Cost math nobody runs first

The gap between Lite and Pro is not a rounding error, and you should size it before you commit. Take a support assistant that processes 2 million input tokens and generates 0.5 million output tokens per day. On Lite that is 2 times $0.06 plus 0.5 times $0.24, which is $0.12 plus $0.12, so $0.24 a day. On Pro the same traffic is 2 times $0.80 plus 0.5 times $3.20, which is $1.60 plus $1.60, so $3.20 a day. Micro, if the task were text only, would be about $0.14 a day. Over a 30 day month that is roughly $4.20 for Micro, $7.20 for Lite and $96 for Pro. Pro is thirteen times the daily cost of Lite for this shape of traffic. If Lite answers the tickets as well, that difference is pure waste.

The inference profile trap, and the fix

Back to that 400. For on-demand Nova in most regions, Bedrock does not want the base model ID. It wants a cross-region inference profile ID, which is the base ID with a geography prefix, so amazon.nova-lite-v1:0 becomes us.amazon.nova-lite-v1:0 in the US. The inference profile lets Bedrock route your call across several regions to absorb traffic bursts, and for these models it is the on-demand path. Call the base ID and you get a ValidationException telling you to use an inference profile. Here is the working call with the boto3 SDK and the Converse API, which is the modern, model-agnostic way to talk to Bedrock.

import boto3

client = boto3.client('bedrock-runtime', region_name='us-east-1')

resp = client.converse(
    modelId='us.amazon.nova-lite-v1:0',   # inference profile, not the base id
    messages=[{'role': 'user',
               'content': [{'text': 'Summarize our refund policy in two sentences.'}]}],
    inferenceConfig={'maxTokens': 400, 'temperature': 0.2},
)

print(resp['output']['message']['content'][0]['text'])
print(resp['usage'])   # inputTokens, outputTokens, totalTokens

Expected output: a two sentence summary, then a usage dict such as inputTokens 41, outputTokens 38, totalTokens 79. Failure mode: set modelId to amazon.nova-lite-v1:0 and the same code raises a ValidationException that says to use an inference profile. Swap in the us. prefixed ID and it clears. The usage block is also where your cost tracking should start, because those token counts feed straight into the price math above.

Regions and availability, the part that bites late

Nova is not in every region, and this is where a design that looked fine on a whiteboard fails in the account. As a base model, Premier runs only in US East (N. Virginia) and AWS GovCloud (US-West). Pro, Lite and Micro are broader, adding Asia Pacific (Sydney), Asia Pacific (Tokyo) and Europe (London) on top of N. Virginia and GovCloud. Cross-region inference profiles stretch that reach much further, to Ohio, Oregon, Ireland, Frankfurt, Paris, Stockholm, Singapore, Seoul and Mumbai, by routing a call to whichever region has spare capacity. The creative and speech models sit in their own footprints: Canvas and Reel in N. Virginia, Ireland and Tokyo, and Sonic in N. Virginia, Stockholm and Tokyo.

That routing is exactly what bites a regulated workload. The same inference profile that saves you during a traffic burst can send a request to a region your compliance team never signed off on, so if data residency is a hard requirement, treat cross-region inference as something to constrain, not a free win. Two habits save real pain here. Check model availability per region before you pick a deployment region, because a model that shows up in the console can still refuse in the region you chose. And confirm the residency behavior of any profile you enable, since the whole point of a profile is to move your traffic around. A model ID that works in a demo account in N. Virginia is not a promise it works in eu-west-2 next quarter.

Where Nova fits against Claude and Llama

Nova does not win every task, and pretending otherwise wastes your time. On the hardest reasoning and coding, the top Anthropic Claude models in Bedrock still lead, and for some open-weight and fine-tuning work a Llama model is the better base. What Nova wins is the price-to-quality ratio on the ordinary majority of work: classification, extraction, summarization, routing, straightforward chat, and multimodal reads where you feed a document image and ask a question. For those, Lite is hard to beat on cost, and Micro is cheaper still when the input is pure text. Premier earns its keep in two narrow spots: when you truly need a 1M token context in one call, and when you are distilling a smaller custom model and want the strongest teacher.

