On June 26, 2026, OpenAI unveiled the GPT-5.6 family. The flagship model, Sol, scored 88.8% on TerminalBench 2.1 — matching Anthropic’s Claude Mythos 5 while producing only a third of the output tokens. The mid-range Terra targets GPT-5.5-level price-performance. The low-end Luna is priced at $1/$6 per million tokens. But what truly ignited the developer community was buried in the second-to-last paragraph of the blog post: GPT-5.6 Sol will land on Cerebras inference chips in July, hitting 750 tok/s. That same day, The Washington Post broke the news that the US government would impose a vetting regime on GPT-5.6 users — only government-pre-screened “trusted partners” would receive access. The top-voted HN comment began with a single sentence: “This is regulatory capture in action.”
Put these two things together, and you have a complete story. On one side, an acceleration in engineering performance — 750 tok/s means getting a frontier-model response in the browser faster than a human can read. On the other, a tightening policy gate — the government decides who gets to use it. The tension between the two points to a judgment that makes the technical community uneasy: regulatory capture is moving from a political science concept to an engineering reality.
The Technical Triggers for the Vetting Regime
To understand why the US government acted now, you need to look at GPT-5.6’s performance on cybersecurity benchmarks. In its safety system card, OpenAI disclosed that Sol achieved unprecedented success rates on both “automated vulnerability research” and “exploit generation” tasks — strong enough that the company itself described the model as having the ability to “shift the performance-efficiency frontier for long-horizon security tasks.” In other words, this model can not only find vulnerabilities — it can plan multi-step exploit chains and autonomously execute them over extended time windows.
OpenAI’s countermeasure was model-level hardening — Sol was designed to be defense-oriented, prioritizing remediation over attack code, and equipped with “the most robust security stack yet” for jailbreak resistance. But the US government was clearly not satisfied with corporate self-policing. In early June, President Trump signed an executive order requiring frontier AI labs to submit models for government review 30 days before release, while promising this would be a “voluntary process.” Two weeks earlier, Anthropic, under a government export-control directive, was forced to take Mythos 5 and Fable 5 completely offline — even the company’s own foreign-national employees lost access.
By the time GPT-5.6 launched, this “voluntary framework” did not actually exist. OpenAI executives conceded in a media briefing that there is currently no formal review standard to follow — the company merely sends its customer list to the government and receives feedback. Former White House AI advisor Dean Ball, soon to join OpenAI, described it bluntly as a “de facto involuntary licensing regime.” From an engineering standpoint, a vetting process with no defined safety benchmarks, no transparent approval criteria, and no appeals mechanism is, at its core, an arbitrary power interface. Anyone who has ever called an API knows that an interface without an SLA is unreliable — the same holds for policy interfaces.
The Regulatory Capture Argument: Both Sides
Regulatory capture refers to a situation where a regulatory agency is co-opted by the industry it regulates, transforming from a guardian of public interest into a defender of industry interests. In the GPT-5.6 case, the applicability of this concept needs to be examined from both directions.
Those who support the capture thesis point to several lines of evidence. First, the current president’s senior AI advisor, David Sacks, is himself a partner at Craft Ventures — and Craft is an OpenAI investor. Second, the vetting regime effectively grants GPT-5.6 and Mythos 5 a “government-endorsed” label — already-approved enterprises gain a competitive moat, while newcomers must prove they are “trustworthy” to enter. HN user jmward01 wrote: “This will make it hard/impossible for new vendors to come into the market and only established companies will get to play, and charge, for LLMs.” Third, the two stories that broke the same day form an ironic juxtaposition: GPT-5.6 requires vetting to ship, while Anthropic’s Mythos 5 blockade was lifted — the Commerce Department sent Anthropic a letter authorizing release to over 100 US institutions, on the condition that Anthropic commit to working with the government on future agreements and release standards. One HN commenter put it starkly: the vetting regime doesn’t lock down safety — it locks down who gets to make money.
The voices that resist casually labeling this as capture also have their logic. They argue that frontier models possess capabilities that exceed the scope of traditional software tools — a model that can autonomously discover and exploit zero-day vulnerabilities has national security implications clearly different from a better code-completion tool. Drugs, chemicals, and explosives all require vetting — why shouldn’t models? HN user coffeemug drew the analogy with drugs, chemicals, and explosives, while adding: “I wouldn’t say it’s a good idea.” Commerce Department spokesperson Benno Kass emphasized that the speed of the government’s action reflected responsible urgency: “In just two weeks, we’ve worked to ensure the United States remains the global leader in AI while safeguarding our security.”
The weak point in this logic is: what are the vetting standards? If the standards are undefined, then “security” can degrade into “security as we define it,” and “as we define it” — in the absence of transparent rules — equals arbitrary discretion. From a technical governance perspective, this is a classic “security-justification trap”: invoking security to bypass the obligation to establish clear rules.
