GPT-5.6 'Sol' and 'Luna': Anatomy of a Hype Cycle Wrapped in Inference Sliders

On August 7, someone leaked a roadmap that OpenAI never published. It describes GPT-5.6 Sol, GPT-5.6 Luna, and a free tier with supposedly unlimited text chat. The fact-error rate, we're told, dropped 62% in finance and 68% in medical and legal questions. Strong numbers. Zero methodology. I have audited enough yield contracts to know that every metric needs a ledger. Here, the ledger is missing. No test set. No sample size. No third-party reproduction. Just a version number that collides with nothing in the public release history, wrapped in the kind of confident marketing language I normally associate with a whitepaper that ends in a post-mortem. The code didn't — and in this case, the data doesn't either. Let me be clear about what this leak is, and what it isn't. It's a product strategy document disguised as a model announcement. It is not a technical specification, not a benchmark report, and not a verified event. But leaks are like on-chain transaction histories: they reveal intent even when they withhold proof. So let's dissect the body and see what's actually pumping under the hood. For those catching up, the leaked narrative describes what appears to be OpenAI's next consumer-frontier revision. All users — from free to Pro — are reportedly slated to run on a “single model” capable of switching between instant responses and deep reasoning. A new slider would let users dial how much “thinking” each reply receives. A dedicated Think button would reportedly be offered to free and Go users, while file uploads, images, and other tools remain gated by usage limits. The rollout plan was to ship Sol to Plus and Pro first, then push Luna as the default for Free and Go, with unlimited text chat unlocked the following week. There is a credible internal logic to all of this. But credibility of logic is not the same as credibility of fact. The model names — Sol and Luna — collide with the token names of a blockchain project whose collapse taught a generation of investors a bitter lesson about confidence intervals. In 2022, Terra's Luna was supposed to keep UST pegged via an algorithmic arbitrage loop. The math was beautiful on paper. The market made it ugly in practice. Every block hides a confession — and the confession here is that OpenAI is leaning hard on “internal evaluations” to move the narrative. Let me start the technical autopsy. The core claim is that this is not a two-model architecture but a single base model with a control dial for inference-time compute. That framing is actually the most plausible part of the entire leak. The industry has been moving from a fast-model/slow-model binary toward what researchers call inference-time compute scaling. Instead of training a tiny model for speed and a giant model for depth, you train one strong base and let the decoding budget expand and contract per query. The slider is not a magic switch; it is a meter for how many tokens of chain-of-thought the model burns before answering. That has a direct cost consequence. When a free user hits the Think button, they are consuming potentially three to five times the compute of a normal reply. The infinite chat offer is not an act of charity. It is a cost-delivery architecture in which the high-margin, low-reasoning baseline keeps the lights on, and deep reasoning becomes the premium accelerant. Anyone who has run a GPU cluster knows this trade-off. I wrote about the same dynamic in 2020 when I quantified slippage risk in SushiSwap's fork mechanics: the community wanted the LP yields, but the real cost was buried in the AMM invariants. Same shape here. The true price of 'infinite' is hidden in the inference budget. Now let's audit the factuality claim, because this is where the leak behaves like a classic bait-and-switch metric. A 62% relative reduction in error rate sounds heroic until you ask: from what baseline? If the old model answered 20 out of 100 finance questions incorrectly and the new one answers 7.6, that's a 62% drop — but it still fails 8 times out of 100. In a medical or legal context, an 8% failure rate is a lawsuit generator, not a green light. Conversely, if the baseline was 3% and it dropped to about 1%, the improvement is real but the headline is inflated. The leak deliberately omits the absolute number. That omission is not an accident. It is the same trick used by unaudited stablecoin reserves: claim the ratio, hide the base. From my experience auditing Harvest Finance's early alpha in 2018, I learned that what matters is not the discovery of one re-entrancy bug but the pattern of how the team discloses the rest. When a protocol tells you only about the fix and never about the tests, you become the test. The same principle applies here. OpenAI's internal evaluation could be meaningful, but without details on annotation methodology, domain difficulty, or cross-lingual coverage, it is a public relations artifact, not a scientific measurement. Let's move to the commercial anatomy, because