My verdict: make Nova Lite the house default and justify every deviation. Route the high-volume text-only paths to Micro, escalate only the measured-hard prompts to Pro, reserve Premier for the two cases above, and move new builds onto Nova 2 where an equivalent exists. If you want the vendor-neutral view of how these choices generalize, the GenAI Series covers the model-selection logic without the AWS labels. For how Nova sits in the wider Bedrock catalog next to the third-party models, go back to Part 3.

Which Nova to reach for first

Nova is the value tier of Bedrock, and treating it that way pays off. Learn the four understanding models by modality and context, default to Lite, keep Micro for cheap text and Premier for the big-context and teacher jobs, and remember the inference profile prefix so your first call does not 400. Next I will get into how Nova and the rest of Bedrock compare on pricing across on-demand, provisioned throughput and batch, because the per-token numbers here are only half the cost story. Try one thing this week: run the same ten prompts through Lite and Pro, diff the answers, and see whether Pro is actually buying you anything.

References

• What is Amazon Nova, model categories and specifications (AWS docs)
• What is new in Amazon Nova 2 (AWS docs)
• Amazon Nova pricing (AWS)
• Amazon Nova foundation models overview (AWS)

What Nova 2 changes

Nova 2 is the current generation, and it is where new work should go. The public models are Nova 2 Lite, Nova 2 Sonic and Nova Multimodal Embeddings. Nova 2 Lite adds extended thinking, meaning the model can spend extra tokens reasoning before it answers, with a three-level intensity control so you decide how much thinking to pay for. It also ships built-in tools: web grounding, which pulls current facts from the web to cut hallucinations, and a code interpreter that runs Python for calculations. Both remove integrations you used to build yourself. Nova 2 Sonic upgrades the speech model to seven languages. Nova Multimodal Embeddings is genuinely new: one model that turns text, documents, images, video and audio into vectors in a shared space, which makes cross-modal search and agentic retrieval much simpler.

A caveat worth stating plainly. As of this writing Nova 2 Pro is in preview with access gated behind Nova Forge, so if you need a top-tier Nova 2 reasoning model today you may still be on first-generation Premier. The original Micro, Lite, Pro, Premier, Canvas, Reel and Sonic have moved to Legacy status, which means they still run but AWS is steering new builds toward Nova 2. Legacy is not deprecated. Your existing Lite calls keep working. But when you plan a new service, check whether a Nova 2 equivalent exists before you commit to a first-generation ID.

Creative and speech models

Three models cover pixels and sound. Nova Canvas (amazon.nova-canvas-v1:0) generates and edits images from text or image prompts, up to about 4.19 million pixels, which covers 2048 by 2048 and similar sizes. Nova Reel (amazon.nova-reel-v1:1) makes short video at 1280 by 720 and 24 frames per second, and it runs through the asynchronous invoke API because a clip takes real time to render, so you submit a job and poll for it rather than blocking on a response. Nova Sonic (amazon.nova-sonic-v1:0) is a speech-to-speech model: you stream audio in and get audio plus a text transcript back, over a bidirectional streaming API, with a connection timeout around 8 minutes that you renew by replaying the conversation history. The first-generation Sonic supports five languages; Nova 2 Sonic extends that to seven.

How I pick a Nova model

Modality comes first and it is a hard filter. If you need to read images or video, Micro is out, full stop. Context is the second filter: if a single call must hold more than roughly 300K tokens, only Premier fits. Once those two constraints leave you with candidates, price and quality decide. My rule of thumb is to prototype on Lite, because it handles the large majority of tasks at a price close to Micro, then A/B the specific prompts where Lite falls short against Pro and keep Pro only there. Reaching for Pro or Premier everywhere is the most common way teams burn money on Nova.

