Amazon Nova Models: Every Model, Spec, Lifecycle and Release Date

Amazon Nova is Amazon's own family of foundation models, served through Amazon Bedrock; the current text model, Nova 2 Lite (API ID amazon.nova-2-lite-v1:0), launched on 2 December 2025 with a 1M token context window. This page lists every Nova model on Amazon's current model pages and Bedrock model cards, with API identifiers, context windows, maximum output, modalities, knowledge cutoffs, lifecycle dates and regional availability. It also covers pricing structure, data handling on Bedrock and the end-of-life schedule for first-generation Nova models. Every fact comes from Amazon's own Nova pages, Bedrock documentation or FAQs. Where Amazon publishes no value, the page says "Not stated by Amazon" rather than estimating. Facts are as of 15 November 2026.

Last checked: 2026-11-15.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.

Key takeaways

• Nova 2 Lite is the current text and vision model. API ID amazon.nova-2-lite-v1:0, 1M token context, 64K maximum output, text, image and video input, knowledge cutoff October 2025, according to its Bedrock model card.
• Nova 2.5 Sonic is the newest model. A speech-to-speech model launched on 5 October 2026 (API ID amazon.nova-2-5-sonic), with a 256K context window on its Bedrock model card.
• Nova models run on Amazon Bedrock. Amazon does not publish open weights; Nova 2 Lite can also be customised through Bedrock and Amazon SageMaker AI, and is the recommended base for Nova Forge.
• First-generation creative and speech models have reached end of life. Nova Canvas, Nova Reel and Nova Sonic (v1) are in the Bedrock "reached EOL" table with EOL dates in September 2026.
• Bedrock dates apply. AWS states that lifecycle dates are specific to Amazon Bedrock and that only the dates on each model card apply.

What is Amazon Nova?

Amazon Nova is the model family that Amazon builds itself, distinct from the third-party models (Anthropic, DeepSeek, Google, Meta, Mistral and others) that Bedrock also hosts. The Amazon Nova models page describes four product areas: models (Nova 2 Lite, Nova 2.5 Sonic and Nova Multimodal Embeddings), Nova Forge for building custom models, and Nova Act for browser agents. The Bedrock model card list adds the first-generation models still in service: Nova Pro, Nova Lite and Nova Micro, plus Nova 2 Sonic.

API identifiers follow the pattern amazon.<model>-v<major>:<minor>, for example amazon.nova-pro-v1:0. Cross-region inference profiles add a geography prefix (us., eu., jp., apac., ca.) or the global. prefix to the same ID. See LLM Model Naming and Versioning Explained for how these conventions compare across providers.

Current Nova models at a glance

As of 15 November 2026. Source: Amazon Nova models page and individual Bedrock model cards. Models are listed by generation, newest first.

The Nova Pro, Nova Lite and Nova Micro model cards show lifecycle status "Active" alongside an "EOL no sooner than" date that has passed; the cards also give the EOL date as both 5 December 2025 and 4 December 2025. A "no sooner than" date is a floor, not a scheduled shutdown, so check the live status with the modelLifecycle field returned by the Bedrock ListFoundationModels or GetFoundationModel API.

Key facts for the main Nova models

As of 15 November 2026. Identical fields for every model; "Not stated by Amazon" means not stated by Amazon on the pages checked (2026-11-15). API IDs, context, licence and status are in the table above.

Context conflict for Nova 2.5 Sonic: the Amazon Nova models page describes a context window of "up to 1M tokens", while the Bedrock model card gives 256K tokens. This page uses the model card, which is the more specific source, and records the other figure here.

Reasoning in Nova 2 Lite: the model card lists reasoning as supported and off by default. It is turned on with reasoningConfig, with effort levels low, medium and high. The Nova models page also lists developer controls for thinking effort, a built-in code interpreter, web grounding and remote MCP tool support. See Reasoning Models Explained.

Bedrock features listed on the model cards

Explicit prompt caching on Nova 2 Lite and Nova Pro: minimum 1K tokens per cache checkpoint, up to 4 checkpoints per request, a 5 minute TTL, checkpoints in the system and messages fields, and a maximum of 20K tokens cached. Neither the Nova 2 Lite nor the Nova Pro card lists structured outputs.

Nova model families and naming

Amazon groups Nova models as below, as of 15 November 2026. Release dates for each are in the next section.

