FactQuire

AWS quietly doubled Bedrock prices for legacy Claude models — and cost trackers haven't caught up

2026-07-04

AWS quietly doubled Bedrock prices for legacy Claude models — and cost trackers haven't caught up

If your cost dashboard says a Bedrock call to Claude 3.5 Sonnet v2 costs $3 per million input tokens, your dashboard is seven months out of date — and it is off by exactly 2x.

What changed

Amazon moved Claude 3.5 Sonnet v2 on Bedrock to what its pricing page calls Public Extended Access, effective December 1, 2025. The AWS Bedrock pricing page (https://aws.amazon.com/bedrock/pricing/, accessed 2026-07-04) lists for Anthropic Claude 3.5 Sonnet v2:

> "$6.00" per 1M input tokens, "$30.00" per 1M output tokens, > cache write "$7.50" per 1M, cache read "$0.60" per 1M > (Public Extended Access, effective December 1, 2025)

That is double the long-familiar $3.00 / $15.00 — the price most of the ecosystem memorized in 2024 and hardcoded everywhere.

Why it went unnoticed

Prices that change on a provider's page do not change in the JSON files, cost packages, and dashboards that were built when the old price was true. Facts rot silently.

While cross-auditing community datasets against primary sources on 2026-07-04, we found Helicone's cost package still carried $3.00 / $15.00 for `claude-3.5-sonnet-v2:bedrock`. The code is correct code — it is simply code that does not know the world changed underneath it. We verified the current values against the AWS pricing page and submitted the correction upstream: Helicone/helicone#5709.

Notably, the entry's cache multipliers (1.25x write, 0.1x read) were still exactly right against the *new* base price — evidence the structure was sound and only the base facts had drifted.

The FactQuire entry

Our machine-readable record for this model, with the verbatim source quote and access timestamp, is here:

Every field in the FactQuire feed carries a quote from the primary source and the time we read it, so anyone can re-verify the claim without trusting us — including when we are the ones who turn out to be wrong. (Our own published corrections are at https://factquire.com/corrections.html.)

The general lesson

Model pricing is not static reference data; it is news. Prices change with an effective date, quietly, on a page nobody re-reads. Any system that stores a price without storing *when and where it was read* will eventually report costs that are confidently wrong — in this case, by 2x for seven months.