Cost math nobody runs first

The gap between Lite and Pro is not a rounding error, and you should size it before you commit. Take a support assistant that processes 2 million input tokens and generates 0.5 million output tokens per day. On Lite that is 2 times $0.06 plus 0.5 times $0.24, which is $0.12 plus $0.12, so $0.24 a day. On Pro the same traffic is 2 times $0.80 plus 0.5 times $3.20, which is $1.60 plus $1.60, so $3.20 a day. Micro, if the task were text only, would be about $0.14 a day. Over a 30 day month that is roughly $4.20 for Micro, $7.20 for Lite and $96 for Pro. Pro is thirteen times the daily cost of Lite for this shape of traffic. If Lite answers the tickets as well, that difference is pure waste.

The inference profile trap, and the fix

Back to that 400. For on-demand Nova in most regions, Bedrock does not want the base model ID. It wants a cross-region inference profile ID, which is the base ID with a geography prefix, so amazon.nova-lite-v1:0 becomes us.amazon.nova-lite-v1:0 in the US. The inference profile lets Bedrock route your call across several regions to absorb traffic bursts, and for these models it is the on-demand path. Call the base ID and you get a ValidationException telling you to use an inference profile. Here is the working call with the boto3 SDK and the Converse API, which is the modern, model-agnostic way to talk to Bedrock.

import boto3

client = boto3.client('bedrock-runtime', region_name='us-east-1')

resp = client.converse(
    modelId='us.amazon.nova-lite-v1:0',   # inference profile, not the base id
    messages=[{'role': 'user',
               'content': [{'text': 'Summarize our refund policy in two sentences.'}]}],
    inferenceConfig={'maxTokens': 400, 'temperature': 0.2},
)

print(resp['output']['message']['content'][0]['text'])
print(resp['usage'])   # inputTokens, outputTokens, totalTokens

Expected output: a two sentence summary, then a usage dict such as inputTokens 41, outputTokens 38, totalTokens 79. Failure mode: set modelId to amazon.nova-lite-v1:0 and the same code raises a ValidationException that says to use an inference profile. Swap in the us. prefixed ID and it clears. The usage block is also where your cost tracking should start, because those token counts feed straight into the price math above.

Regions and availability, the part that bites late

Nova is not in every region, and this is where a design that looked fine on a whiteboard fails in the account. As a base model, Premier runs only in US East (N. Virginia) and AWS GovCloud (US-West). Pro, Lite and Micro are broader, adding Asia Pacific (Sydney), Asia Pacific (Tokyo) and Europe (London) on top of N. Virginia and GovCloud. Cross-region inference profiles stretch that reach much further, to Ohio, Oregon, Ireland, Frankfurt, Paris, Stockholm, Singapore, Seoul and Mumbai, by routing a call to whichever region has spare capacity. The creative and speech models sit in their own footprints: Canvas and Reel in N. Virginia, Ireland and Tokyo, and Sonic in N. Virginia, Stockholm and Tokyo.

That routing is exactly what bites a regulated workload. The same inference profile that saves you during a traffic burst can send a request to a region your compliance team never signed off on, so if data residency is a hard requirement, treat cross-region inference as something to constrain, not a free win. Two habits save real pain here. Check model availability per region before you pick a deployment region, because a model that shows up in the console can still refuse in the region you chose. And confirm the residency behavior of any profile you enable, since the whole point of a profile is to move your traffic around. A model ID that works in a demo account in N. Virginia is not a promise it works in eu-west-2 next quarter.

Where Nova fits against Claude and Llama

Nova does not win every task, and pretending otherwise wastes your time. On the hardest reasoning and coding, the top Anthropic Claude models in Bedrock still lead, and for some open-weight and fine-tuning work a Llama model is the better base. What Nova wins is the price-to-quality ratio on the ordinary majority of work: classification, extraction, summarization, routing, straightforward chat, and multimodal reads where you feed a document image and ask a question. For those, Lite is hard to beat on cost, and Micro is cheaper still when the input is pure text. Premier earns its keep in two narrow spots: when you truly need a 1M token context in one call, and when you are distilling a smaller custom model and want the strongest teacher.

