Skip to content

Amazon Nova and AWS Bedrock models explained

  • How Amazon Bedrock (the platform) differs from Amazon Nova and Titan (Amazon’s own models).
  • How the Nova family is organized across two generations, and which first-generation models have been retired.
  • How Bedrock model IDs, inference profiles and lifecycle states work, and how to read the current state from the API.
  • How to choose between Nova tiers and the third-party models Bedrock also hosts.
  • What fails in production: end-of-life dates, regional gaps, quotas and legacy-access rules.

Amazon Bedrock is AWS’s managed service for calling foundation models through one API, with AWS identity, networking and billing. It hosts models from several providers, including Anthropic, Meta and Mistral, alongside Amazon’s own.

Amazon Nova is Amazon’s first-party model family. The first generation began in December 2024 (Micro, Lite and Pro, plus Canvas for images and Reel for video), with Premier and the speech model Sonic added later. The Nova 2 generation was announced in December 2025 (Nova 2 Lite, Nova 2 Pro, Nova 2 Omni, Nova 2 Sonic and Nova Multimodal Embeddings). Amazon Titan is the older family that preceded Nova.

During 2026 AWS moved much of the first Nova generation to legacy or end-of-life status. If you are reading a 2025 tutorial, several of the models it names can no longer be called. The Bedrock model cards and the ListFoundationModels API are the source of truth, and this page shows how to use them.

A note on sources: this page was verified in November 2026 from AWS documentation and announcement pages found through search, and from the boto3 SDK’s service definitions. AWS’s pages did not always agree with one another on dates, so where they conflict the page says so and points you to the live API.

Name What it is
Amazon Bedrock The managed platform: one API (Converse, InvokeModel) across providers, with IAM, VPC endpoints, guardrails, knowledge bases, agents and evaluation
Amazon Nova Amazon’s own model family for text, multimodal understanding, speech, embeddings and (in the first generation) image and video generation
Amazon Titan Amazon’s earlier text, embedding and image models
Nova Forge A service for building custom models from Nova checkpoints with your own data
Nova Act A service and model for agents that automate browser user-interface workflows
Third-party models on Bedrock Models from other providers behind the same API, each with its own terms, prices and lifecycle

Choosing “AWS for AI” is therefore two decisions: whether to use Bedrock as the platform, and which model to call on it. A common pattern is to run a third-party frontier model on Bedrock for the hardest tasks and use Nova for cost-sensitive or speech workloads.

Model Role Status Notes
Nova 2 Lite Fast, cost-effective reasoning model for everyday workloads Active Generally available since December 2025; 1M-token context; model card lists lifecycle Active with end of life “no sooner than” December 2, 2026 (a floor, explained below)
Nova 2 Pro Amazon’s most capable model for complex, multistep tasks Preview at launch Early access for Nova Forge customers; check the model card for current availability
Nova 2 Omni Multimodal reasoning and image generation in one model Preview at launch Check the model card for current availability
Nova 2 Sonic Speech-to-speech model for real-time voice conversations Active Generally available since December 2025; model ID amazon.nova-2-sonic-v1:0; 1M-token context; end of life no sooner than December 2, 2026
Nova 2.5 Sonic Successor to Nova 2 Sonic, with the same API surface, voices and languages Active Launched October 5, 2026; model card: 256K-token context, 64K output, bidirectional streaming only, end of life no sooner than October 5, 2027
Nova Multimodal Embeddings Embeddings for text, image, document, video and audio Generally available at launch For retrieval across modalities; check the model card for current status

Nova 2 introduced extended reasoning, a 1M-token context window and built-in tools such as a code interpreter and web grounding. Nova 2.5 Sonic is called only through the bidirectional streaming API (InvokeModelWithBidirectionalStream), not through Converse or InvokeModel, so voice code is a separate integration from text code.

