OpenAI's Three-Tier Model Family Is Now Complete. Here's What GPT-6 Sol and Luna's 50% Price Cut Means for Every Enterprise AI Budget.

At 10 a.m. New York time on September 23, 2026, the UN Security Council convenes its first dedicated session on artificial intelligence, chaired by French Minister Jean-Noël Barrot under the "Maintenance of international peace and security" agenda item. OpenAI CEO Sam Altman will be at the UN in person. Anthropic CEO Dario Amodei will join remotely. Hugging Face CEO Clément Delangue and Turing Award winner Yoshua Bengio — co-chair of the UN's Independent International Scientific Panel on AI — will also present. DeepSeek and Chinese AI startup Moonshot have been invited to make statements. It is the first time the 15-member Council has directly hosted frontier AI developers — American and Chinese — in a dedicated session on the technology's risks to international security.

Two days later, on September 25, President Trump meets Chinese President Xi Jinping in a bilateral summit where AI hardware export controls, chip restrictions, and bilateral AI research cooperation are central agenda items. The timing is not incidental. The session that opens today is setting the diplomatic terms of a conversation that the Trump administration will carry into that summit.

What Makes September 23 Different From Every AI Panel Before It

The UN has been discussing AI governance for years. The General Assembly passed a non-binding AI resolution in 2024. The G7 Hiroshima process produced voluntary codes of conduct for advanced AI developers. Bletchley, Seoul, and Paris each generated communiqués, commitments, and convening records. None of it materially changed the behavior of frontier AI developers or established binding rules that any government was required to enforce.

The Security Council is structurally different. It is the UN's primary enforcement body, with Chapter VII powers to authorize mandatory compliance measures. It is the forum where great power conflicts over dual-use technology have historically been negotiated — the Nuclear Non-Proliferation Treaty, the Chemical Weapons Convention, and biological weapons governance all moved through Security Council frameworks before binding international rules existed. The decision to convene AI developers in a Security Council session, rather than a General Assembly working group or a GPAI roundtable, signals that at least some member states have concluded that advanced AI poses a security risk serious enough to warrant the Council's jurisdictional authority.

That does not mean binding rules are coming immediately. China and Russia hold permanent seats with veto power, and both countries' AI development priorities diverge from the regulatory preferences of the US, UK, and France. A Security Council resolution on AI governance would require either unanimous P5 agreement or abstention from dissenting permanent members — both difficult to achieve on a topic where US-China AI competition is explicit and deepening. But the venue itself — the Security Council, not the ITU or UNESCO — changes the status of the governance conversation in ways that enterprise AI compliance functions should begin to track now.

The Participants and What Each Brings

The participant list rewards careful reading. Yoshua Bengio provides scientific authority that no founder or CEO can claim — his advocacy for binding AI safety standards, including calls for a global treaty on frontier AI, has been consistent and credible. Sam Altman and Dario Amodei together represent the companies that hold roughly 67% of enterprise LLM API spend as of mid-2026. Clément Delangue's inclusion explicitly represents the open-source ecosystem — not just as an afterthought, but as a counterweight to proprietary frontier lab positions. The Chinese AI developers' presence is structurally new: this is the first time both sides of the US-China AI competition have appeared in the same formal international security venue with an explicit mandate to discuss shared risks.

OpenAI's Pre-Brief Proposal: US-Led Standards for RSI

On September 21, two days before the Council session, OpenAI published a policy proposal calling for the United States to lead an international effort to develop technical standards for frontier AI, focused specifically on recursive self-improvement (RSI) — the process by which AI systems increasingly contribute to the development of successive generations of AI.

The proposal has four main components:

1. Standardized capability benchmarking. A common framework for measuring what frontier AI models can do — evaluations that national AI safety institutes can run independently and compare across providers. This would replace the current situation where each provider publishes its own benchmarks on its own evaluation sets, making cross-company comparisons unreliable. Standardized benchmarks are already how regulators evaluate drugs, aircraft, and nuclear plants — extending the framework to AI capability would be the most direct translation of existing regulatory practice.