My verdict: make Nova Lite the house default and justify every deviation. Route the high-volume text-only paths to Micro, escalate only the measured-hard prompts to Pro, reserve Premier for the two cases above, and move new builds onto Nova 2 where an equivalent exists. If you want the vendor-neutral view of how these choices generalize, the GenAI Series covers the model-selection logic without the AWS labels. For how Nova sits in the wider Bedrock catalog next to the third-party models, go back to Part 3.

Which Nova to reach for first

Nova is the value tier of Bedrock, and treating it that way pays off. Learn the four understanding models by modality and context, default to Lite, keep Micro for cheap text and Premier for the big-context and teacher jobs, and remember the inference profile prefix so your first call does not 400. Next I will get into how Nova and the rest of Bedrock compare on pricing across on-demand, provisioned throughput and batch, because the per-token numbers here are only half the cost story. Try one thing this week: run the same ten prompts through Lite and Pro, diff the answers, and see whether Pro is actually buying you anything.

References

• What is Amazon Nova, model categories and specifications (AWS docs)
• What is new in Amazon Nova 2 (AWS docs)
• Amazon Nova pricing (AWS)
• Amazon Nova foundation models overview (AWS)

Four understanding models

These are the ones you will use most, so learn them cold. Micro is text only, the cheapest and fastest, with a 128K token context window (the amount of text it can read in one call). Lite adds image and video input and jumps to 300K context, at a price only a little above Micro. Pro shares the 300K window and the same modalities but is tuned for accuracy on harder tasks, and it costs more than ten times Lite on output tokens. Premier is the heavyweight: a 1M token context and the best reasoning, and it is the designated teacher when you distill a smaller custom model. All four cap output at 10K tokens per call, which surprises people who expect long generations.

What Nova 2 changes

Nova 2 is the current generation, and it is where new work should go. The public models are Nova 2 Lite, Nova 2 Sonic and Nova Multimodal Embeddings. Nova 2 Lite adds extended thinking, meaning the model can spend extra tokens reasoning before it answers, with a three-level intensity control so you decide how much thinking to pay for. It also ships built-in tools: web grounding, which pulls current facts from the web to cut hallucinations, and a code interpreter that runs Python for calculations. Both remove integrations you used to build yourself. Nova 2 Sonic upgrades the speech model to seven languages. Nova Multimodal Embeddings is genuinely new: one model that turns text, documents, images, video and audio into vectors in a shared space, which makes cross-modal search and agentic retrieval much simpler.

A caveat worth stating plainly. As of this writing Nova 2 Pro is in preview with access gated behind Nova Forge, so if you need a top-tier Nova 2 reasoning model today you may still be on first-generation Premier. The original Micro, Lite, Pro, Premier, Canvas, Reel and Sonic have moved to Legacy status, which means they still run but AWS is steering new builds toward Nova 2. Legacy is not deprecated. Your existing Lite calls keep working. But when you plan a new service, check whether a Nova 2 equivalent exists before you commit to a first-generation ID.

Creative and speech models

Three models cover pixels and sound. Nova Canvas (amazon.nova-canvas-v1:0) generates and edits images from text or image prompts, up to about 4.19 million pixels, which covers 2048 by 2048 and similar sizes. Nova Reel (amazon.nova-reel-v1:1) makes short video at 1280 by 720 and 24 frames per second, and it runs through the asynchronous invoke API because a clip takes real time to render, so you submit a job and poll for it rather than blocking on a response. Nova Sonic (amazon.nova-sonic-v1:0) is a speech-to-speech model: you stream audio in and get audio plus a text transcript back, over a bidirectional streaming API, with a connection timeout around 8 minutes that you renew by replaying the conversation history. The first-generation Sonic supports five languages; Nova 2 Sonic extends that to seven.