Pax Silica: The Geopolitical Extension of Vetting
The US vetting regime is not an isolated domestic event. In June, the US State Department-led Pax Silica agreement gained ten new signatories, including the European Union as a bloc. HN user rzerowan’s comment captured the practical effect of this framework with precision: “EU will be a renter of the LLMs that the US allows them to use.” Pax Silica is nominally a multilateral framework for coordinating chips, semiconductors, data centers, and AI supply chains — but in practice, it functions first and foremost as an institutional tool to bar Chinese models from allied markets. With the EU signing on, European enterprises will effectively select their AI models from a list approved by Washington.
This is not a conspiracy theory. Semafor reported that European officials have already expressed frustration at being “dependent on decisions made in Washington.” The vetting regime, layered on top of Pax Silica, transforms AI access from a market question into a licensing question. For startups outside the US, this means they must both compete with established American incumbents and satisfy US government security review standards — and the latter, by institutional design, leaves no room for foreign newcomers.
The Open-Source Counter-Window
Against this backdrop, a quantitative analysis by Doubleword blogger Jamie Dborin offers a counterintuitive timeline. He tracked 18 benchmark metrics from Artificial Analysis, measuring the time lag for open-weight models to match closed-source models on each capability dimension. The core finding: the gap between the open-weight frontier and the closed-source frontier has been steadily narrowing since summer 2024, and at the current regression trend, the gap will reach zero on December 3, 2026.
I approach this prediction with caution — it is based on a single institution’s set of benchmarks, and the regression assumes linear extrapolation, whereas real-world progress is typically nonlinear. But the directional signal deserves serious attention: if open-source models are indeed catching up across 18 metrics, the effective window for a vetting regime may be only six months. The shorter the half-life of a regulatory moat, the more pronounced the side effects of market distortion.
This is also why the HN community repeatedly invoked the historical analogy of MySQL/PostgreSQL defeating Oracle. When MySQL launched in the mid-1990s, nobody believed it could compete with Oracle’s enterprise-grade database. But because MySQL was good enough, open, and freely deployable, it generated network effects among developers and ultimately underpinned the infrastructure layer of the internet. A parallel narrative is forming in the LLM space: open-source models like Qwen, DeepSeek, and Kimi are iterating continuously in markets outside the US, while the vetting regime turns the US domestic market into a closed laboratory — and the open ecosystem accelerates its evolution externally.
rzerowan put it this way: “In the long run OpenSource will dominate as it did in the DB (MySQL/Postgres) / ServerOS (Linux/BSDs) versus Proprietary rent seeking alts like Oracle and Microsoft et al.” But they also added a crucial caveat: “the transition period will be ugly.” The small startups and independent developers who can’t get vetting approval during that transition will bear the ugliest side of it most directly.
Don’t Overestimate the Stability of Vetting Regimes
From a broader perspective, the vetting regime faces at least three structural pressures. First, the US itself is internally contradictory — the same executive branch is simultaneously demanding slower release cadences, pushing global deployment through Pax Silica, and worrying about China pulling ahead in the AI race. Dean Ball’s warning bears repeating: the absence of clearly defined safety standards could lead to “an endless cycle of release delays,” which could not only cede first-mover advantage to China but also jeopardize the hundreds of billions of dollars invested in AI infrastructure.
Second, the compliance costs of vetting naturally favor large companies. An OpenAI or Anthropic with a legal and policy team of hundreds can engage in “daily intense negotiations” (as Commerce Secretary Lutnick put it) to secure release; a five-person startup can hardly afford the same level of government-relations investment. Complexity itself is a barrier — a side effect of how the system operates, not deliberate discrimination.
Third, the technology itself does not wait. Cerebras’s 750 tok/s opens the door to a new phase — the leap in inference speed will unlock real-time agent workflows that are currently infeasible. The time constants of the technology capability curve and the policy response curve are not synchronized; the former is usually shorter. Policymaking is a high-friction process; engineering iteration does not need consensus.
On the day GPT-5.6 launched, the community saw more than a model release. It saw an industry’s competitive rules being rewritten in real time. Whether the vetting regime solidifies incumbent advantages, as commenters fear, ultimately hinges on a question that remains unresolved as of this writing: what, exactly, determines whose name appears on the approval list. If the criteria remain opaque, unreviewable, and untraceable, then “regulatory capture” is an accurate description of the power structure. If — and this is a very big “if” — the government can produce a publicly defined, measurable set of safety benchmarks and a transparent vetting process within weeks, then the current friction may prove to be nothing more than the growing pains of institutional磨合.
The analysis above is based on currently available public information and community discussion. If you have a different perspective or additional information, discussion is welcome.