What “end of life no sooner than” means. Bedrock model cards can show an “EOL no sooner than” date. It is only a floor on the end-of-life date: AWS commits not to retire the model before it. It is not a retirement date, and it does not stop AWS from moving the model to Legacy earlier. Under the lifecycle policy for models launched before September 7, 2026, a model spends at least six months in Legacy, and its end-of-life date can be no earlier than the floor. Claude Sonnet 4 on Bedrock shows how this plays out: its card listed end of life no sooner than May 23, 2026, it became Legacy on April 14, 2026 (before the floor), and it reached end of life on October 14, 2026. So the December 2, 2026 floor on Nova 2 Lite and Nova 2 Sonic says only that neither can stop working before that date. Either could turn Legacy at any time, and from that moment new customers cannot adopt it and idle accounts can lose access. Read modelLifecycle on a schedule now, not from the floor date, and keep a qualified replacement ready.

Nova 2 Lite is called through inference profile IDs, which AWS documents as us.amazon.nova-2-lite-v1:0 for US routing and global.amazon.nova-2-lite-v1:0 for global routing.

First-generation Nova status, as shown on AWS pages in November 2026
Model Role Model ID Status (Nov 2026)
Nova Micro Text only, lowest latency and cost; 128K context amazon.nova-micro-v1:0 Not confirmed: AWS’s model card did not show a current status in November 2026. Read modelLifecycle
Nova Lite Low-cost multimodal (text, image, video input); 300K context amazon.nova-lite-v1:0 Legacy, end of life September 30, 2026, on AWS’s Regional availability page; the model card still showed Active. Read modelLifecycle
Nova Pro Balanced multimodal; 300K context amazon.nova-pro-v1:0 Legacy in us-east-1, eu-west-1 and ap-northeast-1, end of life September 30, 2026, on AWS’s Regional availability page; the model card still showed Active. Read modelLifecycle
Nova Premier Most capable first-generation model; teacher for distillation amazon.nova-premier-v1:0 Legacy; AWS pages showed end-of-life dates of September 14 and October 31, 2026. Read modelLifecycle
Nova Canvas Image generation See model card End of life September 30, 2026; no longer available
Nova Reel Video generation See model card End of life September 30, 2026
Nova Sonic (v1) Speech-to-speech amazon.nova-sonic-v1:0 End of life September 14, 2026 (Legacy from March 13, 2026)

Two consequences follow. Amazon’s dedicated image and video generation models from this generation are gone from Bedrock, so creative workloads need Nova 2 Omni or a third-party model. And any system still configured with a first-generation ID needs an owner and a migration date now. AWS publishes a guide for moving from Nova 1 to Nova 2.

Titan text models (Text Lite, Express, Premier) and Titan embeddings came before Nova. Titan Text Embeddings V2 appears in many older retrieval tutorials. This page could not confirm the November 2026 lifecycle state of each Titan model from a current AWS page, so do not start new work on Titan without checking its model card, and plan to evaluate Nova Multimodal Embeddings or another embedding model for new retrieval systems. Remember that changing embedding models means re-embedding your whole corpus.

Model IDs and inference profiles. A base model ID looks like amazon.nova-micro-v1:0. Many newer models are invoked through a cross-Region inference profile instead, whose ID adds a geography prefix such as us. or global.. The profile routes each request to a Region with capacity inside that geography. This matters for data residency: a global. profile can process a request outside your home Region.

Lifecycle states. Bedrock models are Active, then Legacy, then End-of-Life. AWS’s documentation states:

  • After the end-of-life date the model is removed from all Regions and requests fail. Migration is not automatic.
  • New customers cannot start using a Legacy model, and existing customers may lose access to a Legacy model after 15 days of inactivity.
  • Lifecycle state can differ by Region.
  • Bedrock’s dates are its own. For a third-party model they may differ from the dates the provider publishes, and only the Bedrock model card applies to Bedrock usage.
  • A revised lifecycle policy applies to models launched on Bedrock on or after September 7, 2026.

Reading the state from the API. The fragment below requires AWS credentials and Bedrock permissions and was not executed. The operation, filter and field names were checked against the boto3 1.44.3 service definitions, where modelLifecycle.status is ACTIVE or LEGACY and the same structure carries legacyTime and endOfLifeTime.