How I pick a Nova model

Modality comes first and it is a hard filter. If you need to read images or video, Micro is out, full stop. Context is the second filter: if a single call must hold more than roughly 300K tokens, only Premier fits. Once those two constraints leave you with candidates, price and quality decide. My rule of thumb is to prototype on Lite, because it handles the large majority of tasks at a price close to Micro, then A/B the specific prompts where Lite falls short against Pro and keep Pro only there. Reaching for Pro or Premier everywhere is the most common way teams burn money on Nova.

Cost math nobody runs first

The gap between Lite and Pro is not a rounding error, and you should size it before you commit. Take a support assistant that processes 2 million input tokens and generates 0.5 million output tokens per day. On Lite that is 2 times $0.06 plus 0.5 times $0.24, which is $0.12 plus $0.12, so $0.24 a day. On Pro the same traffic is 2 times $0.80 plus 0.5 times $3.20, which is $1.60 plus $1.60, so $3.20 a day. Micro, if the task were text only, would be about $0.14 a day. Over a 30 day month that is roughly $4.20 for Micro, $7.20 for Lite and $96 for Pro. Pro is thirteen times the daily cost of Lite for this shape of traffic. If Lite answers the tickets as well, that difference is pure waste.

The inference profile trap, and the fix

Back to that 400. For on-demand Nova in most regions, Bedrock does not want the base model ID. It wants a cross-region inference profile ID, which is the base ID with a geography prefix, so amazon.nova-lite-v1:0 becomes us.amazon.nova-lite-v1:0 in the US. The inference profile lets Bedrock route your call across several regions to absorb traffic bursts, and for these models it is the on-demand path. Call the base ID and you get a ValidationException telling you to use an inference profile. Here is the working call with the boto3 SDK and the Converse API, which is the modern, model-agnostic way to talk to Bedrock.

import boto3

client = boto3.client('bedrock-runtime', region_name='us-east-1')

resp = client.converse(
    modelId='us.amazon.nova-lite-v1:0',   # inference profile, not the base id
    messages=[{'role': 'user',
               'content': [{'text': 'Summarize our refund policy in two sentences.'}]}],
    inferenceConfig={'maxTokens': 400, 'temperature': 0.2},
)

print(resp['output']['message']['content'][0]['text'])
print(resp['usage'])   # inputTokens, outputTokens, totalTokens

Expected output: a two sentence summary, then a usage dict such as inputTokens 41, outputTokens 38, totalTokens 79. Failure mode: set modelId to amazon.nova-lite-v1:0 and the same code raises a ValidationException that says to use an inference profile. Swap in the us. prefixed ID and it clears. The usage block is also where your cost tracking should start, because those token counts feed straight into the price math above.

Regions and availability, the part that bites late

Nova is not in every region, and this is where a design that looked fine on a whiteboard fails in the account. As a base model, Premier runs only in US East (N. Virginia) and AWS GovCloud (US-West). Pro, Lite and Micro are broader, adding Asia Pacific (Sydney), Asia Pacific (Tokyo) and Europe (London) on top of N. Virginia and GovCloud. Cross-region inference profiles stretch that reach much further, to Ohio, Oregon, Ireland, Frankfurt, Paris, Stockholm, Singapore, Seoul and Mumbai, by routing a call to whichever region has spare capacity. The creative and speech models sit in their own footprints: Canvas and Reel in N. Virginia, Ireland and Tokyo, and Sonic in N. Virginia, Stockholm and Tokyo.

That routing is exactly what bites a regulated workload. The same inference profile that saves you during a traffic burst can send a request to a region your compliance team never signed off on, so if data residency is a hard requirement, treat cross-region inference as something to constrain, not a free win. Two habits save real pain here. Check model availability per region before you pick a deployment region, because a model that shows up in the console can still refuse in the region you chose. And confirm the residency behavior of any profile you enable, since the whole point of a profile is to move your traffic around. A model ID that works in a demo account in N. Virginia is not a promise it works in eu-west-2 next quarter.