this is where the leak reveals its cleanest logic. The combination of free unlimited text chat and gated multimodal tools is a textbook freemium funnel. Text-only inference is far cheaper than image understanding or file parsing. By making text unlimited, OpenAI converts the free tier into a habit loop. The user returns daily, develops dependence, and then hits the file-upload limit or the reasoning cap. At that exact friction point, the Think button is the elegant upsell: a taste of deep reasoning available to the free user, deliberately rationed, designed to demonstrate what Pro would feel like. For Plus and Pro users, the slider serves a different function. It lets power users optimize for their own latency and quality trade-off, effectively shifting the compute-cost decision onto the user. That is a sophisticated cost-transference mechanism. OpenAI exposes the dial, the user absorbs the responsibility, and OpenAI captures the behavioral data. Free tier provides the scale, paid tier provides the margin, and the slider provides the excuse when inference costs spike. It would be a beautiful model — if it were real. But, as with most leaked marketing, the commercial story is missing the hard numbers. No pricing changes, no API information, no indication that the consumer tier will absorb external developer traffic. The document's silence on API pricing is strategic. If OpenAI truly achieved a 68% factuality improvement with cost parity, the API would be the most aggressive place to prove it. The leak does not. That tells me the enterprise and developer stories are still being managed, which means the consumer announcement is the tip of a much larger commercial wedge. Now, the infrastructure question. An “unlimited” text tier is an infrastructure declaration, not a product feature. For almost any other AI company, offering uncapped free chat would be an existential risk. The fact that the leak even floats this possibility suggests OpenAI has made genuine progress on inference efficiency — likely through a mix of quantization, speculative decoding, KV-cache optimization, and aggressive batch scheduling. I would not be surprised if the deeper story here is not model quality but unit economics. The version names Sol and Luna — day and night — hint at load-shifting: serving the world's peak demand by moving compute across two hemispheres, using cheaper energy and idle cycles. That is a classic demand-response pattern, and it is exactly what a company that intends to run millions of free requests per minute must engineer. My skepticism on this point is purely Bayesian. I have seen too many “unlimited” offers in crypto turn out to be soft-capped by hidden rate limits. The leak's mention of anti-abuse mechanisms is a tell. Free unlimited will likely mean “unlimited with a token bucket, per IP, per account, deprioritized during peak hours, and throttled for bots.” That is not a lie. It is just not the whole truth. And in this industry, the partial truth is the most expensive version of false. Here is the uncomfortable angle the bulls have latched onto, and I will give credit where it is due: the unified-model-plus-slider design could be a genuinely superior consumer experience. Every competitor, from Google to Anthropic, forces users to choose between a fast model and a smart model. The UI asks you to know beforehand how hard your question is. That is bad design. The slider removes the pre-selection burden. You ask the question, the model decides the budget, and you can always override with a toggle. It is the difference between buying a second-hand GPU and renting elastic compute. As someone who has spent years working across both worlds, I recognize the elegance. The product is the architecture. What the bulls also get right: free unlimited text chat is the most potent customer-acquisition weapon in the consumer AI wars. We have seen the pattern before. Every time ChatGPT opened a free tier, traffic jumped and the closed-loop data flywheel accelerated. The training signal from millions of daily interactions is a moat that no benchmark can capture. Even at an elevated inference cost, that data has long-term value far beyond the price of a single GPU-hour. The bulls are not wrong to say that this, if true, strengthens OpenAI's position. Liquidity flows, but integrity stagnates — and in a free market, flow matters more than fidelity. Yet here is the contrarian counterpoint the same bulls ignore. The factuality claim, even if true, cuts both ways. When you position a model as better at finance, medical, and legal questions, you are placing a target on its back. Regulators are now watching for the exact class of error the leak celebrates. The EU AI Act, the FDA, and financial regulators will not read “62% improvement” as a victory; they will read it as an admission that the model still produces deterministic-looking lies in regulated domains. The more confident the model becomes, the more dangerous its mistakes. In my consultation work with Australian banks on Bitcoin ETF exposure, I