The AI Underclass Will Be Built With Access Controls Tony Malott · Published 2026-08-20 https://shareplane.malott.ai/artifacts/the-ai-underclass-will-be-built-with-access-controls/ AI access · Cognitive opportunity · August 2026 The AI Underclass Will Be Built With Access Controls The question is not whether everyone gets AI. It is who gets the best intelligence while it still creates an advantage. The AI underclass will be defined less by who has AI than by who can obtain the best intelligence, how much they can afford, how freely they can use it, and whether governments or providers permit access to its full capabilities. By Tony Malott Updated 2026-08-20 Protected Development candidate AI is everywhere. Capability is not. A capability field progresses from broad access toward restricted capability while ordinary access controls interrupt the path. 01 Broad access 02 Frontier 03 Scale 04 Restricted capability Price Plan Quota Identity Geography Trust Compute I wrote the first version of this article on July 7, 2026. At the time, the argument felt aggressive even to me. Frontier AI was starting to behave less like ordinary software and more like strategic infrastructure. Government intervention, trusted-access programs, model restrictions, pricing tiers, API gates, safety classifiers, compute ownership, and geopolitics were beginning to determine who could use which capability. My concern was that the AI underclass would not be created by taking AI away from people. It would be created by giving different people different intelligence. Six weeks later, I think that distinction matters even more. The important question is no longer whether everyone will have an AI assistant. They probably will. Cheap models will improve. Open models will improve. Free products will improve. The baseline will keep moving upward. The harder question is: Who gets the best intelligence while it still creates an advantage? That is a different problem. And there is a second distinction worth putting on the table early: Money does not buy wisdom. It buys cognitive opportunity. That means access is no longer only about whether a model is available. It is about how much useful reasoning work a person or organization can actually bring to a problem. 02 Everyone Can Have AI and Still Not Have Equal Intelligence The phrase “AI underclass” can sound like science fiction if we imagine a world where one group has artificial intelligence and everybody else gets nothing. I do not think that is the likely outcome. The likely outcome is more ordinary, which is precisely why it is easier to normalize. One person gets a capable free model. Another pays for a stronger model. Another gets the highest reasoning tier. A company buys API capacity and runs thousands of model calls. A large enterprise runs persistent agents, long contexts, internal tools, private data, and automated evaluation around the clock. A trusted research organization receives access to capabilities that are not available to ordinary subscribers at all. Everyone can truthfully say they “have AI.” Those statements can all be true while the productive capability available to each person is materially different. We already accept this in computing. Owning a laptop and owning a data center both count as access to computers. Nobody pretends they create equal computational capacity. AI is heading toward the same distinction, except the thing being metered is increasingly useful cognitive work. Access is binary. Capacity is not. Four ways to truthfully say “I have AI.” Every lane begins with AI access. The usable capability, scale, duration, and qualification posture diverge. 01 Consumer / free capable model bounded usage general access 02 Frontier subscription stronger reasoning higher limits paid access 03 Scaled agent / API parallel execution sustained throughput organizational capacity 04 Restricted / trusted specialized capability qualified use identity + trust gate 03 The Permission Layer Became Literal The original article focused heavily on government and provider controls because the pattern was already visible. Then the pattern became painfully literal. Anthropic launched Claude Fable 5 and Claude Mythos 5 on June 9. Anthropic described Fable 5 as more capable than any model it had previously made generally available. Mythos 5 used the same underlying model with fewer safeguards in sensitive cybersecurity and biology domains and was restricted to trusted-access partners. Both were priced at $10 per million input tokens and $50 per million output tokens. [1] Three days later, Anthropic said the U.S. government had issued an export-control directive requiring the company to suspend access to Fable 5 and Mythos 5 by foreign nationals. Anthropic said it could not reliably verify nationality in real time, so it disabled both models for all customers. [2] The controls were later lifted. Fable returned globally. Mythos access returned for a limited set of U.S. organizations following government approval and remains restricted through trusted-access programs. [3][4] That sequence matters even if we believe the safety concerns were legitimate. A frontier model was released. Government action changed who could use it. The provider removed access. The provider later restored access under a different structure. The most capable version in sensitive domains remained available only to selected organizations. That is not a hypothetical architecture for access control. It is an observed release history. The policy question is not whether governments should ever intervene in genuinely dangerous capabilities. Sometimes they should. The question is whether we can build legitimate safety controls without quietly converting advanced intelligence into a permission economy. Observed release history The permission layer became literal. Release, government intervention, suspension, and a split restoration changed who could use the same frontier-model family. JUN 09 Release Fable 5 and Mythos 5 launch JUN 12 Directive Government export-control intervention JUN 12 Suspension Access removed globally while nationality could not be verified JUN 30 Controls lifted Fable return announced JUL 01+ Split return General Fable access resumes; Mythos remains trusted-access 04 The Economic Layer Is Now Harder to Ignore The part I understated