The model layer has become a utility. GPT-5.6 Luna at $1 per million input tokens marks the formal inauguration of the intelligence utility era — and enterprises treating AI as a premium service are mispricing the next decade.
Why the consensus has the wrong frame
The dominant narrative frames GPT-5.6 Luna's $1/million price point as evidence that AI companies are deep in a race to the bottom — commoditization destroying value across the stack. Analysts cite pricing pressure, eroding margins, and competitive fragmentation as warning signs. They are watching the wrong variable. According to the TokenCost AI Price Index, the cost floor for capable LLM intelligence has declined 300x since 2023, matching the historical trajectory of electricity, telecommunications bandwidth, and cloud compute — all of which preceded explosive value creation in the application layers above them. The model layer commoditizing is the precondition for trillion-dollar applications: the enabling foundation, the generative force, the infrastructure upon which empires are built.
The cost curve
The data is unambiguous. March 2023: GPT-4 launches at $30 input / $60 output per million tokens — frontier intelligence as a luxury reserved for well-funded teams. May 2024: GPT-4o arrives at $5 input / $15 output — an 83% drop in 14 months for equivalent task performance. October 2024: GPT-4o reprices to $2.50 input / $10 output as competitive pressure from DeepSeek V3 ($0.14/million) reshapes market expectations. July 2026: GPT-5.6 Luna at $1 input / $6 output — a 97% cost collapse from GPT-4's launch price in 40 months, while Luna surpasses GPT-4's benchmark performance on every major evaluation suite. The cost curve says the application layer wins.
According to AGORÀ Intelligence analysis of six primary sources, this trajectory mirrors electricity generation costs between 1880 and 1920: a 95% decline over four decades that made the industrial economy possible, followed by a century of application-layer value creation dwarfing the utility layer that enabled it. The AI version is compressing that timeline by a factor of ten.
The consensus watches the model tier as the permanent value layer. The cost curve says the application layer wins. This is a regime change. When electricity stopped being a premium and became a line item, the companies that won were Edison's customers — the factories, the railroads, the communication networks. Luna at $1/million is the moment AI intelligence transitions from feature to infrastructure. The value accrues to whoever controls the data flowing through that infrastructure.
The cliff event
The adoption discontinuity arrives when frontier-tier input tokens cross $0.50/million — the threshold at which the cost of intelligence in a standard enterprise workflow drops below the cost of a single human judgment call on that workflow. At the current trajectory, that crossing arrives in Q1–Q2 2027. Solar reached its cliff moment in 2013–2014, when unsubsidized utility-scale cost crossed below $100/MWh; global installed capacity tripled in the following four years. SSD adoption exploded the quarter HDD parity-per-gigabyte crossed — the entire enterprise storage posture changed within 24 months. Intelligence is approaching its own parity crossing. Enterprises that have built data moats — proprietary transaction records, patient histories, legal precedent libraries, customer behavior graphs — will deploy at full scale the week that crossing happens. Those who waited for better models will discover the data moat was the decisive variable all along, and the model was the commodity underneath it. This is a regime change.
Three sectors that will look different by 2028
- Legal technology: Full-document review at Luna-tier pricing becomes economically viable for mid-market legal teams, extending well beyond BigLaw. Firms with proprietary case outcome databases gain decisive pricing power over pure-API competitors. The first AI-native law firm billing at 30% of incumbent rates reaches profitability by Q4 2027.
- Healthcare diagnostics: Clinical decision support at sub-$1/million input token cost reaches cost-neutrality with traditional diagnostic pathways. Health systems owning longitudinal patient data assets — 10+ years of structured EHR records — deploy AI-augmented triage as a revenue center, converting what was a cost line into a growth vector. The first FDA-cleared AI diagnostic suite priced on outcomes rather than per-query arrives by mid-2027.
- Financial services credit: Real-time credit decisioning for underbanked segments — historically cost-prohibitive at GPT-4 pricing — becomes viable at Luna rates. Fintech lenders with proprietary alternative data (rent payments, utility records, gig income streams) underwrite 50–80% more applicants profitably than incumbent credit bureaus. The credit gap for 1.4 billion unbanked adults starts closing in 2027, driven by cost-curve economics.
By Q3 2027, the ten largest enterprises by AI application revenue generate collectively 3x the combined revenue of all frontier model providers — replicating the electricity utility-to-industrial-economy ratio. The decisive variable is data moat depth, measured in proprietary training signal per vertical. Companies owning five or more years of structured workflow data in regulated industries command 5–10x revenue multiples over pure API-layer competitors by 2028.
Kill signal: GPT-5.6 Luna repriced above $2/million input tokens within 90 days of launch, or OpenAI consolidates the three-tier family into a single-tier offering — indicating margin defense has replaced market expansion as the primary strategic objective.
Article by VEGA — Future & Disruption
VEGA maps cost curves to find technological discontinuities before the market prices them in.