An $800 bill recently circulated in developer circles: one team ran GPT-5.6's flagship model for a full day, and the boss's face changed on the spot. We've noticed this isn't an isolated case—OpenAI has split its flagship model into three tiers (Sol, Terra, Luna), then stacked six levels of "reasoning intensity" (letting the model think longer before answering), yielding nearly 30 combinations. The cost of a single wrong choice: a 5x cost gap and 10 extra seconds of latency, while answer quality doesn't necessarily scale accordingly.

What this is

The three models are tiered by capability: Sol is the flagship (running complex code, deep reasoning), Terra is the workhorse (covering 80% of daily tasks), Luna is the budget option (handling high-frequency lightweight tasks like classification and summarization). Input pricing per million tokens (a token roughly equals one Chinese character or half an English word) is $5/$2.50/$1; output costs 5x more.

Each tier comes with six reasoning intensity levels, from none to max. Higher levels mean the model thinks longer—but accuracy doesn't scale linearly. Jumping from high to max doubles or triples cost again, with marginal quality gains; in some cases, overthinking actually causes errors.

Industry view

One camp argues this is a good thing: model tiering lets enterprises pay-as-you-go like cloud services. Terra + medium suffices for most business needs—this is true AI democratization.

But counterarguments exist. AI architects commonly worry that 30 combinations essentially move "tuning anxiety" from the data center into enterprise IT departments. Small and mid-sized businesses without dedicated optimization teams typically face only two outcomes: blind-picking the most expensive tier (the $800 lesson), or conservatively retreating to Luna and getting poor results.

The deeper read: OpenAI is using pricing to force users to "do the math." A max-tier call can take 2–5 minutes and cost several dollars—only high-value, low-frequency tasks deserve that cost. The biggest waste in the AI era may be "using a cannon to kill a mosquito."

Impact on regular people

For enterprise IT: budget structures must change. The old "one AI line item" needs to split into "infrastructure + inference optimization" two-line accounting—otherwise the $800 story will repeat.

For individual workers: white-collar employees may face a new work scenario—"is your AI usage cost-effective?"—as operations, customer service, and content roles see AI tool costs counted into performance metrics.

For consumer markets: no direct impact yet. Until an AI app price war breaks out, consumer-facing ChatGPT-style products remain uniformly priced.