# Fragment: requires AWS credentials with bedrock:ListFoundationModels; not executed
# pip install boto3
import boto3
bedrock = boto3.client("bedrock", region_name="us-east-1")
for model in bedrock.list_foundation_models(byProvider="Amazon")["modelSummaries"]:
lifecycle = model["modelLifecycle"]
print(
model["modelId"],
lifecycle["status"],
lifecycle.get("endOfLifeTime", "no end-of-life date"),
)

Run a check like this in every Region you deploy to, on a schedule, and alert when a model you use turns LEGACY.

Worked example: calling Nova through the Converse API

Section titled “Worked example: calling Nova through the Converse API”

An invented shop, OrdersPortal, already runs on AWS and wants ticket summaries without sending data to another vendor. The Converse API gives one request shape across models, so the model can be swapped by changing an ID.

This fragment requires AWS credentials and model access and was not executed; parameter and response field names were checked against the boto3 1.44.3 service definitions.

# Fragment: requires AWS credentials and access to the model; not executed
import os
import boto3
runtime = boto3.client("bedrock-runtime", region_name="us-east-1")
response = runtime.converse(
modelId=os.environ.get("SUMMARY_MODEL_ID", "us.amazon.nova-2-lite-v1:0"),
system=[{"text": "Summarize the support ticket in one sentence."}],
messages=[{
"role": "user",
"content": [{"text": "Order A1001 arrived with the box crushed. I want a replacement, not a refund."}],
}],
inferenceConfig={"maxTokens": 200},
)
print(response["output"]["message"]["content"][0]["text"])
print(response["usage"]["inputTokens"], response["usage"]["outputTokens"], response["stopReason"])

The design choices that matter are outside the call: the model ID comes from configuration, the IAM policy allows only the models and Regions in use, and the same code can be pointed at a different Bedrock model for comparison.

Need Start with Why
Everyday assistants, extraction, summarization on AWS Nova 2 Lite Generally available, 1M-token context, positioned for price performance
The hardest reasoning on AWS Nova 2 Pro if available to you, or a third-party frontier model on Bedrock Compare on your own tasks; preview models carry limits
Real-time voice agents Nova 2.5 Sonic Speech in and speech out in one model, instead of chaining transcription, a text model and text-to-speech; the newest Sonic, with the latest end-of-life floor
Retrieval across text, images, audio and video Nova Multimodal Embeddings One embedding space for mixed content
Image or video generation Nova 2 Omni (check status) or a third-party model The first-generation Canvas and Reel models reached end of life
A model specialized on proprietary data Nova Forge, or Bedrock fine-tuning and distillation Only when prompting and retrieval are measurably insufficient
Automating browser workflows Nova Act Purpose-built for user-interface automation

Prices change and differ by Region, tier and routing, so this page lists none. Use the Bedrock pricing page and estimate with your own token counts. Bedrock offers on-demand, batch and provisioned throughput modes, and the right one depends on volume and latency needs.

Choice Gain Cost
Bedrock as platform One API and bill across providers; IAM, VPC endpoints, CloudTrail logging, data stays within AWS terms New model features can arrive later than on a provider’s own API; quotas and Region availability vary
Nova models Price performance, native AWS integration, speech-to-speech Fewer independent evaluations than the most widely used model families; several capable models have been preview-only
Third-party models on Bedrock Frontier capability inside your AWS account Provider-specific terms; Bedrock lifecycle dates may differ from the provider’s
A provider’s own API Newest features first Another vendor, contract and security review
Self-hosted open models Full control You run the infrastructure. See Running LLMs locally

For comparison across providers, see LLM model selection, Anthropic models, Google models and OpenAI models.

