Amazon Nova pricing: every model, checked on AWS.

IHow Amazon Nova pricing works

Nova uses the standard Bedrock pricing model. What makes it different is where its prices sit: at the low end of the catalog by design.

For the text and multimodal models (Micro, Lite, Pro, Premier and the Nova 2 models), the unit is the token, about three quarters of an English word. Each request is metered twice: input tokens (prompt, system instruction, history, retrieved context, and any images, documents or video you send to a multimodal model) and output tokens (what the model writes). Output costs about four to eight times input on Nova.

Canvas (images) is billed per generated image, Reel (video) per second of generated video, and Sonic (speech-to-speech) per million speech and text tokens. Budget those by volume, not by prompt length.

Four things change the effective rate. The service tier: Standard, Flex (cheaper, for work that can wait), Priority (faster, costs more) and Batch (half price for asynchronous jobs). Prompt caching: cached input reads cost 75% less than normal input. Provisioned Throughput: a flat hourly charge per model unit. And customization: a training charge plus monthly storage for a custom model.

IINova prices per tier: input and output, per 1M tokens

The table most people come for. Standard on-demand prices, US East (N. Virginia), checked 10 October 2026 on the <a href="https://aws.amazon.com/bedrock/pricing/">AWS Bedrock pricing page</a>.

IIIFlex, Priority, Batch and prompt caching

The same model can cost half or nearly double the standard price, depending on how fast you need the answer and whether context repeats.

• Batch: Submit many requests as one asynchronous job (usually a file in S3) and pay half: Nova Micro $0.0175 / $0.07, Nova Lite $0.03 / $0.12 per 1M tokens. The default for bulk classification, extraction, summarization and offline evaluation.
• Flex: Cheaper on-demand for work that can tolerate slower responses: Nova Pro $0.40 / $1.60 and Nova Premier $1.25 / $6.25 per 1M tokens.
• Priority: Faster responses at a premium: Nova Pro $1.40 / $5.60, Nova Premier $4.375 / $21.875 and Nova 2 Lite $0.5775 / $4.8125 per 1M tokens.
• Prompt caching: Cached input reads cost 75% less than normal input. For Nova Lite that is $0.015 per 1M cached tokens instead of $0.06. It pays off when a long system prompt, document or tool list repeats across calls.
• Provisioned Throughput: A flat $60.50 per hour per model unit with no commitment, $55 with a 1-month commitment, or $30.25 with a 6-month commitment. Worth it only for steady, high volume or to serve a custom model.

All prices in this section: US East (N. Virginia), checked 10 October 2026.

IVNova Canvas, Reel, Sonic and embeddings

The creative and speech models are not priced like the text tiers. Each has its own unit.

VA worked monthly cost example

Per-token prices are hard to feel until you run a workload through them. The math below uses the standard prices from section II.

The workload: a multimodal support and extraction assistant with 200,000 requests a month. Each request reads a message plus a screenshot (about 2,000 input tokens) and writes about 400 tokens. That is 400M input and 80M output tokens a month.

On Nova Lite: 400 × $0.06 = $24 of input and 80 × $0.24 = $19.20 of output, so about $43 a month. Run it as batch and it is about $22. On Nova Pro: $320 + $256 = about $576. On Nova Premier: $1,000 + $1,000 = about $2,000.

For comparison, an illustrative frontier model at $3 input and $15 output per 1M tokens would cost about $2,400 for the same work. Check the Anthropic rows on the Bedrock pricing page for Claude's current prices. The spread is the whole case for routing: send the easy majority of requests to a cheap Nova tier and escalate only the hard minority.

VIWhy Nova is the cost-optimization pick on Bedrock

If the question is "how do I run generative AI on AWS for the least money without wrecking quality," the answer usually involves Nova, as one tier in a router rather than the only model.

• Right-size each request: Route easy requests to Micro or Lite, the middle to Nova 2 Lite or Pro, and only the hard minority to Premier or a frontier model. The single biggest lever.
• Batch the bulk work: Anything non-interactive at half price.
• Cache repeated context: Long fixed system prompts, shared documents and tool definitions at 75% less on cached reads.
• Distill for narrow tasks: Use Nova Premier as a teacher to train a small custom model for one high-volume task. For reference, supervised fine-tuning on Nova Lite costs $0.002 per 1,000 training tokens, plus $1.95 a month to store the custom model.
• Trim input, cap output: Shorter retrieved context and sensible max-output limits cut token cost directly.

The honest counterweight: Nova is the value pick, not always the frontier pick. Where a task is genuinely hard, paying for a stronger model is the right call. The discipline is "the cheapest model that clears the bar for each task."

VIIWho pays: AWS credits and partner-filed funding

Everything above is what Nova costs if you pay AWS directly. Nova usage is on your AWS bill, so AWS credits apply to it, and because Nova is cheap, credits last a long time.

AWS Activate credits are the ones you apply for yourself, and they work on Bedrock (AWS announcement). Activate Founders is $1,000 to start, up to $5,000, with no investor needed. Activate Portfolio is up to $200,000 with the Org ID from your VC, accelerator or angel group if it is an Activate Provider. A consulting partner cannot apply for Activate on your behalf.

For most startups the larger route is partner-filed funding. A vetted AWS partner can file POC (proof of concept) funding to test a Nova build, or incremental workload (IW) funding if you already run on AWS and are launching a new AI feature. For startups that is $50K to $150K. Depends on the workload; AWS approves.