in July was money. Access control is not only political. It is economic. Anthropic currently describes Fable 5 as its most capable generally available model and offers it on paid Pro, Max, Team, and Enterprise surfaces. Mythos remains available to a smaller trusted-access population. [4][5] OpenAI now makes a similar capability distinction explicit across GPT-5.6. Free and Go users receive Luna in standard ChatGPT use. Eligible paid plans receive Sol. Extra High reasoning and Sol Pro are reserved for higher tiers. OpenAI also separates Sol, Terra, and Luna across ChatGPT, Codex, Work, and API surfaces by capability and cost. [6][7] OpenAI has also created a Trusted Access path for more sensitive cybersecurity capability. Its July 9 GPT-5.6 release says qualified individuals and organizations can receive more precise safeguards for verified defensive work, with identity verification, stronger account-security requirements, and restrictions for high-risk entities and jurisdictions. [7] Again, that may be entirely defensible as safety engineering. It is also another example of the same architecture: capability changes with trust status, identity, account posture, and institutional qualification. None of this is scandalous by itself. Training and serving frontier models costs money. Capacity is finite. Providers need sustainable businesses. Higher-cost products have existed since the first person discovered that customers will pay more for the version with the useful button enabled. But AI is not just another productivity application. We are increasingly buying access to problem-solving capacity . That changes the economics. 05 Cognitive Opportunity Is the Real Economic Unit It is tempting to reduce this argument to “the rich get the smart model.” That is too simplistic and, more importantly, wrong often enough to weaken the real point. A cheaper model can outperform an expensive model on a particular task. An expert using a modest model can outperform a novice using the frontier model. A good local model with the right tools, data, and workflow can beat a badly designed expensive system. Judgment, taste, domain expertise, courage, curiosity, and the ability to recognize nonsense remain stubbornly unavailable as subscription upgrades. Money does not purchase wisdom. What money increasingly purchases is more cognitive opportunity . It can buy a stronger model for difficult tasks, more attempts at the same problem, longer contexts, longer-running agents, more parallelism, more tool calls, more private data integration, more evaluation and verification, lower latency, fewer quota interruptions, and specialized access that ordinary users do not receive. That does not guarantee a better outcome. It changes the probability of getting one. The advantage is not that a wealthy person presses a button and becomes smarter. The advantage is that a person or organization with more capital can increasingly bring more high-quality cognitive work to the same problem in the same amount of time. That advantage can compound. Better access can support faster iteration or stronger execution. Stronger execution can create economic returns. Those returns can finance still more inference, agents, compute, data, and experimentation. That loop is not inevitable. Plenty of well-funded organizations will use expensive AI to produce extremely sophisticated garbage. Capital has never been a vaccination against bad judgment. But even a small advantage, repeated across research, engineering, analysis, software delivery, and decision-making, can become a large difference over time. The deeper economics of that compounding effect deserve their own argument. The point here is narrower: economic capacity is itself becoming an access control. The economic unit Cognitive opportunity Economic capacity can increase the amount and quality of cognitive work brought to a problem. It does not purchase judgment. 01 Model capability 02 Attempts 03 Context 04 Duration 05 Parallelism 06 Tools 07 Private data 08 Verification 09 Throughput 10 Quota freedom Money can buy more opportunity to apply high-quality cognitive work. It cannot buy wisdom. The bounded handoff 01 Access 02 More cognitive opportunity 03 Potential execution advantage The full capital-intelligence flywheel belongs to the companion thesis, not this article. 06 The Strongest Counterargument Is Also True There is a serious argument against this thesis. AI is getting cheaper. Open-weight models are improving. Inference hardware is improving. Competition is brutal. Capabilities that were frontier six months ago can become ordinary surprisingly quickly. That matters enormously. Open models and cheap inference are the strongest pressure against a permanent AI aristocracy. They give individuals, universities, startups, small businesses, public-interest organizations, and countries outside the richest technology blocs an alternative path. They also make access controls harder to enforce once capable weights and techniques are widely distributed. I hope that pressure wins. But falling absolute cost does not automatically eliminate relative advantage. Computers became radically cheaper and more available. That did not make a laptop equivalent to a hyperscale cluster. Cloud computing democratized infrastructure. It also created companies capable of spending billions of dollars on compute that smaller competitors could not match. The same thing can happen with intelligence. The floor can rise while the ceiling rises faster. That is the scenario worth watching. Counterforce Concentration and diffusion are happening at the same time. Capital, restricted access, and compute can concentrate capability while open models, competition, hardware improvement, and lower inference cost push it outward. Toward concentration Frontier cost Trusted access Scale Compute ↔ Toward diffusion Open weights Competition Hardware improvement Falling inference cost 07 Safety Cannot Become a Capability Aristocracy There are legitimate reasons to restrict certain capabilities. Cyber offense is real. Biological misuse is real. National-security competition is real. A serious access policy cannot simply shout “freedom” and pretend every capability has identical consequences. But the opposite failure is equally serious. If access to advanced intelligence is governed by opaque government relationships, selected corporate partnerships, pricing barriers, discretionary trust programs, geographic restrictions, and institutional status, then we can create a capability aristocracy while insisting that everybody still has access to AI. That is the trap. A defensible safety framework should make the control logic inspectable. What capability is being restricted? What evidence justifies the restriction? Who qualifies? Who decides? How long does the restriction last? What appeal or review path exists? What changes when the risk falls? What prevents a temporary safety boundary from becoming a permanent competitive moat? Those questions matter because intelligence is becoming economically consequential infrastructure. The more consequential the capability becomes, the less comfortable we should be with “trust us” as the governance model. 