Where Nova fits against Claude and Llama

Nova does not win every task, and pretending otherwise wastes your time. On the hardest reasoning and coding, the top Anthropic Claude models in Bedrock still lead, and for some open-weight and fine-tuning work a Llama model is the better base. What Nova wins is the price-to-quality ratio on the ordinary majority of work: classification, extraction, summarization, routing, straightforward chat, and multimodal reads where you feed a document image and ask a question. For those, Lite is hard to beat on cost, and Micro is cheaper still when the input is pure text. Premier earns its keep in two narrow spots: when you truly need a 1M token context in one call, and when you are distilling a smaller custom model and want the strongest teacher.

My verdict: make Nova Lite the house default and justify every deviation. Route the high-volume text-only paths to Micro, escalate only the measured-hard prompts to Pro, reserve Premier for the two cases above, and move new builds onto Nova 2 where an equivalent exists. If you want the vendor-neutral view of how these choices generalize, the GenAI Series covers the model-selection logic without the AWS labels. For how Nova sits in the wider Bedrock catalog next to the third-party models, go back to Part 3.

Which Nova to reach for first

Nova is the value tier of Bedrock, and treating it that way pays off. Learn the four understanding models by modality and context, default to Lite, keep Micro for cheap text and Premier for the big-context and teacher jobs, and remember the inference profile prefix so your first call does not 400. Next I will get into how Nova and the rest of Bedrock compare on pricing across on-demand, provisioned throughput and batch, because the per-token numbers here are only half the cost story. Try one thing this week: run the same ten prompts through Lite and Pro, diff the answers, and see whether Pro is actually buying you anything.

References

• What is Amazon Nova, model categories and specifications (AWS docs)
• What is new in Amazon Nova 2 (AWS docs)
• Amazon Nova pricing (AWS)
• Amazon Nova foundation models overview (AWS)

My first Nova call in production returned a 400. The request was clean, the region was right, the model ID looked correct: amazon.nova-lite-v1:0. Bedrock rejected it and told me to use an inference profile. That one detail trips almost everyone the first time, and I will come back to it with a fix. But before the wiring, you need to know which Nova model to call at all, because AWS now ships nine of them plus a second generation, and the names do not tell you much.

What Amazon Nova actually is

Amazon Nova is the family of foundation models that AWS trained in house, as opposed to the Anthropic, Meta, Mistral and other third-party models that also live in Bedrock. A foundation model here means a large pretrained model you call through an API rather than train yourself. The pitch for Nova is price and speed: for a lot of everyday work it costs a fraction of the frontier third-party models and answers faster, while still handling images and video. That trade is the whole reason Nova exists, and it is why I default to it for high-volume paths and keep the pricier models for the hard 10 percent.

The family splits into three jobs. Understanding models read text, and in most cases images and video, then write text back. Creative models generate images or video. The speech model listens and talks. On top of the original lineup sits Nova 2, a newer generation with better reasoning and tool use, plus two products that are not chat models at all: Nova Act, which drives browser and computer tasks for agents, and Nova Forge, a service for building your own custom frontier model on top of Nova. Keep those last two in a separate mental bucket; this part is about the models you invoke for text, images, video and speech.

Four understanding models

These are the ones you will use most, so learn them cold. Micro is text only, the cheapest and fastest, with a 128K token context window (the amount of text it can read in one call). Lite adds image and video input and jumps to 300K context, at a price only a little above Micro. Pro shares the 300K window and the same modalities but is tuned for accuracy on harder tasks, and it costs more than ten times Lite on output tokens. Premier is the heavyweight: a 1M token context and the best reasoning, and it is the designated teacher when you distill a smaller custom model. All four cap output at 10K tokens per call, which surprises people who expect long generations.

