OpenAI's top model at $4/$20 is a three-month answer to a permanent build decision

Arena's scoring is the part worth sitting with. According to the benchmark platform, Sol's results on agentic coding and work benchmarks improved when the price came down ^[7]. The weights did not change. A leaderboard that moves on a rate card is measuring cost per finished task, which is the unit procurement uses as well.

The headline reduction of more than 20% ^[1] describes the input leg. Output came down by a third ^[14]. On a job with equal input and output volume, the meter reads about $24 per million-token pair where it read $35, roughly 31% off ^[15]. Agent traffic skews heavily to output, so for the workloads OpenAI most wants to win, the real discount sits nearer a third than a fifth.

Then there is the term. A buyer signing twelve months on the strength of this rate gets one quarter at the new price and three at the old one, a blended $32.25 per million-token pair, or 7.9% off list ^[16]. Set that against Gartner's estimate that agents can push inference costs up to fivefold ^[8]: a third off the output leg turns 5x into 3.3x ^[17], and only until the promotion lapses.

The Ramp figures explain the shape of the problem OpenAI is trying to price its way out of. Sol already accounted for 25% of tokens among OpenAI's Ramp customers, about four times the share Anthropic's flagship held on its side of the ledger ^[10][20], and Anthropic's less powerful Opus 5 has passed that flagship in business spending outright ^[11]. Sol now undercuts Opus 5 by 20% on both legs of the meter ^[18]. But buyers did not sort themselves down the tiers because the top tier looked expensive in isolation. They did it because the tasks did not require it, which is the same reason the newsletter credits for the traction of small, domain-specific and open-source models ^[13]. A cheaper premium tier does not make the task harder.

Thomson Reuters sits on the other side of that trade, having trained a domain-specific model on its proprietary data to own the intelligence layer of its business, with lower costs listed as a possible benefit rather than the reason ^[12]. That asset does not reprice when a quarter ends. Meanwhile OpenAI's own lightweight tier took an 80% cut last month, alongside 20% off the mid-tier ^[5], and Tomasz Tunguz of Theory has argued on X that the lightweight model is now competitive with DeepSeek's latest on intelligence per dollar ^[6]. Read the two moves together and the company is defending the floor of its own ladder while renting out the top.