08 The Underclass Will Still Have AI That may be the most important update to my original thesis. The AI underclass will probably have excellent AI. That is what makes the issue easy to miss. They may have models that would have looked miraculous a year earlier. They may have open weights, local inference, assistants in every application, cheap agents, and more computing power than entire companies once possessed. And they may still be operating at a meaningful disadvantage. Because the comparison that matters is not yesterday’s intelligence. It is the intelligence available to the other person today. The dividing lines will move. Free versus paid. Paid versus premium. Premium versus API scale. General access versus trusted access. Rented inference versus owned compute. Permitted capability versus restricted capability. The labels will change because the technology industry has never encountered a hierarchy it could not turn into a pricing table. The underlying issue will remain. If intelligence becomes one of the most important productive resources in the economy, then broad access to capable intelligence is not merely a consumer-product question. It becomes an economic-design question. A competition question. An education question. A national-policy question. And eventually a civic question. I still believe safety matters. I still believe providers should be paid. I still believe some capabilities justify stronger controls. But we should be very careful about building a future where everyone technically has AI while only a narrow class can consistently afford, operate, or receive permission to use the intelligence that creates the largest advantage. The ordinary boundary The underclass does not need to be denied AI. Relative capability can remain unequal across an ordinary access boundary. ACCESS You have AI. BEYOND THE GATE Someone else has more. That would not look like exclusion. It would look like a perfectly ordinary login screen. And a better model on the other side of it. --- 09 Sources Anthropic, Claude Fable 5 and Claude Mythos 5 , June 9, 2026. Anthropic, Statement on the US government directive to suspend access to Fable 5 and Mythos 5 , June 12, 2026. Anthropic, Redeploying Claude Fable 5 , June 30, 2026, updated July 1, 2026. Anthropic, Claude Mythos 5 , accessed August 20, 2026. Anthropic, Claude Fable 5 , accessed August 20, 2026. OpenAI, GPT-5.6 in ChatGPT , accessed August 20, 2026. OpenAI, GPT-5.6 , accessed August 20, 2026. Evidence behind the thesis Check the work, not just the conclusion. Public research, authority, lineage, and author testimony are labeled separately. Sources can corroborate, challenge, or bound the argument; they do not replace Tony Malott's judgment. Portable public record Take the complete artifact with you. The deterministic package contains a self-contained offline article, the exact public-route snapshot, canonical public metadata, receipt, source text when available, plain-text context, claim ledger, source records, and a member-hash manifest. Download full artifact package Read plain-text context Inspect package manifest 3 public sources Sources, authority, and lineage Each record states the role it plays. Research support and governance provenance are not treated as interchangeable. Historical Source Original SharePlane publication Historical predecessor authority. Historical predecessor authority. Open source Migration Source SharePlane Next preservation package Preserved migration and provenance record. Preserved migration and provenance record. Open source Governing Publication Authority Issue #540 Governs the August 20, 2026 refresh, semantic and design locks, implementation boundary, and owner UAT. Governs the August 20, 2026 refresh, semantic and design locks, implementation boundary, and owner UAT. Open source Claim discipline What is asserted—and how it is bounded Research, author analysis, and personal testimony remain distinct. Supporting links and caveats stay attached to each claim. Owner Thesis claim:equal-ai-not-equal-intelligence Universal access to some AI can coexist with materially unequal access to the most consequential intelligence. Support content/artifacts/the-ai-underclass-will-be-built-with-access-controls/EVIDENCE_FRESHNESS.issue-540.json Evidence Bound Synthesis claim:permission-layer-observed Recent frontier-model releases show capability access changing through government action, trusted-access programs, plan tiers, account posture, and provider safeguards. Support content/artifacts/the-ai-underclass-will-be-built-with-access-controls/EVIDENCE_FRESHNESS.issue-540.json Owner Inference claim:cognitive-opportunity Economic capacity can buy more opportunity to apply high-quality cognitive work through capability, attempts, context, duration, parallelism, tools, data, verification, throughput, and fewer usage constraints. Support content/artifacts/the-ai-underclass-will-be-built-with-access-controls/EVIDENCE_FRESHNESS.issue-540.json Boundary This does not claim that money buys judgment, wisdom, or guaranteed outcomes. Public boundary. Current provider facts are dated August 20, 2026. Economic compounding claims are explicitly Tony's bounded inference rather than vendor-published fact. The full intelligence-capital flywheel remains outside this Work under Issue #541. 3 sources 3 governed claims 1 portable package SOURCE REFERENCES Consult the source references retained in the native presentation.