The Number to Model Is the Cached Input Price
Within roughly three days at the end of September, three frontier labs landed on the same headline price. OpenAI's GPT-6.1 Sol launched at $2 per million input tokens and $10 per million output tokens. Google's Gemini 4 Argon followed with introductory pricing of $2 and $10, rising to $4 and $20 after the introductory period. Anthropic's Claude Sonnet 5.5, released September 28, sits at the same $2 and $10. Three vendors, one price card.
When the headline converges, the selection pressure moves to the line below it. Both Sol and Argon list cached input at $0.10 per million tokens. OpenAI halved that number from GPT-6 Sol's $0.20. For an agent, the cached rate is the one that touches almost every token it sends.
The reason is mechanical. An agent loop resends its accumulated context on every call: the system prompt, the tool definitions, the conversation so far, the results of every previous tool call. A task that runs thirty steps pays for the first step's context thirty times. Most of what an agent bills as input is repeated input, and repeated input is exactly what caching discounts.
I want to put numbers on that, with the assumptions stated. Take a coding-agent task that accumulates 2 million input tokens across its run, and set output aside for the moment. At the headline $2 rate with no caching, that is $4.00 of input. With a 90 percent cache hit rate at $0.10, 0.2 million tokens bill at $2 for $0.40 and 1.8 million bill at $0.10 for $0.18. The input cost comes to $0.58. Caching removes 85 percent of the bill, and cached tokens make up only 31 percent of what remains.
Now move each lever by half. Raising the headline rate by 50 percent, to $3, lifts the input bill from $0.58 to $0.78, an increase of about 34 percent. Raising the cached rate by 50 percent, to $0.15, lifts it from $0.58 to $0.67, about 16 percent. The headline rate has twice the leverage on the final bill, even though cached tokens are nine out of every ten tokens sent. The uncached tenth still carries two-thirds of the cost.
The OpenAI halving shows what the cached rate is worth on its own. At the old $0.20 cached rate, the same task costs $0.40 plus $0.36, or $0.76. At $0.10 it costs $0.58. The cut takes 24 percent off the input bill with the headline rate untouched.
The lever that moves the number most is the one neither vendor sets. Hold the prices fixed and vary the hit rate. At 80 percent cached, the input bill is $0.96. At 90 percent it is $0.58. At 95 percent it is $0.39. Ten points of cache hit rate swing the bill by more than a third in one direction, and five points swing it by a third in the other. How the agent orders its context, how stable the prefix stays between calls, and whether tool definitions get reshuffled mid-run all feed that hit rate, and none of them appear on a pricing page.
Output tokens will change the totals, since they bill at five times the input rate and agents produce plenty of them. I have left them out because the shape of the argument holds either way. The cached rate and the hit rate together determine how input behaves, and input dominates long-running agent traffic.
The cheaper tiers sharpen the same point. OpenAI's lineup, per one secondary source, runs gpt-6-luna at $0.10 input and $0.50 output, against gpt-6-astra at $10 and $50. Anthropic's Haiku 5.5 arrived with headlines about a roughly 75 percent price cut for high-volume work, which I have only seen reported at the headline level. A price list spread across a hundredfold range is a routing problem before it is a pricing problem. Which tier handles which step of a task decides the average cost far more than which vendor's flagship has the lower sticker.
What I take from the convergence is that headline price has stopped working as a selector. Vendors will keep matching each other there, because a gap in the visible number is the first thing a buyer notices. Differentiation migrates to the terms that need arithmetic to compare: cache rates, cache time-to-live, how introductory pricing steps up, and whether the discount applies to the context pattern an agent actually produces. The Argon step from $2 and $10 to $4 and $20 shows how provisional the visible number can be.
A comparison done on the sticker alone produces a ranking that survives only until someone reruns it with real traffic. A comparison done with a stated token profile, a measured hit rate, and a routing split produces a number that stays usable until the next price change. That second number takes an afternoon of arithmetic, and I have not yet seen a vendor slide that includes it.
Sources: TechCrunch, "OpenAI launches GPT-6.1 Sol" (Sept 29, 2026); Implicator on Sol's cached-input price (Sept 2026); Google blog and DataCamp on Gemini 4 Argon (Sept 30, 2026); andrew.ooo, "What changed in the latest OpenAI model release" (Oct 15, 2026); Tech Insider via AF.NET on Claude Haiku 5.5 (Oct 2026). Cost figures are my own calculation under the stated assumptions of 2 million cumulative input tokens per task and the stated cache hit rates.