The index behind these numbers blends three input tokens to one output token, and that ratio carries more weight than it looks [2]. V4 Pro's old flat rate charged twice as much for output as for input; the peak rate charges three times [5][6]. Input-only traffic therefore goes up 3.03x. Output-only traffic goes up 4.55x [2]. At a 3:1 blend it lands on 3.64x, which is where the report's "roughly tripled" comes from [1][9].
Peak covers 01:00 to 04:00 and 06:00 to 10:00 UTC, seven hours, leaving seventeen at exactly half [6][3]. Exactly half is the well-built part. There are no tiers or commitments underneath it, so a cron window and a multiplier of two describe the entire schedule [7]. According to the report the change appeared on the pricing page at 16:00 UTC on August 16 with no notification to callers [5][9], which makes that page a dependency with no change feed.
Both increases came from the same lab. V4 Flash took the same treatment the same day, 3.8x blended, with its output rate up 4.71x [15][6]. The floor line did not move, because it reads the cheapest listed price across every host serving the base model rather than the lab's own page [16]. DeepInfra's $0.09 / $0.18 blends to $0.1125, which is the $0.113 the report prints [4][14]. Two things have to hold before that number belongs in your model: GPQA Diamond at 70 has to proxy the work you actually send [13], and DeepInfra has to have the throughput and latency you need at that price. The report's own line is the right caution, that a lab's list price is a policy while a reseller's is a margin [17].
One ratio in the report does not reconcile. It puts GPT-4-class capability 37x under the flagship ceiling on September 1 [18], but the ceiling it names is Claude Opus 5 at $10.00 blended and the floor it prints is $0.113, which are 88.9x apart [19][14][5]. The two pairs of numbers do not match, so 37x describes some other comparison in the report, not this ceiling and this floor.
Sol's cut is the larger headline and the softer input. Minus 29% is the first time in the index's history that a constituent lowered the price of a model it had already shipped [10]. The sentence holding it up says the promotional rate is available at least through November 21, 2026, and OpenAI names no date on which it goes back up [11]. The sentence sets a minimum duration for the rate, not a date when it reverts.
The month netted minus 5.7%, and the index is down 9.4% since February 23 [3], so a spreadsheet tracking only the average saw a quiet August, helped by three flagship handovers where each successor inherited its predecessor's list price and moved nothing [4]. Underneath the average, two of ten flagships now print a number that is only the top of a range [12], and the flagship spread narrowed from 21x to 13x from both ends at once as Sol's cut pulled the top down and V4 Pro's peak rate lifted the bottom onto Mistral Large 3 at $0.75 [19]. An assumption that held across 40 daily readings [1] now needs a UTC clock and an expiry date to express.
