PointFive's 230,000-token coding task produces a fivefold price gap between models

The $1.05 splits into two legs. Reading 200,000 tokens at $3 per million costs $0.60, and writing 30,000 tokens at $15 per million costs $0.45 ^[15][16][4]. Input is 57 percent of the bill even though every output token bills about five times higher ^[17][6]. PointFive's own analogy prices the same burger at $2.62 in India and $9.12 in Switzerland, a spread of 3.5 times ^[12][22]; the same coding task across models spreads five ^[3].

"One task is about a dollar, and that is the number that fools everyone," the post says ^[13]. Its prose gives per-million rates for Sonnet 5 alone, with the rest of the models in a chart, and it does not report how many passes each model needed to finish the task ^[5][24]. That second number decides the ranking. PointFive says real sessions run long, with the agent reading more, retrying, and rerunning its own tool calls ^[14]. At a fivefold spread, a cheapest-in-class model that needs three attempts still costs 60 percent of one clean pass on the priciest, then loses outright past five attempts ^[23]. The only cross-model comparison in the text is that the introductory price puts Sonnet 5 below Gemini and GPT ^[4].

Field data from DX, cited in the index, puts an active engineer at $200 to $600 a month once actual usage is counted. That is $10 to $30 a day, and $2,000 to $6,000 a month for a team of ten ^[8][9]. One of the author's own days, priced at API rates, came to about $240, eight times the top of that daily band ^[10]. Almost all of it was one frontier model doing work a cheaper one could have handled ^[20]. The runaway session PointFive cites reached $81,267 in a week, about 13.5 times a full month of ten engineers at the top of the band ^[11][21].

Caching moves more than the promotion does. If the whole reading pile came back from cache at roughly 10 percent of the input rate, the input leg falls from $0.60 to $0.06. The task falls to $0.51, 51 percent under standard ^[7][19]. The introductory $2/$10 lapses on August 31, 2026, taking the same fixed task from about $0.70 to about $1.05, a 50 percent increase for identical work ^[4][18].

I would use this index for one decision: which class of task gets routed to which model, with the task held fixed ^[1]. The more interesting version of the decision is the opposite policy, a frontier model by default with hard per-session caps. That is defensible if the expensive model finishes in one pass where the cheap one thrashes. The index prices one fixed task, so it cannot tell the two apart ^[1][24]. An organisation that negotiates a rate card and leaves sessions uncapped is buying against $1.05 while its exposure sits at the $81,267 week ^[11].