What Nova 2 changes

Nova 2 is the current generation, and it is where new work should go. The public models are Nova 2 Lite, Nova 2 Sonic and Nova Multimodal Embeddings. Nova 2 Lite adds extended thinking, meaning the model can spend extra tokens reasoning before it answers, with a three-level intensity control so you decide how much thinking to pay for. It also ships built-in tools: web grounding, which pulls current facts from the web to cut hallucinations, and a code interpreter that runs Python for calculations. Both remove integrations you used to build yourself. Nova 2 Sonic upgrades the speech model to seven languages. Nova Multimodal Embeddings is genuinely new: one model that turns text, documents, images, video and audio into vectors in a shared space, which makes cross-modal search and agentic retrieval much simpler.

A caveat worth stating plainly. As of this writing Nova 2 Pro is in preview with access gated behind Nova Forge, so if you need a top-tier Nova 2 reasoning model today you may still be on first-generation Premier. The original Micro, Lite, Pro, Premier, Canvas, Reel and Sonic have moved to Legacy status, which means they still run but AWS is steering new builds toward Nova 2. Legacy is not deprecated. Your existing Lite calls keep working. But when you plan a new service, check whether a Nova 2 equivalent exists before you commit to a first-generation ID.

Creative and speech models

Three models cover pixels and sound. Nova Canvas (amazon.nova-canvas-v1:0) generates and edits images from text or image prompts, up to about 4.19 million pixels, which covers 2048 by 2048 and similar sizes. Nova Reel (amazon.nova-reel-v1:1) makes short video at 1280 by 720 and 24 frames per second, and it runs through the asynchronous invoke API because a clip takes real time to render, so you submit a job and poll for it rather than blocking on a response. Nova Sonic (amazon.nova-sonic-v1:0) is a speech-to-speech model: you stream audio in and get audio plus a text transcript back, over a bidirectional streaming API, with a connection timeout around 8 minutes that you renew by replaying the conversation history. The first-generation Sonic supports five languages; Nova 2 Sonic extends that to seven.

How I pick a Nova model

Modality comes first and it is a hard filter. If you need to read images or video, Micro is out, full stop. Context is the second filter: if a single call must hold more than roughly 300K tokens, only Premier fits. Once those two constraints leave you with candidates, price and quality decide. My rule of thumb is to prototype on Lite, because it handles the large majority of tasks at a price close to Micro, then A/B the specific prompts where Lite falls short against Pro and keep Pro only there. Reaching for Pro or Premier everywhere is the most common way teams burn money on Nova.

Cost math nobody runs first

The gap between Lite and Pro is not a rounding error, and you should size it before you commit. Take a support assistant that processes 2 million input tokens and generates 0.5 million output tokens per day. On Lite that is 2 times $0.06 plus 0.5 times $0.24, which is $0.12 plus $0.12, so $0.24 a day. On Pro the same traffic is 2 times $0.80 plus 0.5 times $3.20, which is $1.60 plus $1.60, so $3.20 a day. Micro, if the task were text only, would be about $0.14 a day. Over a 30 day month that is roughly $4.20 for Micro, $7.20 for Lite and $96 for Pro. Pro is thirteen times the daily cost of Lite for this shape of traffic. If Lite answers the tickets as well, that difference is pure waste.

The inference profile trap, and the fix

Back to that 400. For on-demand Nova in most regions, Bedrock does not want the base model ID. It wants a cross-region inference profile ID, which is the base ID with a geography prefix, so amazon.nova-lite-v1:0 becomes us.amazon.nova-lite-v1:0 in the US. The inference profile lets Bedrock route your call across several regions to absorb traffic bursts, and for these models it is the on-demand path. Call the base ID and you get a ValidationException telling you to use an inference profile. Here is the working call with the boto3 SDK and the Converse API, which is the modern, model-agnostic way to talk to Bedrock.

import boto3

client = boto3.client('bedrock-runtime', region_name='us-east-1')

resp = client.converse(
    modelId='us.amazon.nova-lite-v1:0',   # inference profile, not the base id
    messages=[{'role': 'user',
               'content': [{'text': 'Summarize our refund policy in two sentences.'}]}],
    inferenceConfig={'maxTokens': 400, 'temperature': 0.2},
)

print(resp['output']['message']['content'][0]['text'])
print(resp['usage'])   # inputTokens, outputTokens, totalTokens

Expected output: a two sentence summary, then a usage dict such as inputTokens 41, outputTokens 38, totalTokens 79. Failure mode: set modelId to amazon.nova-lite-v1:0 and the same code raises a ValidationException that says to use an inference profile. Swap in the us. prefixed ID and it clears. The usage block is also where your cost tracking should start, because those token counts feed straight into the price math above.

