GPT-5.6 Sol and Luna: The Marketing Math Behind 'Unlimited' Free Reasoning

On August 7, OpenAI reportedly pushed GPT-5.6 Sol to Plus and Pro subscribers, while GPT-5.6 Luna appeared on the Free and Go tiers. Sandwiched between the two model names is an internal evaluation stat: Luna cuts fact errors by 62%, Sol cuts them by 68% on finance, medical, and legal questions. Two models. Different names. Nearly identical error reduction. That is not a coincidence. It is a tell. The official framing implies a breakthrough in reasoning. The actual change, if the report is accurate, is more likely a product-layer integration: one base model with a dynamic inference budget, exposed through a slider and a Think button. The same model supposedly supports instant responses and deep reasoning, with users deciding how much cognitive effort each reply receives. Free users get unlimited text chat, while file uploads, image tools, and other multimodal features remain capped. Work and Codex, which use GPT-5.6 Sol, will not change with this release. That rollout pattern matters. It tells you this is not strictly a model upgrade; it is a pricing strategy wearing a model name. Let's dissect the mechanics like a systems engineer, not a headline reader. For years, the dominant pattern was to ship two separate models: a fast one for latency-sensitive tasks, and a slow, heavier one for complex reasoning. Google does this with Flash and Pro. Anthropic does it with Sonnet and Opus. OpenAI's described approach collapses that binary into a single state machine. The base weights are likely shared. The only variable is the number of inference steps allocated before tokens are emitted. A slider that adjusts "how much thinking per response" is not a new neural architecture. It is an inference-time compute budget. Luna and Sol probably share the vast majority of parameters. The 62% and 68% error reductions are close enough to suggest the difference between them is not underlying intelligence, but default reasoning budget and output calibration. Math doesn't care about marketing narratives. But relative improvements are dangerous numbers. A 62% reduction from a baseline of 50 errors per 100 leaves 19 errors. A 62% reduction from a baseline of 5 leaves 1.9 errors. The user experience is orders of magnitude apart, yet both can be advertised as "fact error reduction of 62%." The summarized report omits the absolute baseline, the test set composition, the sample size, and whether the evaluation was human-rated or automated. In my audit work, whether I am checking a ZK proof aggregation circuit or a model card, I demand the denominator. No denominator, no conclusion. This is not cynicism. It is basic verification discipline. The second signal is hidden in the cost structure. Free users get "unlimited text chat" and a Think button. Multimodal tools remain throttled. That asymmetry reveals OpenAI's unit economics: plain text inference, especially with a low reasoning budget, is cheap enough to give away. Images and files are not. The Think button is the paywall in product form. Free users can sample deep reasoning, but the default slider position for a free account will be set low. If the slider allows users to crank thinking effort to maximum, each request could consume several times the compute of a standard answer. Unlimited chat with a Think button is not an act of generosity. It is a tiered pricing mechanism redesigned as a UX feature. Unlimited also needs to be audited. The original report mentions anti-abuse restrictions. In practice, "unlimited" always means "unlimited up to a rate limit that the provider can change at any time." There will be per-account request caps, IP-based throttling, and dynamic queueing during peak load. The math of physical infrastructure does not permit true infinite throughput. Liquidity is an illusion until it's withdrawn. The same is true for free inference. Don't confuse a promotional phrase with an infrastructure guarantee. Now look at the missing specification sheet. Parameter count, activation parameters, context window, training data, and total compute are absent. For a model claiming to halve error rates on high-stakes professional questions, that is a glaring gap. The version numbers themselves are suspicious. GPT-5.6 Sol, GPT-5.6 Luna, and GPT-5.5 Instant do not align with OpenAI's public naming rhythm. It is possible the article is built on leaked or fabricated information. That does not automatically invalidate the strategic logic, but it raises the confidence ceiling. No third-party benchmark, no open evaluation set, no independent replication. The entire narrative rests on an internal promise. In crypto, we call that a trusted setup with no ceremony. The "community governance" problem is hiding in plain sight. OpenAI's internal evaluations are not public. There is no way to verify that the 62% and 68% figures were measured on a representative sample of legal, financial, and medical questions. In open-source protocols, we rely on reproducible benchmarks and community governance to validate consensus-level claims. Here, we are asked to accept a vendor's self-report. That is not engineering consensus. That is brand trust. Smart contracts execute. They don't rationalize. LLMs do both, which is exactly why they need adversarial evaluation rather than marketing-adjusted confidence. The contrarian angle is sharpest where the marketing is loudest. Vertical accuracy in finance, medicine, and law is a double-edged sword. If GPT-5.6 truly reduces errors in those fields, it accelerates adoption in high-value workflows. If the improvement is a relative metric on an easy test set, then real-world deployment will produce confident, articulate, and still-wrong answers in places where the cost of being wrong is catastrophic. Reasoning is not factuality. A model can deploy more tokens to construct a more coherent argument for a false conclusion. The Think button increases the complexity of the rationalization machine. That is not a safety improvement. It is a new risk class. Free access also expands the attack surface. Prompt injection, data extraction, and social engineering attempts will scale with the user base. Every free text exchange generates preference signals, correction logs, and adversarial probes from real users. That feedback stream is worth more than a benchmark score. OpenAI's rivals may have model quality; they may not have the same volume of organic free-tier interaction data. That asymmetry compounds over time. But it also creates liability. Users will share sensitive information more casually because the product is free and unlimited. Any leak in training data or conversation memory carries amplified consequences. The ethical exposure grows in direct proportion to the user base. There is also a deeper infrastructure story. A free unlimited text tier implies that OpenAI's marginal inference cost has dropped to a level that makes customer acquisition via giveaway viable. That could come from better hardware, optimized kernels, speculative decoding, KV-cache management, or a combination of techniques. It could also come from tighter default reasoning budgets. The name pair Sol and Luna, day and moon, suggests round-the-clock load balancing. The system may route different reasoning workloads across different infrastructure buckets depending on demand. That is a clever capacity management strategy, but it is not a new model parameter. It is an operational layer. The most important investment signal is not the error reduction. It is the willingness to sacrifice short-term gross margin for user scale and data moat. OpenAI is effectively saying that raw model leadership is less important than controlling the consumer entrance point. If free unlimited text becomes a habit, paid conversions follow through the Think button, file uploads, and tool access. The data flywheel spins faster with every free conversation. That is an aggressive, defensible strategy. It is also a cash-burning one if the cost curve does not cooperate. What should builders watch? Not the 62% or 68% number. Watch the cost per million tokens on the API. Watch whether OpenAI publishes a model card with absolute accuracy rates. Watch whether third-party evaluators can reproduce the internal results. If the API price stays flat while the free tier expands, OpenAI has achieved a structural cost advantage. If the free tier silently degrades under load, the "unlimited" promise was always a loan against future infrastructure. On August 7, the stage belonged to Sol and Luna. The actual story is deeper: a unified model with variable thinking, a free tier with hidden limits, and a data moat built from user interaction. The version number could be wrong. The direction is not. The question that matters is not whether GPT-5.6 Sol and Luna are real. It is whether the economics of unlimited inference survive contact with actual demand. If they do, the AI market shifts permanently. If they do not, the rollout will be remembered as the moment "unlimited" became a throttled product feature with a friendly interface.