Amazon Nova 2 arrives on Bedrock with Lite, Pro and Omni tiers priced below rival models

Amazon has shipped the second generation of its Nova foundation models, splitting the family into three variants — Lite, Pro and Omni — that developers can call through Amazon Bedrock. According to a developer guide published on dev.to, AWS is aiming the family at production AI applications where inference cost, throughput and multimodal handling matter as much as raw capability.

A family split by workload

The three models sit at different points on a cost-versus-capability spectrum. Nova 2 Lite is tuned for high-volume, low-cost workloads; Nova 2 Pro targets deeper reasoning and complex multimodal tasks; and Nova 2 Omni is built for what AWS calls Any-to-Any processing, where an input in one modality can produce outputs in several others.

Across the board, the dev.to guide reports longer context windows, support for more than 200 languages, higher throughput, lower inference costs and speech understanding compared with the original Nova line. A shared new feature is extended thinking, which lets developers dial reasoning depth up or down per request, trading answer quality against speed and cost.

Lite for volume, Pro for reasoning

Lite is aimed at bread-and-butter workloads: chatbots, customer support, classification, summarization, document processing, content generation and straightforward agent automation. The notable change from the previous Lite, per the guide, is the addition of extended thinking.

The benchmark claims come from Amazon itself, as relayed by dev.to: Nova 2 Lite reportedly matched or beat Claude Haiku 4.5 on 13 of 15 benchmarks, GPT-5 Mini on 11 of 17, and Gemini Flash 2.5 on 14 of 18. The guide cautions that results depend heavily on the task, evaluation methodology, prompting and configuration, so they signal which category of model Lite competes in rather than proving universal superiority.

Pro, the higher tier, accepts text, images, video and speech, and carries a one-million-token context window — enough for long documents, enterprise knowledge bases, large codebases, research workflows and multi-step agents. It can also serve as a teacher model for knowledge distillation, helping produce smaller specialized models. Amazon's comparisons put Pro equal to or better than Claude Sonnet 4.5 on 10 of 16 benchmarks, GPT-5.1 on 8 of 16, Gemini 2.5 Pro on 15 of 19, and Gemini 3 Pro Preview on 8 of 18.

Omni goes any-to-any

Omni takes a different route. Rather than being simply a stronger reasoning model, it is designed for workflows where an input such as text can yield text, image, speech or video outputs. The guide lists visual understanding, video analysis, image editing, media pipelines and multimodal agents as candidate use cases, but is blunt that Omni is not the default choice: text-heavy, cost-sensitive applications fit Lite, and reasoning-heavy applications fit Pro. Omni earns its keep only when mixing modalities is central to the product.

Pricing is the pitch

According to a GoML comparison cited by dev.to, Nova 2 Lite costs roughly $0.30 per million input tokens and $2.50 per million output tokens, while Nova 2 Pro runs around $1.25 and $10 respectively. The same comparison lists Claude 4.5 Sonnet at $12-plus per million tokens on both sides, GPT-5 Mini and 5.1 in the $8–10 range, and Gemini 3 Pro at about $18. The guide warns that pricing shifts over time, so developers should verify current Bedrock rates before building a cost model.

It also argues that sticker price per token is not the whole story: latency, throughput, output quality, retry rates, tool-call reliability and cost per successful task all count. A cheap model that needs several attempts to finish a job correctly can end up costing more than a pricier one that succeeds on the first pass.

Agent workflows in focus

Agentic AI is one of the sharper use cases the guide highlights. An agent that must retrieve information, reason over results, call external tools, inspect documents or images and decide its next step strains older and smaller models. The guide suggests Nova 2's controllable reasoning makes those multi-step chains more viable, and that the unified multimodal support means fewer separate models stitched together per workflow — the practical upgrade over Nova 1, which offered limited reasoning, fewer languages, smaller context, little or no speech support and fragmented modality handling.

Why it matters

Amazon is competing for Bedrock workloads on price-performance rather than headline capability alone. If Lite and Pro hold up near their claimed benchmark parity — and the numbers are vendor-supplied, so teams should re-test against their own tasks — the token-price gap against Anthropic, OpenAI and Google models is wide enough to reshape cost models for high-volume applications. The three-tier split also gives teams a cleaner upgrade path: start on Lite, move to Pro when context length or reasoning depth becomes the bottleneck, and reach for Omni only when modality mixing is the product itself.