On July 10, OpenAI publicly released GPT-5.6. The model is strong, Codex is gone, pricing comes in three tiers — all worth reporting. But not the model specs.
I noticed a detail that most coverage of GPT-5.6 overlooked, distracted by the Sol/Terra/Luna solar-system naming, nostalgia over Codex’s disappearance, and the “four times cheaper than Fable 5” value pitch. None of that matters. The truly important thing hides in one detail of the release — OpenAI replaced its billing logic from “sell a monthly ticket” to “install a water meter.”
The facts first. After GPT-5.6, the new ChatGPT desktop app packs Chat, Work, and Codex modes into a single interface. The door is free — even the free tier enters. But once inside, every task draws from a usage pool. OpenAI’s billing docs are clear: ChatGPT Work, Codex, and the Excel agent share the same usage pool. Each task burns credits proportional to scale and complexity; heavier work drains faster. Conversation, notably, does not drain the meter. Plus, Pro, and Enterprise subscriptions still exist — but their role changed. Subscription tiers now determine your pool size. When the pool runs dry: buy credits, upgrade, or wait for reset.
Previously the membership fee was the product itself — pay and chat freely. Now membership is an admission ticket. The real pricing power has been handed to the water meter. Alongside this, OpenAI set three-tier model pricing: flagship Sol at $5/$30 per million tokens (level with GPT-5.5); Terra at $2.50/$15 (half-price volume push); Luna at $1/$6 (floor-price scale play). The strongest model holds its price, the middle tier goes half-price, the volume tier hits the floor — an arrangement with only one explanation: it is built for scale.
The Adobe Precedent — and Why This Time Is Different
If “water meter” sounds like an exaggeration, compare Adobe’s 2013 playbook. Adobe killed the perpetual-license CS suite and forced Creative Cloud subscriptions. The outrage was deafening; some predicted suicide. Office later became 365. The result? Subscription pricing rebuilt the entire software industry’s market cap. That time, the shift was from perpetual license to membership fee. This time, it is from membership fee to water meter. The script barely changed — even the outrage dialogue is recycled.
But there is one fundamental difference: the water meter’s cost structure is completely unlike a membership fee’s. The beauty of membership pricing was that light users’ fees were nearly pure profit — pay and don’t use, free money. The water meter inverts this: every character generated burns real compute, so revenue scaling forces cost scaling in lockstep.
OpenAI’s Bind
This drags OpenAI into an extremely awkward position. Per The Information’s S-1 draft review: OpenAI’s 2025 full-year revenue was roughly $13 billion against a net loss of about $39 billion. Q1 2026: $5.7 billion revenue, $3.7 billion cash burn. Its own February budget: $25 billion to burn in 2026, $57 billion in 2027. The same filing shows a balance sheet with no debt and only $46 million in quarterly capital expenditure — looking like an asset-light company. But another corner of the filing discloses future procurement commitments for chips, energy, and datacenters totaling $665 billion, mostly off-balance-sheet.
See it? To sell work-hours per unit, you must first stockpile astronomical work-hour capacity. That $665 billion is the water-meter model’s sunk cost. And OpenAI’s stated target is an IPO within the year, with Altman repeatedly stating internally that any proposal lowering the trillion-dollar valuation target is “completely unacceptable.”
The three-tier model pricing is the other half of the same mechanism: Sol holds the price ceiling (anchoring perceived value), Terra runs the volume discount (pulling mid-market), Luna hits the floor (flooding the market and building usage habits that feed the meter). Subscription determines the pool; the meter determines the burn; the model tiers determine what a unit of burn buys. Every layer feeds the next — and every layer’s design assumption is that usage will grow. This is not a pricing model for a company hoping customers use less. It is a pricing model that only pays off if customers use vastly more — which is why the $665 billion in procurement commitments exists.
The Deeper Question: Who Bears the Risk?
The water meter shifts risk from vendor to customer in a way the subscription model never did. Under subscriptions, the vendor carried the risk of heavy users consuming more than their fee covered. Under the meter, the customer carries the risk of usage spikes, unexpected complexity, and model behavior changes that increase token consumption. That risk transfer is subtle but profound — and it changes how enterprises must budget for AI. Instead of a predictable monthly line item, AI becomes a variable cost that tracks workload in real time, requiring the same monitoring and optimization discipline as cloud computing. Finance teams that never thought about “AI cost management” will need to start, the same way they learned to manage AWS bills a decade ago.
There is also a competitive dimension. If OpenAI’s meter succeeds, it validates usage-based billing for the entire AI industry — and every competitor will follow. That is good for vendors, who gain revenue predictability. But it is ambiguous for customers: usage-based pricing rewards efficiency (which is good) while punishing experimentation (which is bad, because experimentation is how teams discover AI’s value in the first place). The vendors who solve this tension — who make the meter feel predictable while still capturing usage growth — will define the industry’s pricing norms for the next five years.
For the software industry, the shift is tectonic. If OpenAI’s water-meter model succeeds — and Adobe’s precedent suggests it will — every AI tool vendor faces the same transition: from selling access to selling consumption. The vendors who figure out how to make the meter feel fair while margins hold will own the next decade of enterprise AI. The ones who get the meter wrong will watch customers defect to whichever competitor’s water runs cheaper. Either way, the era of “pay once, use freely” is ending — for AI, and possibly for software as a whole.
