ANDREJ KARPATHY: Access tiers, from free to paid to internal, gate how much capability each person can observe

The Gist

What you can see depends on what you can use. Free users get smaller models, paying users get better ones, and the very best stay inside the labs or with a few partners. So most people never see the strongest systems. This is a steelman: a charitable reconstruction of Andrej Karpathy's posts in their strongest form, not an endorsement of their claims.

Conclusion

Tiered access means most people can observe only systems well below the frontier, which by itself sustains a gap in shared understanding of capability.

Premises

1. Frontier developers offer different models at different tiers: free users receive smaller default models, paid users receive stronger models and agent tools, and the strongest systems remain internal or gated to vetted partners.
2. As of October 2026, free ChatGPT users receive a smaller default model rather than the top paid model, Anthropic's Mythos Preview is limited to participating organizations, and OpenAI's mathematics model has no public release date.
3. A person can observe only the capability of the tier they can reach.
4. Therefore the access structure itself ensures that most people see less capable systems than the frontier, independent of their ambition or domain.

Assumptions

• Gating of the strongest models is partly for safety, as with defensive-only access for vulnerability discovery, not only commercial.
• Free tiers have improved over time, so the size of the access gap varies across releases.
• Some capable open-weight models are freely downloadable, so free access to strong models exists for technically able users.

Analysis

Argument type: Inductive.

Potential Fallacies

• Overstated sufficiency (single-cause attribution) (Conclusion and P4) — The premises support access tiers as one contributing factor to uneven impressions of capability. 'By itself sustains' and 'ensures' claim they are enough on their own, leaving out skill, use case, media framing, prior beliefs, benchmarks and open-weight availability.
• Equivocation between observing and understanding (P3 to Conclusion) — P3 is about what an individual can personally use, while the conclusion is about collective understanding. Knowledge of capability also travels through testimony, published evaluations, demos and downstream products. A bridging premise is needed: that beliefs track direct experience and that indirect channels do not close the gap.
• Absolute quantifier / overgeneralization (P3 and P4) — 'Only' and 'independent of their ambition or domain' state a tendency as a law. Paying for a tier, joining a vetted organization, or running open-weight models (A3) all depend on resources, skill and motivation, so ambition does affect what a person can observe.
• Unsupported quantifier shift (existence to 'most people') (P1/P2 to P4 and Conclusion) — The premises show that gated tiers exist, not how users are distributed across them or how large the capability difference is. 'Most people' and 'well below the frontier' are plausible but unquantified.
• Snapshot treated as structure (P2 to Conclusion) — Three examples at one date support a claim about a lasting condition. A2 concedes that the gap varies across releases and that free tiers keep improving, so the word 'sustains' implies more durability than the evidence shows.

Counterarguments

These challenges are not ranked.

• Premise 3 (High impact) — Capability is observed indirectly through public benchmarks, system cards, independent evaluations, published demos, journalism and the reports of people with access. Direct use is not the only route to knowledge, so limited access does not mean limited understanding. The argument's own claims about gated systems show this.
• Conclusion (High impact) — Divergent views of AI capability persist even among people with equal access. Prompting skill, domain-specific jaggedness, use case, motivated skepticism and hype shape beliefs strongly. Equalizing access would not close the gap, so access is a contributing factor, not a sufficient one.
• Premise 2 (Medium impact) — Free tiers typically lag the frontier by months, competition keeps narrowing the lag, and gated releases are often staged rollouts. The gap is therefore a transient, shrinking time delay, not a durable structural feature, as A2 acknowledges.
• Premise 4 (Medium impact) — Ambition and resources partly determine the tier a person reaches. Paying for a subscription, joining a partner program, or running open-weight models (A3) are choices. 'Independent of ambition or domain' is therefore too strong, and the opinion-shaping groups (journalists, researchers, developers) are often not access-limited at all.
• Conclusion (Medium impact) — 'The frontier' is not a single ranking. Capability varies by task, so 'well below the frontier' may hold in one domain and not another, and for many everyday uses a smaller model may be adequate. The size and significance of the gap are never measured.