2. RSI governance checkpoints. Rules specifying when and how AI systems can contribute to the development of successor AI generations. OpenAI was explicit that fully autonomous RSI is not currently happening and should not be pursued unless it can be conducted safely. The proposal calls for human oversight checkpoints and incident reporting requirements for any automated AI research that approaches RSI boundaries. This is the first time OpenAI has publicly called for external governance of its own research roadmap at this level of specificity.

3. Network of national AI safety institutes. Rather than creating a new international body — which would take years and face political resistance — OpenAI proposes leveraging the existing network of institutes established in Australia, Canada, France, Germany, India, Japan, Kenya, Korea, Singapore, and the UK. These institutes already have staff, government relationships, and some technical capacity. Using them as the implementation vehicle for international standards bypasses the institution-building problem and creates a faster path from proposal to enforcement.

4. Voluntary technical foundation, not licensing. The standards would operate as a technical baseline that responsible developers adopt voluntarily, not as a pre-market approval or government licensing requirement. This is the critical constraint that makes the proposal palatable to the US government: it does not require regulatory approval for AI development, which would face strong opposition from the domestic AI industry.

The strategic timing is precise: published two days before Altman briefs the Council, with enough lead time for member states to read it before the session, but close enough to shape what the session discusses. OpenAI is trying to define the terms of the governance conversation before other parties — particularly France and the EU, which favor binding rules — can set the terms instead.

Anthropic's Position: Slow Down and Restrict China's Chips

Amodei's public position entering the session is notably different from OpenAI's. In public statements ahead of the session, he has called for slowing the overall pace of AI development — a position that puts Anthropic at odds with OpenAI's implicit message that accelerating with the right standards is the correct response to AI safety concerns. He has also publicly supported maintaining US chip export restrictions on China, making Anthropic the most prominent frontier AI company to explicitly endorse the export control regime rather than arguing for its relaxation.

The Amodei-Altman split on these questions is real and will be visible to Security Council members. Altman's proposal envisions voluntary standards facilitated by an international network of safety institutes — a governance structure that allows AI development to continue at current pace under better oversight. Amodei's position implies binding constraints on development pace itself, and sustained restrictions on Chinese access to US AI hardware as a necessary complement to safety standards. The gap between them is not a communication disagreement — it reflects genuinely different views about whether the speed of AI development itself is a safety problem, independent of the standards governing that development.

For enterprise AI teams, Amodei's position has a direct practical implication: Anthropic's enterprise frontier safeguards — the zero-data-retention architecture announced September 1 — are consistent with a company that believes AI deployment needs to be governed carefully enough that safety infrastructure can operate at the same pace as capability deployment. Anthropic's product architecture is an expression of its policy position: if you are going to deploy frontier AI in enterprise environments, you need controls that governments and regulators can inspect and verify. Enterprise CISOs evaluating AI vendors should factor in that Anthropic's product decisions are informed by a governance thesis, not just a commercial one.

DeepSeek's Invitation: What the Presence of Chinese Frontier Labs Signals

The inclusion of DeepSeek and Moonshot in the Security Council session is the most structurally significant aspect of the meeting, and the one that most enterprise AI teams have underweighted in their governance planning.

DeepSeek's V3 and R1 models, released in late 2024 and 2025, demonstrated that Chinese AI capability had reached frontier levels at dramatically lower compute cost than Western equivalents — running at roughly 1/30th the compute cost of comparable US models at release. That finding undermined the premise of the US chip export control policy: if Chinese developers could reach frontier capability with restricted hardware, then chip restrictions alone could not maintain a US AI advantage. DeepSeek's presence at the Security Council is a reminder that the governance challenge is not just about governing Western AI companies — it is about governing a genuinely global technology where multiple countries have independent frontier-level capability.

Their invitation signals one of two possibilities. Either Western governments have concluded that AI safety standards must include Chinese developers to be meaningful — a voluntary standard adopted only by US and European companies is a constraint on Western developers that creates competitive disadvantage without actually improving global safety. Or China's government has concluded that participating in the emerging governance framework serves Chinese interests better than non-participation, since being inside the process gives China influence over what standards are written, rather than having those standards written by US and European regulators alone.

