Buying direct from the developer is not always the cheapest route. 22% of models cost more from their own developer than from independent hosts, by a median 24.2%. In May it was 35.7%. On $10M, that premium is $2.4M a year.
https://t.co/xilBIoB20u
#AtticStandard
Gemini 4 Hydrogen from @GoogleDeepMind enters our assessment at $0.0015 input, $0.0075 output per 1,000 tokens. It is the core market’s premium option: top 2 of 32 on input at 5x spot, top 2 of 32 on output at 6.4x.
https://t.co/cU8tHXyqlU
#AtticStandard
Gemini 4 Hydrogen entered our assessment at $0.0015 per 1,000 input tokens and $0.0075 output. Google calls it its everyday model, yet on both it ranks among the two most expensive of 32 core models. One of four vendors charges 25% above list. https://t.co/EZR9Y23H51 #AtticStandard
Gemini 4 Hydrogen from @GoogleDeepMind enters our assessment at $0.0015 input, $0.0075 output per 1,000 tokens. It is the core market’s premium option: top 2 of 32 on input at 5x spot, top 2 of 32 on output at 6.4x.
https://t.co/cU8tHXyqlU
#AtticStandard
35.5% of models sold on a cloud marketplace cost more there than from independent hosts, by a median 64.4%. For those models, $10M of usage at host rates costs about $16.4M on the marketplace.
https://t.co/xilBIoB20u
#AtticStandard
Where you buy an AI model matters. 35.5% of 124 models sold on a cloud marketplace and by independent hosts cost more on the marketplace, by a median 64.4%. $10M of usage there becomes about $16.4M.
How the clause works: https://t.co/xilBIoBzQ2
#AIpricing#Inference
35.5% of models sold on a cloud marketplace cost more there than from independent hosts, by a median 64.4%. For those models, $10M of usage at host rates costs about $16.4M on the marketplace.
https://t.co/xilBIoB20u
#AtticStandard
A channel discount on a flagship model ended this week and its rate there doubled. The flagship index rose 0.64%, as 3 of 363 SKUs repriced. Tied to that seller, a $10M contract pays 100% more; tied to the index, about $64K more.
https://t.co/xilBIoB20u
#AtticStandard
A channel discount on a flagship AI model ended this week and its rate doubled. A contract tied to that seller doubled too. One settled on our flagship index rose 0.64%, about $64,000 a year on $10M.
How the clause works: https://t.co/xilBIoBzQ2
#AIpricing#Inference
A channel discount on a flagship model ended this week and its rate there doubled. The flagship index rose 0.64%, as 3 of 363 SKUs repriced. Tied to that seller, a $10M contract pays 100% more; tied to the index, about $64K more.
https://t.co/xilBIoB20u
#AtticStandard
Gemini 4 Carbon entered our assessment this week at $0.004 per 1,000 input tokens and $0.02 output, at list. A half-price offer runs to November 30, so a $10M commitment priced at the offer costs $20M once it ends.
https://t.co/KJnH6CkyGY
#AIpricing#Inference
Gemini 4 Carbon from @GoogleDeepMind enters our assessment at $0.004 input, $0.02 output per 1,000 tokens: list, not the half-price offer to Nov 30. It sits high in the flagship market, top 7 of 29 on both, at 2x and 3.3x spot.
https://t.co/cU8tHXyYbs
#AtticStandard
Most AI prices sit still for weeks: 3.2% of 1,303 listings changed list price in 13 weeks. But when they move, the median change is 20.1%. On a $10M contract, that’s $2M a year, on a date the seller picks.
How the clause works: https://t.co/xilBIoBzQ2
#AIpricing#Inference
Only 3.2% of AI inference listings changed their price in 13 weeks, but the median change was 20.1%, and one in three was a rise. On a $10M contract, one such move is $2M a year, on a date the seller picks.
https://t.co/xilBIoBzQ2
#AtticStandard
A launch price often doesn’t last. 11.1% of AI models launched 4 to 13 weeks ago already sell below their launch week price, by a median 15.6%. On a $10M contract signed at launch, that’s $1.56M a year.
How the clause works: https://t.co/xilBIoB20u
#AIpricing#Inference
1 in 9 models launched in the last three months already costs less than in its launch week, by a median 15.6%. In May it was 1 in 16. A contract priced at launch keeps paying the launch rate unless it follows the market.
https://t.co/xilBIoBzQ2
#AtticStandard
The Inference Price Benchmark for the week of November 16 is published.
Across our 29 indexes, the largest move came in the flagship tier, up 0.64% on input and output and 0.79% on cached input, its first rise after four weeks of falls.
The rise came from just 3 of the 363 SKUs in the flagship basket, where a discounted channel rate returned to the developer’s list price and doubled. On a $10 million commitment priced on flagship output, the move is worth about $64,000 a year.
Across the whole assessment, 24 of 7,754 SKUs changed list price, 19 of them by 10% or more. The image per megapixel spot fell 50% to $0.015 after 22 of its 63 models were withdrawn from sale, while its benchmark held level.
For anyone agreeing a price today, a discount written into the contract as a fixed term against the benchmark stays in place for as long as the contract runs, whatever happens to channel rates.
How index-linked pricing works for inference contracts: https://t.co/PeznOXL6Ce
How to read the inference market: https://t.co/ZCYgCTkyWF
Live market coverage: https://t.co/EZR9Y24eUz
Newsletter: https://t.co/xq1Llw00lu
#AI #AIpricing #Inference #LLM #GenAI #Benchmark #AtticStandard
Flagship AI prices rose 0.64% this week, the first rise since mid-October, after a discounted channel rate returned to list price. On a $10M commitment priced on flagship output, it’s about $64,000 a year.
The Inference Price Benchmark for the week of November 16 is published.
Across our 29 indexes, the largest move came in the flagship tier, up 0.64% on input and output and 0.79% on cached input, its first rise after four weeks of falls.
The rise came from just 3 of the 363 SKUs in the flagship basket, where a discounted channel rate returned to the developer’s list price and doubled. On a $10 million commitment priced on flagship output, the move is worth about $64,000 a year.
Across the whole assessment, 24 of 7,754 SKUs changed list price, 19 of them by 10% or more. The image per megapixel spot fell 50% to $0.015 after 22 of its 63 models were withdrawn from sale, while its benchmark held level.
For anyone agreeing a price today, a discount written into the contract as a fixed term against the benchmark stays in place for as long as the contract runs, whatever happens to channel rates.
How index-linked pricing works for inference contracts: https://t.co/PeznOXL6Ce
How to read the inference market: https://t.co/ZCYgCTkyWF
Live market coverage: https://t.co/EZR9Y24eUz
Newsletter: https://t.co/xq1Llw00lu
#AI #AIpricing #Inference #LLM #GenAI #Benchmark #AtticStandard
38.1% of 465 token models priced in the week of November 16 charge over four times their input rate for output, up from 35.2% in May. At five times, output is 83% of a bill with equal tokens in and out.
How the clause works: https://t.co/xilBIoB20u
#AIpricing#Inference
38.1% of token models now charge more than 4x their input rate for output, up from 35.2% in May. At 5x, a workload with equal input and output volumes spends 83% of its bill on output. One blended rate hides that.
https://t.co/xilBIoB20u
#AtticStandard