resources. Zhipu’s MIT-licensed drop treats them as public infrastructure.
European enterprises are already adjusting. Their multi-model strategies are not political statements. They are operational insurance.
For the rest of the industry, the question is shifting. It used to be “which model is strongest?” Now it is also “which model can I actually run?” The answer to the second question is starting to matter as much as the first.
Footnotes
-
The Information / NDTV / Bloomberg — Trump Tightens Grip On AI, Asks OpenAI To Limit GPT-5.6 Release https://www.ndtv.com/world-news/donald-trump-tightens-grip-artificial-intelligence-asks-openai-to-limit-gpt-5-6-release-11689469 ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9
-
The White House — Promoting Advanced Artificial Intelligence Innovation and Security (Executive Order, June 2, 2026) https://www.whitehouse.gov/presidential-actions/2026/06/promoting-advanced-artificial-intelligence-innovation-and-security/ ↩ ↩2 ↩3 ↩4
-
Zhipu AI / Z.ai — GLM-5.2: Built for Long-Horizon Tasks https://z.ai/blog/glm-5.2 ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8
-
Fortune — Anthropic disables Fable and Mythos AI models following U.S. government export ban https://fortune.com/2026/06/13/anthropic-disables-fable-mythos-export-controls-national-security-threat/ ↩ ↩2
-
South China Morning Post — China’s Zhipu AI sparks new ‘DeepSeek moment’ with cost-effective coding model https://www.scmp.com/tech/big-tech/article/3358434/chinas-zhipu-ai-sparks-new-deepseek-moment-cost-effective-coding-model ↩ ↩2 ↩3
-
Reuters — US curbs on AI spur European firms to spread the risk https://www.zawya.com/en/news/insights/us-curbs-on-ai-spur-european-firms-to-spread-the-risk-bgaknvky ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9
-
CryptoBriefing — Kimi AI releases open-source K2.7 Code model with 1 trillion parameters https://cryptobriefing.com/kimi-k2-7-code-open-source-release/ ↩