repeatedly warned about model risk in exactly these terms. Stronger models create sharper tail risks. There is also a structural contradiction in the leak's own rollout plan. Sol for Plus/Pro and Luna for Free/Go suggests two different post-training configurations, which contradicts the “one model” narrative. If the base is identical, the difference must be in safety tuning and output style. But that opens a regulatory hole: the free model, with less guard-rail investment, gets exposed to the most high-stakes questions from the least protected users. The leak mentions anti-abuse controls, but not the red-team results, the alignment methodology, or the bias evaluations. In security terms, that is like a smart contract audit that only checks for reentrancy and ignores flash-loan manipulation. The known unknown is the one that kills you. Let me now return to the name collision, because I think there is a deeper metaphor the leaker did not intend. Terra Luna presented itself as a dynamic stabilization mechanism. Its own community called the equilibrium “mathematically inevitable.” What killed it was not a flaw in the arithmetic but a mismatch between the model's assumptions and real-world liquidity withdrawals. When I conducted my post-mortem of UST's depeg, I calculated the liquidity depth required to defend the peg under panic conditions. The answer had six zeros. The market produced nine. In the same way, OpenAI’s internal evaluation is a controlled environment. It cannot measure how the model performs under adversarial users, prompt injection attacks, or prolonged hallucinations that sound increasingly reasonable. Sol and Luna, as names, may be a warning rather than a promise. The takeaway from this leak, if we step back, is not about OpenAI's roadmap. It is about the industry's relationship with evidence. A company leaks a narrative. The narrative contains an impressive-sounding percentage. No one demands the baseline. No one asks for the test set. The market begins pricing the outcome before a single third-party replication exists. Sound familiar? That is exactly how we bought into the algorithmic stablecoins. That is how we bought into the open-source security theater. The cycle repeats because the incentives repeat: attention flows to the strongest claim, not the best proof. And the strongest claim is always the one without a method section. As a cold dissector, my own position is simple. I do not reject the possibility that OpenAI has made real progress. I reject the possibility that this leak proves it. The technical patterns I see — unified models, inference-budget sliders, and free-tier compute rationing — are all plausible industry trends. The commercial logic of “infinite text, limited tools” is sharp. But the factuality claim is unaudited, the infrastructure details are absent, and the regulatory risk is understated. We chased the glow, not the ledger. The glow here is the word “infinite.” The ledger is the hidden rate limit, the red-team report, and the absolute accuracy number that will arrive, conveniently, only after the model has captured the market. History is written in hex, not headlines. But this one is written in marketing. Until OpenAI publishes its own evaluation methodology, I will treat GPT-5.6 Sol and Luna the way I treated the Terra whitepaper: as an interesting theory that deserves verification, not investment. The burden of proof is on the party making the extraordinary claim. A 62% reduction in financial errors is an extraordinary claim. Leaked slide decks are not extraordinary evidence. Here is what I will be watching in the next 90 days. First, the third-party benchmarks: MMLU, GPQA, HumanEval, SWE-bench, and any independent factual-accuracy suite. If the model appears at the top of those without gaming, the leak's core premise gains real weight. Second, the API pricing sheet. If OpenAI lowers the price per token while claiming better accuracy, that is the real revolution — it means the efficiency gains are structural, not rhetorical. Third, the regulatory filing. The moment a financial regulator references ChatGPT in a formal guidance document, the risk profile of this whole narrative changes. That will be the true tell. That, and the first widely publicized hallucination in a legal contract. Until then, I remain what I have always been: the person who reads the smart contract before the summary, the person who checks the slippage before the yield, the person who asks for the audit report before the deposit. The new model may be remarkable. The slider may be a revolution. The Think button may actually teach the free tier what depth feels like. But none of it is verified. None of it is ledger-backed. And in a market that has burned more investors on confidence than on computation, that is the only fact that matters. Minted in hope, burned in regret. The next burning — if it comes — will not be caused by the model's error rate. It will be caused by our willingness to believe a percentage without a source. That is the one error that no slider can fix.