Regions and availability, the part that bites late

Nova is not in every region, and this is where a design that looked fine on a whiteboard fails in the account. As a base model, Premier runs only in US East (N. Virginia) and AWS GovCloud (US-West). Pro, Lite and Micro are broader, adding Asia Pacific (Sydney), Asia Pacific (Tokyo) and Europe (London) on top of N. Virginia and GovCloud. Cross-region inference profiles stretch that reach much further, to Ohio, Oregon, Ireland, Frankfurt, Paris, Stockholm, Singapore, Seoul and Mumbai, by routing a call to whichever region has spare capacity. The creative and speech models sit in their own footprints: Canvas and Reel in N. Virginia, Ireland and Tokyo, and Sonic in N. Virginia, Stockholm and Tokyo.

That routing is exactly what bites a regulated workload. The same inference profile that saves you during a traffic burst can send a request to a region your compliance team never signed off on, so if data residency is a hard requirement, treat cross-region inference as something to constrain, not a free win. Two habits save real pain here. Check model availability per region before you pick a deployment region, because a model that shows up in the console can still refuse in the region you chose. And confirm the residency behavior of any profile you enable, since the whole point of a profile is to move your traffic around. A model ID that works in a demo account in N. Virginia is not a promise it works in eu-west-2 next quarter.

Where Nova fits against Claude and Llama

Nova does not win every task, and pretending otherwise wastes your time. On the hardest reasoning and coding, the top Anthropic Claude models in Bedrock still lead, and for some open-weight and fine-tuning work a Llama model is the better base. What Nova wins is the price-to-quality ratio on the ordinary majority of work: classification, extraction, summarization, routing, straightforward chat, and multimodal reads where you feed a document image and ask a question. For those, Lite is hard to beat on cost, and Micro is cheaper still when the input is pure text. Premier earns its keep in two narrow spots: when you truly need a 1M token context in one call, and when you are distilling a smaller custom model and want the strongest teacher.

My verdict: make Nova Lite the house default and justify every deviation. Route the high-volume text-only paths to Micro, escalate only the measured-hard prompts to Pro, reserve Premier for the two cases above, and move new builds onto Nova 2 where an equivalent exists. If you want the vendor-neutral view of how these choices generalize, the GenAI Series covers the model-selection logic without the AWS labels. For how Nova sits in the wider Bedrock catalog next to the third-party models, go back to Part 3.

Which Nova to reach for first

Nova is the value tier of Bedrock, and treating it that way pays off. Learn the four understanding models by modality and context, default to Lite, keep Micro for cheap text and Premier for the big-context and teacher jobs, and remember the inference profile prefix so your first call does not 400. Next I will get into how Nova and the rest of Bedrock compare on pricing across on-demand, provisioned throughput and batch, because the per-token numbers here are only half the cost story. Try one thing this week: run the same ten prompts through Lite and Pro, diff the answers, and see whether Pro is actually buying you anything.

References

• What is Amazon Nova, model categories and specifications (AWS docs)
• What is new in Amazon Nova 2 (AWS docs)
• Amazon Nova pricing (AWS)
• Amazon Nova foundation models overview (AWS)

About The Author

Dr. Pranay Jha

Dr. Pranay Jha is a Cloud and AI Consultant with 18+ years of experience in hybrid cloud, virtualization, and enterprise infrastructure transformation. He specializes in VMware technologies, multi-cloud strategy, and Generative AI solutions. He holds a PhD in Computer Applications with research focused on Cloud and AI, has published multiple research papers, and has been a VMware vExpert since 2016 and a VMUG Community Leader.

See author's posts