The $0.00799 fine-tuning rate in your Bedrock budget belongs to Llama 2

The filter is the whole experiment. `list_foundation_models(byCustomizationType="FINE_TUNING")` returns only the models an account can train in that region; drop the filter and you get the invoke catalogue, which is much longer and answers a different question ^[12]. That is also why the console agrees with the sweep. The us-east-1 fine-tuning picker on 8 September 2026 offered one provider, Amazon, and the same six models the API returned ^[11][9]. The gap is in the pricing page, not in the catalogue.

Thirteen of the 33 regions never gave an answer at all. They returned `UnrecognizedClientException`, which is what an opt-in region returns to an account that has never enabled it, so their zero is not evidence of absence ^[13]. That leaves 20 real answers ^[7]. Two of those 20 came back non-zero ^[15] and the write-up counts seventeen zeros ^[16], which leaves one region unaccounted for in the tally ^[14]. It does not move the conclusion: one of the 20 answering regions can train Llama 3.3 70B, or 5% of them ^[19].

Now the number people are budgeting from. For $0.00799 per 1,000 tokens to be the Llama 3.3 70B training rate, AWS would have to charge for a Llama 3.3 job exactly what it charges for a Llama 2 70B job, and the customization page does not say that; it lists Llama 2 and stops ^[4][1]. The table is not even consistent in units, since two models on it are priced by the hour ^[2]. Meanwhile the capability exists and produces a line item: the author reports a fine-tuned Llama 3.3 70B in production on Bedrock and a training charge on the bill, with no published price to check it against ^[3].

What the evidence does not establish is the actual rate, or how far $0.00799 sits from it. It could be identical, and nothing published settles that either way, which leaves the invoice from your own first job as the only number you can defend ^[5].

On durability, two runs five weeks apart returned the same table ^[17]. That is a narrow window to generalise from, and the author says so, though the probe reproduces in about a minute ^[18]. The planning consequence is narrow too. A Llama 3.3 70B fine-tune is a us-west-2 job, alongside Claude 3 Haiku and the Llama 3.1 pair ^[8], London offers exactly one tunable model and it is not a Llama ^[10], and the first training job you run is the price list ^[5].