Both readings are important for enterprise AI procurement. If Western-Chinese AI governance cooperation deepens — even on limited voluntary standards — the risk calculus for deploying Chinese AI models in enterprise environments changes. If cooperation fails and governance frameworks diverge, enterprises using both US and Chinese AI vendors will face increasingly incompatible compliance requirements.

France's Presidency and the Geopolitical Architecture

France's decision to use its September Security Council presidency to convene an AI session reflects both genuine safety concerns and explicit geopolitical strategy. France has been the leading advocate within the EU for binding AI regulation: the EU AI Act's enforcement machinery, including Article 50 transparency requirements for AI-generated content, has been active since August 2026. The first EU AI Act enforcement actions demonstrated that France and the broader EU are willing to enforce binding rules, not just threaten them.

By pulling the AI governance conversation into the Security Council venue — where binding instruments are legally possible — rather than keeping it in the G7 voluntary track or the emerging CAISI network, France is signaling a long-term intent to establish international AI governance with the same juridical status as nuclear or chemical weapons governance. The Security Council is not where that governance will ultimately live — a binding international AI treaty would require years of negotiation and would likely operate through a different UN body — but the Security Council session establishes political precedent: AI is a matter of international security, not just trade or technology policy.

For the US, this precedent is double-edged. On one hand, Security Council jurisdiction over AI aligns with the US position that AI poses national security risks and should be treated accordingly. On the other hand, Security Council governance creates exposure to Russian and Chinese veto power over any future binding measures — a dynamic that OpenAI's proposal, with its emphasis on voluntary standards through national institutes, explicitly avoids.

The Trump-Xi Summit: Two Days Later

The September 25 Trump-Xi bilateral summit is the near-term event that will actually move AI policy, and the Security Council session's primary function may be to set the terms of the AI conversation that the two presidents will carry into that meeting. Amodei's public support for maintaining chip export restrictions — the most prominent AI executive to take that position on the record — provides the Trump administration with a domestic industry endorsement for the export control regime going into the summit.

The chip export restrictions have been the most contested element of US AI policy. American semiconductor companies have lobbied against them — the lost revenue to Chinese customers is real and immediate. American AI companies have been more divided. OpenAI's silence on the question has been notable; Anthropic's explicit endorsement of restrictions is the clearest signal yet that at least one major frontier lab believes the strategic logic of maintaining hardware access restrictions outweighs the commercial costs to the US chip industry.

If Trump and Xi reach an agreement that relaxes chip restrictions — trading hardware access for concessions on other issues — the competitive landscape for enterprise AI shifts within weeks. Chinese AI models would gain access to US-grade compute, accelerating their capability development. If restrictions tighten or remain unchanged, the existing infrastructure advantages of US AI providers — built on hardware that Chinese competitors cannot access — compound.

What Global AI Standards Would Mean for Enterprise Buyers

Enterprise AI compliance teams should track this session not because standards will emerge from it immediately — they will not — but because it establishes the governance venue and vocabulary for standards that will emerge over the next 12-24 months. The White House frontier AI model review framework, published earlier in 2026, established domestic review mechanisms for the most capable AI models. If the Security Council session generates momentum toward international analogues of that framework, enterprises will face compliance requirements spanning multiple jurisdictions simultaneously.

The specific standards OpenAI is proposing are directly relevant to enterprise deployment architecture:

Capability benchmarking standards would give procurement teams a vendor-neutral framework for evaluating model claims. Currently, each provider publishes its own benchmarks on its own evaluation sets — making it impossible to make fair cross-provider comparisons. Standardized external benchmarks, if they emerge, would function like standardized drug trials: giving enterprise buyers reliable third-party evidence about what each AI model can actually do.

RSI governance checkpoints and incident reporting would create documentation requirements for AI development processes that enterprises would need to understand when assessing provider risk. An AI provider that has undergone RSI governance review and can document its process represents a different compliance risk profile than one that has not — analogous to how SOC 2 certification functions for cloud software security.