  • End-of-life outages. Requests to a retired model fail and nothing migrates for you. Several first-generation Nova models reached that point in September 2026.
  • Losing legacy access through inactivity. A disaster-recovery or seasonal workload that calls a Legacy model rarely can lose access after 15 idle days.
  • Conflicting dates. In November 2026, different AWS pages showed different end-of-life dates for the same model, and some model cards did not show a current status. Trust the API response for your Region.
  • Reading the floor as a schedule. An “EOL no sooner than” date is the earliest possible retirement, not the retirement itself, and Legacy can start before it. Plan from the Legacy transition and the actual end-of-life date, which you learn from modelLifecycle.
  • Wrong ID form. Calling a base model ID where an inference profile is required, or the reverse, returns a validation or access error. Copy the exact ID from the model card.
  • Region assumptions. A model, a feature or a quota available in one Region may be missing in another. Verify per Region before committing an architecture.
  • Data residency with cross-Region profiles. A us. or global. profile processes requests in other Regions. Confirm that this fits your compliance rules, and restrict with IAM and service control policies where it does not.
  • Quotas and throttling. Default requests-per-minute and tokens-per-minute quotas are low for new accounts and new models. Request increases early, and implement retries with backoff.
  • Preview dependence. Preview models can change or be withdrawn and may have no service-level commitment.
  • Access not enabled. Model access and IAM permissions are separate from having an AWS account. An access-denied error on first call is usually a missing permission or agreement, not a code bug.
  • Embedding lock-in. Vectors from one embedding model cannot be compared with another’s. Retiring an embedding model means re-indexing everything.
  • Bedrock, when you are not on AWS and have no reason to add it: a provider’s own API is simpler.
  • Nova, when a task needs the strongest available reasoning and your evaluation shows a third-party model doing clearly better. Bedrock lets you switch.
  • A Legacy model, for any new system.
  • A preview model, for a workload with an uptime commitment.
  • Custom model training with Nova Forge, before you have shown that prompting, retrieval and a larger model do not solve the problem.

1. What is the difference between Amazon Bedrock and Amazon Nova? Bedrock is the managed platform that exposes many providers’ models through one API with AWS security and billing. Nova is Amazon’s own model family, one of several available on Bedrock. You can use Bedrock without Nova.

2. What is an inference profile, and why does it matter? It is an ID, usually prefixed with a geography such as us. or global., that routes a request to any Region with capacity inside that geography. It improves availability and throughput, but it means data may be processed outside your home Region, and some models can only be called this way.

3. Describe the Bedrock model lifecycle. Active, Legacy, then End-of-Life. Legacy models take no new customers and can be lost after inactivity. After end of life, requests fail in every Region and there is no automatic migration. State can differ by Region, so you read modelLifecycle from ListFoundationModels per Region instead of relying on a blog post.

4. Why prefer the Converse API over InvokeModel? Converse uses one message, tool and usage format across models, so switching models is a configuration change and tool-calling code is written once. InvokeModel passes the provider’s native request body, which is needed for model-specific features but couples code to one model.

5. How would you design a Bedrock application so that one model’s retirement is a configuration change, not a code change? Call models through Converse, so request and tool formats are shared. Map each workload to a model ID or inference profile in configuration, per Region. Keep a graded evaluation set per workload so a replacement can be qualified in days. Run a scheduled job that reads modelLifecycle in every Region you use and opens a ticket when a model turns Legacy. Grant IAM access to the replacement model in advance, and record in the configuration who owns each model choice.

6. Senior scenario: security requires that customer data never leaves the EU, and the team wants a Nova 2 model. What do you check? Whether the model is offered in an EU Region and through which ID: an in-Region model ID, an EU geography profile, or only a global profile. A global profile would breach the requirement. Then enforce it: IAM and service control policies that allow only EU Regions and the approved profile, VPC endpoints for Bedrock, logging through CloudTrail, and a review of whether features such as web grounding send data elsewhere. Confirm quotas in the EU Region, and document a fallback model that meets the same constraint.

  • Bedrock is the platform; Nova and Titan are Amazon’s models on it, next to third-party models.
  • Nova 2 (December 2025) is the current generation; much of the first generation was retired during 2026, including Canvas, Reel and the original Sonic.
  • Model IDs, inference profiles and lifecycle state are Region-specific: read them from the model card and the API.
  • Legacy models take no new customers and can be lost after 15 idle days; end of life means hard failures.
  • Use the Converse API and keep model IDs in configuration so models can be swapped.
  • Compare Nova with third-party Bedrock models on your own evaluation before committing.

Found and checked through search on November 13 and 15, 2026 (docs.aws.amazon.com is not directly reachable from the verification environment).