AI・テクノロジー
Amazon "Nova Pro 2.0" Officially Launched — 68% Reduction in Inference Costs Reshapes Cloud AI Procurement
機械翻訳 / Machine-translated
AWS officially launched its in-house LLM "Nova Pro 2.0" on August 21, 2026. The model achieves an MMLU score of 91.7 while cutting input token costs by 68% compared to the previous generation, and comes standard with multimodal capabilities including a 128K context window and support for up to 30 minutes of video. Availability through Amazon Bedrock began simultaneously across all regions, giving companies with existing AWS infrastructure an immediate opportunity to rethink their cost structures.
On August 21, Amazon Web Services officially launched its proprietary LLM "Nova Pro 2.0" through Amazon Bedrock, available in all regions.
Key specifications are as follows:
Immediately after launch, real-world reports from developers flooded X (formerly Twitter).
Nova Pro 2.0 delivered output at the same level as GPT-4o with the same prompt, but clearly cheaper. Being able to run it on Bedrock means existing IAM management carries over directly — that's a major practical advantage. (Engineer building enterprise AI infrastructure)
AWS first unveiled the Nova series at re:Invent in November 2025, adopting a three-model lineup of Nova Micro, Nova Lite, and Nova Pro, with a strategy prioritizing Bedrock optimization and price competitiveness. Nova Pro 2.0 represents the second generation of that lineup.
Underlying this release is the "legibility" problem of AI costs. While large-scale enterprise AI pilots began in earnest in the latter half of 2025, cases of opaque monthly cloud bills were on the rise. Nova Pro 2.0's 68% cost reduction is seen as AWS's own answer to this challenge.
Because it runs through Bedrock, existing AWS IAM policies apply as-is. Model access control, log collection, and cost allocation can all be managed at the AWS Organizations level, and the ability to use a high-precision LLM without additional contracts is a real-world differentiator.
Use cases that previously required separate specialized models — such as meeting recording summarization, manufacturing line anomaly detection, and medical imaging diagnostic assistance — can now be covered by Nova Pro 2.0 alone. The more unstructured data an organization holds, the greater the benefit of consolidating to a single model.
An input price of $0.0008 / 1K tokens represents a 3–4x cost advantage over GPT-4o ($0.0025) and Claude Sonnet 4.6 ($0.003). As performance levels across providers converge, this gap is increasingly likely to become the central axis of procurement decision-making.
Nova Pro 2.0 presents a clear economic rationale in response to the question of where to source cloud AI. In 2026, as the practical quality of LLMs across providers approaches parity, the basis for selection is shifting from "intelligence" to "cost, governance, and existing integration." It is fair to say AWS read that shift and set today's specifications and pricing accordingly.
For Japanese companies in particular, the video support deserves close attention. Many large enterprises hold vast quantities of unstructured data in video format — meeting recordings, training footage, equipment inspection videos — and reaching a practical level where a single API can process both video and text could serve as a trigger that accelerates adoption decisions.
That said, caution is warranted. Much of the benchmark data is based on AWS's own measurements, and the figures should not be taken at face value until independent evaluations from third parties are available. The step of running task-specific benchmarks — for code generation, long-form summarization, multilingual support, and so on — in your own environment cannot be skipped.
Nova Pro 2.0 is a structural bet by AWS to win the AI procurement race not on performance, but on cost and integration. Organizations that already have an environment on Bedrock for cross-model comparison are well served by running a real-world comparison against equivalent workloads this week.
Where is your organization's AI spending currently concentrated, on which models, and for what reasons? — The arrival of Nova Pro 2.0 is also an opportunity to revisit that question. The next phase to watch is whether Azure and GCP will follow suit with equivalent cost levels.
This article was written by an AI writer (AI News) from the Mirai News editorial team.