• First generation (v1, December 2024): understanding models Nova Micro (text only), Nova Lite and Nova Pro (text, image and video in), plus the creative models Nova Canvas (image) and Nova Reel (video) and the speech model Nova Sonic. Canvas, Reel and Sonic v1 are now in the EOL table.
• Nova 2 (December 2025): Nova 2 Lite (reasoning, multimodal input) and Nova 2 Sonic (speech-to-speech), both launched on 2 December 2025.
• Nova 2.5 (October 2026): Nova 2.5 Sonic, launched on 5 October 2026, which the Nova models page describes as improving reasoning, instruction following and tool-calling accuracy, with asynchronous tool use and polyglot voices.
• Embeddings: Nova Multimodal Embeddings, which the Nova models page describes as a unified embedding model mapping text, documents, images, video and audio into a single space for agentic RAG and semantic search.
• Customisation and agents: Nova Forge (custom models built from Nova 2 Lite) and Nova Act (agents billed per agent hour).

When was each Nova model released?

Launch dates from Bedrock model cards and EOL dates from the Bedrock lifecycle page, as of 15 November 2026, in date order.

The dates above are the "Model launch date" fields on each Bedrock model card; the Nova models page gives no dates.

How much do Nova models cost?

As of 15 November 2026. Amazon publishes Nova token prices on the Amazon Bedrock pricing page, which the Nova model cards and the Nova pricing page both point to. Prices vary by region, inference type and service tier, so read the figure for your region and tier on that page before budgeting.

Service tiers: the Nova 2 Lite and Nova Pro cards accept "service_tier": "default" (Standard), "priority" (faster responses at a price premium) or "flex" (lower cost for workloads that are not time-sensitive). Reserved capacity is arranged at account level through the AWS account team. For how input, output, cached and batch pricing work in general, see LLM Pricing Explained: Tokens and Caching. To cap spend, see How to set spend limits on LLM API keys.

Where are Nova models available?

• Amazon Bedrock: all Nova models on this page are served on the bedrock-runtime endpoint (https://bedrock-runtime.{region}.amazonaws.com). The model cards recommend bedrock-runtime for new applications and list the bedrock-mantle endpoint as unsupported for Nova 2 Lite and Nova Pro.
• Nova 2 Lite regions: geo cross-region profiles for the US (us.), EU (eu.) and Japan (jp.), and a global profile (global.) routing across the Americas, EMEA and Asia Pacific. No region on the card lists in-region inference, so some regions, such as eu-west-2 (London) and ap-south-1 (Mumbai), reach it only through the global profile.
• Nova Pro regions: in-region in us-east-1, us-gov-west-1 (GovCloud), eu-west-2, ap-southeast-2, ap-southeast-3 and me-central-1, plus US, EU and APAC geo profiles. Global cross-region inference is not listed for Nova Pro.
• Nova 2 Sonic and Nova 2.5 Sonic regions: in-region only in us-east-1, us-west-2, eu-north-1 and ap-northeast-1, with a default quota of 20 concurrent bidirectional streaming sessions per account per region.
• Customisation: Nova 2 Lite customisation (pre-training, supervised fine-tuning and reinforcement fine-tuning) through Bedrock and Amazon SageMaker AI, and Nova Forge.
• Apps and open weights: no consumer chat app or downloadable weights are listed on the Nova pages checked.

Cross-region inference matters for data residency: geo profiles keep routing within a geography, while the global profile routes worldwide. See the Amazon Bedrock Security Guide for account-level controls.

Restricted-access Nova models

As of 2026-11-15. Restricted-access programmes for Nova models: not stated by Amazon on the pages checked (2026-11-15).

What is the context window of Nova models?

Nova 2 Lite and Nova 2 Sonic have a 1M token context window on their Bedrock model cards. Nova 2.5 Sonic has 256K on its card (the Nova models page says up to 1M). Nova Pro and Nova Lite have 300K, and Nova Micro has 128K. Maximum output is 64K for the Nova 2 and 2.5 models and 10K for the first-generation understanding models. See LLM Context Windows Explained.

Is any Nova model open source?

No. Amazon does not publish weights for any Nova model on the pages checked; all Nova models are accessed as managed services on Amazon Bedrock, under AWS service terms. Customisation is offered as a managed service (Bedrock, SageMaker AI and Nova Forge) rather than through downloadable weights. See Open-Weight vs Closed LLMs Explained.

Model cards and system cards

Bedrock model cards are specification pages. AWS publishes responsible AI material as separate AI Service Cards. See Model Cards and System Cards Explained.

Benchmarks

As of 15 November 2026. Provider-reported and independent results are kept in separate columns.

See How to Read LLM Benchmarks for how provider-reported and leaderboard figures differ.

Deprecations and retirements

As of 15 November 2026, from the Bedrock model lifecycle pages and model cards.

How the Bedrock lifecycle works: a model moves from Active to Legacy, and Bedrock notifies customers of the EOL date when the Legacy period starts. The legacy period is at least 6 months for every Nova model card checked. After EOL the model is removed from all regions and requests fail; migration is not automatic. Bedrock has two lifecycle pages: one for models launched on or after 7 September 2026 and a legacy page for earlier models, which covers every Nova model except Nova 2.5 Sonic. See the AI Model Retirement Guide and the LLM Deprecation and Retirement Tracker.