Human oversight requirements would mandate review processes that overlap with the governance controls enterprise CISOs are already trying to implement for AI agent deployments. The governance infrastructure that Anthropic's enterprise frontier safeguards already provide — activity logs under customer-controlled encryption, misuse detection that operates without data retention — is exactly what these requirements would specify.

The Enterprise Playbook: What to Do Before Standards Are Defined

1. Map your AI supplier concentration by national jurisdiction now. If US-China bilateral AI agreements change chip export rules or research cooperation terms, the risk profile of Chinese AI providers — DeepSeek, Moonshot, Baidu Ernie, Alibaba Qwen — in your supplier stack changes with them. Know what you are using, at what volume, before you need to rapidly assess that risk.

2. Connect your compliance function to the relevant national AI safety institute. OpenAI's proposal routes international standards through national institutes — AISI in the UK, NIST's AI Safety Institute in the US, equivalent bodies in other jurisdictions. These institutes are the information source for what standards are coming before they become regulatory requirements. Connecting now means being in the information flow early rather than discovering requirements when enforcement begins.

3. Treat EU AI Act compliance as the floor, not the ceiling. The EU has demonstrated willingness to enforce binding AI rules. If the Security Council session generates momentum toward international standards that resemble the EU framework, being ahead of EU compliance now means being ahead of the emerging global baseline. Organizations that have invested in EU AI Act compliance infrastructure are better positioned for whatever international standard emerges than those that have not.

4. Build audit trails for AI decision processes before requirements are defined. Whatever specific standards emerge from international governance processes, they will almost certainly require documentation of how AI systems make decisions in high-stakes contexts: what models were used, what data they accessed, what decisions they influenced, and what human review occurred. Build the audit trail infrastructure now, while you have time to design it thoughtfully, rather than retrofitting it when specific documentation requirements arrive.

5. Monitor the Trump-Xi summit outcome for near-term policy changes. The September 25 bilateral summit is the highest-probability event for near-term changes to AI hardware export policy. If chip restrictions loosen, Chinese AI providers' cost and capability trajectory changes quickly. If restrictions tighten, the existing US provider infrastructure advantage compounds. Either outcome changes the enterprise AI vendor landscape materially — the governance forum today sets the terms, and the bilateral summit two days later sets the actual rules.

The Governance Framework That Is Actually Being Proposed

Step back from the diplomatic detail and the governance structure on the table is clear: voluntary technical standards for capability measurement and RSI governance, implemented through national AI safety institutes, with the US leading the framework and China participating in the process but not being subject to US regulatory authority.

This is substantially less than binding international AI regulations would look like — but it is substantially more than the current situation. Today there is no common framework for comparing AI capabilities across providers, no agreed standards for RSI governance, and no international incident reporting requirement for AI safety events. The Security Council session does not create these things — but it formally puts them on the agenda of a body with enforcement authority, for the first time.

The governance conversation that starts today in New York will shape the compliance requirements that enterprise AI teams face in 2027 and 2028. Organizations that treat this session as a diplomatic curiosity rather than an early regulatory signal will be less prepared when those requirements arrive than organizations that begin tracking the process now.

Takeaway: Today's UN Security Council AI session is not where binding international AI standards get created — it is where the political process that creates them gets formally launched. OpenAI's pre-brief proposal for US-led voluntary standards through national AI safety institutes is the most concrete governance proposal on the table, focused on capability benchmarking, RSI governance checkpoints, and human oversight documentation. The Amodei-Altman split on development pace and chip restrictions previews a governance debate that will shape the AI vendor landscape through 2027. DeepSeek's presence signals that AI governance, to be meaningful, must include Chinese developers — and that both sides have concluded engagement serves their interests better than exclusion. The Trump-Xi summit two days later is the event that actually moves AI hardware policy. Enterprise teams should take five actions now: map AI supplier concentration by jurisdiction, connect compliance functions to national AI safety institutes, treat EU AI Act compliance as the baseline, build audit trails before requirements are specified, and monitor the bilateral summit for near-term export control changes. The governance infrastructure you build before standards are defined determines how ready you are when they arrive.