Suggested Improvements

• Weaken the sufficiency claim — Replace 'by itself sustains' and 'ensures' with 'is a significant structural contributor to' a gap in shared understanding, and drop or qualify 'independent of ambition or domain'. The reframed claim fits the evidence and the stated assumptions (A2, A3), and it keeps the insight that access is a distinct, non-individual driver of divergent views.
• Bridge from observation to understanding — Add a premise that most people form beliefs about capability mainly from their own direct use, and explain why indirect channels (benchmarks, reports, demos) fail to close the gap, for example because they are second-hand, selective or hard to trust. Revise P3 to say that direct experience is limited by tier. This removes the equivocation between observing and understanding and replaces a false absolute with a plausible empirical claim.
• Quantify prevalence and magnitude — Provide data on the share of users on free, paid and gated access, and measure the capability difference between tiers on common benchmarks over time. 'Most people' and 'well below the frontier' are central claims that currently rest on assumption. A one-generation lag and a multi-year lag have very different implications.
• Show the understanding gap empirically — Cite survey or behavioral evidence that capability beliefs differ by tier, such as users' estimates compared with benchmark results, or a comparison of free and paid users' forecasts. The outcome variable, 'shared understanding', is asserted and never measured, so the main inference is untested.
• Define terms and add scope — Say whether 'frontier' means the best public model, the best internal model, or the best model in a given domain. State the time scope explicitly and add a comparison class, such as earlier technologies with stratified access. This keeps the claim from being trivially true, stale after the next release, or too vague to test.

Scenario Tests

• Free-tier users are given temporary access to the top paid model, and their estimates of AI capability are compared with those of a control group. (Supports) — If beliefs shift substantially, access is a real causal driver of perceived capability. If they shift only modestly, access is one factor among several, which supports weakening the sufficiency claim.
• Two groups with identical paid access hold sharply different views of what AI can do, because of prompting skill and use case. (Challenges) — Access is not sufficient to explain divergent understanding. 'By itself' fails, though access can still contribute.
• A lab publishes detailed system cards and independent evaluators release benchmark results for a gated model, and public estimates of capability move toward those results. (Challenges) — Indirect observation can close much of the gap, so P3 fails as stated. Even so, the argument survives in a weaker form if such channels are slow or uneven.
• Free tiers converge to near-frontier capability within weeks of each release, as competition intensifies. (Challenges) — The gap shrinks to a brief lag, so 'sustains' gives way to a time-varying, partly transient condition. This is consistent with A2 but weakens the structural claim.
• Heavy users of paid agentic coding tools report rapid progress while free-chatbot users report stagnation, and the divergence tracks the tier used. (Supports) — This matches the 'two worlds' pattern and supports tier-limited experience as a driver of disagreement, though the evidence is mostly anecdotal and skill and task differences may also contribute.

Coherence & Relevance

The argument is a clear chain: tiers exist, they are illustrated with dated examples, observation is limited by reach, so most people see sub-frontier systems, and a gap in shared understanding persists. The first links are coherent and empirically grounded. The final inference, from limited direct access to a collective gap in understanding sustained by access alone, needs unstated premises about belief formation, prevalence and the absence of compensating channels. The stated assumptions (variable gap, open-weight access, partly safety-motivated gating) support a more modest contributing-factor claim than the sufficiency wording suggests. As written, the argument is plausible and informative but overconfident, and it is not valid as a deductive argument.

• Frontier developers offer different models at different tiers: free users receive smaller default models, paid users receive stronger models and agent tools, and the strongest systems remain internal or gated to vetted partners. (Strong) — Establishes that tiers exist but not how users are distributed across them or how large the capability differences are.
• As of October 2026, free ChatGPT users receive a smaller default model rather than the top paid model, Anthropic's Mythos Preview is limited to participating organizations, and OpenAI's mathematics model has no public release date. (Strong) — Concrete support for P1, but a small, time-bound sample of specialized models. It does not show persistence or typical gap size.
• A person can observe only the capability of the tier they can reach. (Moderate) — Links access to individual observation but ignores indirect channels. The move from observation to shared understanding needs a premise about how beliefs form.
• Therefore the access structure itself ensures that most people see less capable systems than the frontier, independent of their ambition or domain. (Moderate) — This is an intermediate conclusion rather than an independent premise. It needs a distribution premise for 'most people', and its universal wording conflicts with A3 and with the role of resources and skill.

View this argument on LogicFirst.ai