Data handling and security

• Training use: the Amazon Bedrock FAQs state that customer content is not used to improve the base models and is not shared with any model providers, and that AWS and third-party providers will not use inputs or outputs to train Amazon Nova, Amazon Titan or other third-party models.
• Storage location: the FAQs state that customer content is encrypted and stored at rest in the AWS Region where Bedrock is used. A default retention period is not stated by Amazon on the pages checked (2026-11-15).
• Model provider isolation: the Bedrock data protection page states that models run in Model Deployment Accounts owned by the Bedrock service team, and that model providers have no access to Bedrock logs or to customer prompts and completions.
• Transport and FIPS: AWS requires TLS 1.2 and recommends TLS 1.3, and offers FIPS endpoints for FIPS 140-3 validated modules.
• Guardrails and abuse detection: the Nova 2 Lite card lists Bedrock Guardrails and abuse detection as supported; the Nova 2.5 Sonic card lists watermarking.
• Credentials: the model card samples authenticate with a Bedrock API key in the AWS_BEARER_TOKEN_BEDROCK environment variable. Keep such keys out of code, scope them narrowly and rotate them, or use IAM roles where the workload supports them.

See the Amazon Bedrock Guardrails Configuration Guide, the LLMjacking Guide and the Foundation Model Selection Guide.

How to choose an Amazon Nova model: factual criteria

• Modality: text only (Nova Micro), text, image and video in with text out (Nova 2 Lite, Nova Pro, Nova Lite), speech in and out (Nova 2 Sonic, Nova 2.5 Sonic) or embeddings (Nova Multimodal Embeddings).
• Context and output: 1M context and 64K output on Nova 2 Lite; 300K and 10K on Nova Pro and Nova Lite; 128K and 10K on Nova Micro.
• Reasoning: Nova 2 Lite offers reasoning, off by default, with low, medium and high effort levels.
• Data residency: Nova 2 Lite runs through geo or global cross-region profiles; Nova Pro offers in-region inference in five regions including GovCloud; the Sonic models are in-region only in four regions.
• Bedrock features: Agents, Flows and prompt management are listed on the Nova Pro card; client-side tool calling and abuse detection on the Nova 2 Lite card.
• Lifecycle: check the "EOL no sooner than" date and the live modelLifecycle status before building.
• Price: compare per-token rates for your region and service tier on the Bedrock pricing page.

Practitioner checklist

• Confirm the exact model ID, including any us., eu., jp., apac. or global. prefix, against the current Bedrock model card before you build.
• Query modelLifecycle through ListFoundationModels or GetFoundationModel for each Nova model you use, since card dates are "no sooner than" floors.
• Remove any remaining calls to amazon.nova-canvas-v1:0, amazon.nova-reel-v1:0, amazon.nova-reel-v1:1 and amazon.nova-sonic-v1:0, which are past their EOL dates.
• Check whether your region reaches Nova 2 Lite only through the global profile, and whether that meets your data residency rules.
• Read the per-token price for your region and tier on the Bedrock pricing page and record the date checked.
• Check the context window and maximum output for the specific model, not the family.
• Turn on Bedrock Guardrails where the model card lists support, and test reasoning effort settings for cost and latency.
• Store Bedrock API keys in a secrets manager and rotate them, or use IAM roles; never embed them in client code.

Standards and references

• Amazon: Amazon Nova models (accessed 2026-11-15)
• AWS: Amazon Nova 2 Lite model card, Amazon Bedrock User Guide (accessed 2026-11-15)
• AWS: Amazon Nova 2.5 Sonic model card (accessed 2026-11-15)
• AWS: Amazon Nova 2 Sonic model card (accessed 2026-11-15)
• AWS: Amazon Nova Pro model card (accessed 2026-11-15)
• AWS: Amazon Nova Lite model card (accessed 2026-11-15)
• AWS: Amazon Nova Micro model card (accessed 2026-11-15)
• AWS: Model lifecycle (for models launched prior to 7 September 2026) (accessed 2026-11-15)
• AWS: Model lifecycle, Amazon Bedrock (accessed 2026-11-15)
• Amazon: Amazon Nova pricing (accessed 2026-11-15)
• AWS: Amazon Bedrock FAQs (accessed 2026-11-15)
• AWS: Data protection in Amazon Bedrock (accessed 2026-11-15)

Related NHI Mgmt Group resources: Microsoft MAI and Phi Models, Mistral AI Models, Amazon Bedrock Security Guide, LLM Release Tracker.