CtrlK
Amazon Nova is a family of foundation models developed by Amazon and offered through Amazon Bedrock. The model family was announced on December 3, 2024, during the AWS re:Invent conference in Las Vegas. Amazon Nova includes text-only and multimodal understanding models (Nova Micro, Nova Lite, Nova Pro, and Nova Premier), an image generation model (Nova Canvas), a video generation model (Nova Reel), and a speech-to-speech model (Nova Sonic). In December 2025, Amazon introduced the second generation of the family, Nova 2, with updated models and new capabilities including extended thinking and built-in tools.
Amazon CEO Andy Jassy and AWS CEO Matt Garman introduced Amazon Nova at the 2024 AWS re:Invent keynote on December 3, 2024. The announcement positioned Nova as Amazon's first-party answer to models from OpenAI, Google, Anthropic, and Meta. Before Nova, Amazon had relied primarily on third-party models available through Bedrock (including Anthropic's Claude, Meta's Llama, and Mistral's models) rather than offering competitive in-house alternatives. The Titan model family, Amazon's earlier generation of proprietary models, had not gained significant traction against the competition.
At launch, Nova Micro, Nova Lite, and Nova Pro were made generally available in Amazon Bedrock. Nova Premier, the largest model in the family, was still in training at the time of announcement and was released later on April 30, 2025.
Amazon stated that Nova models are "at least 75 percent less expensive than the best performing models in their respective intelligence classes in Amazon Bedrock," and positioned the family as offering a strong balance between cost and capability. The models are available exclusively through Amazon Bedrock, Amazon's managed service for accessing foundation models via API. Early enterprise adopters included SAP, which integrated Nova models into its SAP AI Core generative AI hub to power skills for Joule, SAP's AI copilot.
The first-generation Amazon Nova family consists of several models, each targeting a different point on the cost-performance spectrum.
Nova Micro is a text-only model designed for the lowest latency and lowest cost among the Nova family. It accepts text input and generates text output, with a context window of 128,000 tokens. Amazon optimized Micro for tasks such as text summarization, translation, content classification, interactive chat, brainstorming, and basic mathematical reasoning. Because it processes only text (no images or video), it is the fastest model in the lineup and the cheapest per token. At $0.035 per million input tokens and $0.14 per million output tokens, Micro is one of the least expensive commercial large language model APIs on the market. This makes it attractive for high-volume, latency-sensitive applications like real-time chatbots and content filtering pipelines.
Nova Lite is a low-cost multimodal model that processes text, images, video, and documents. It supports a context window of 300,000 tokens, which is equivalent to roughly 30 minutes of video input. Amazon designed Lite as a budget-friendly option for multimodal workloads where speed and cost matter more than peak accuracy. Common use cases include document understanding, visual question answering, and video summarization. At $0.06 per million input tokens, Lite costs less than two cents per typical request, making it viable for processing large volumes of visual content in automated pipelines.
Nova Pro is the mid-range multimodal model, positioned as offering the best combination of accuracy, speed, and cost across a broad range of tasks. Like Lite, it supports text, image, and video inputs with a 300,000-token context window. Amazon reported that Nova Pro performed equal to or better than GPT-4o on 17 out of 20 benchmarks and outperformed Gemini 1.5 Pro on 16 out of 21 benchmarks. Nova Pro is particularly strong in agentic workflows that involve calling APIs and tools to complete multi-step tasks. On instruction-following evaluations, including the Comprehensive RAG Benchmark (CRAG) and the Berkeley Function Calling Leaderboard, Pro scored well against competing models in its price class.
In a summarization cost comparison published by Amazon, Nova Pro cost approximately $0.00068 per task compared to GPT-4o's $0.00122, a 44% cost reduction for equivalent work. With throughput of approximately 200 tokens per second, Pro is also roughly twice as fast as GPT-4o on inference speed.
Nova Premier is the most capable model in the first-generation Nova family. It was released on April 30, 2025, after being announced as "still in training" at the December 2024 launch event. Amazon had originally targeted Q1 2025 for Premier's release, but the model arrived at the end of April. Premier supports a context window of 1,000,000 tokens, allowing it to process large codebases, long documents, and extended videos in a single prompt. In addition to being the strongest Nova model for complex reasoning, Premier was designed to serve as a "teacher" model for model distillation, transferring its capabilities to smaller, faster Nova models for specific use cases.
On Amazon's internal benchmarks, Premier scored 87.4 on MMLU (undergraduate-level knowledge), 87.4 on MMMU (multimodal understanding), and 86.3 on SimpleQA. On IFEval (instruction following), Premier scored 91.5%. On the Berkeley Function Calling Leaderboard, Premier surpassed both Claude 3.5 Sonnet and Claude 3.7 Sonnet, though it fell short of GPT-4.5.
However, independent evaluations found Premier trailing competitors on several harder tasks. On SWE-bench Verified (a coding benchmark that tests a model's ability to resolve real GitHub issues), Premier achieved 42.4%. While this exceeded GPT-4.5's score on the same benchmark, it fell short of both Claude Sonnet versions. Premier also scored lower than leading competitors on GPQA Diamond and AIME 2025, which test graduate-level science reasoning and math competition problems respectively. Artificial Analysis, an independent benchmarking organization, ranked Nova Premier 34th out of 63 models on its Intelligence Index, with a score of 19 out of a median of 22 for non-reasoning models.
Amazon published a technical report titled "The Amazon Nova Family of Models: Technical Report and Model Card" through Amazon Science in March 2025. According to the report, Nova Micro, Lite, and Pro are all based on the Transformer architecture. Amazon did not publicly disclose the parameter counts for any Nova model, nor did the report detail specific architectural modifications beyond the base Transformer design.
The models went through a multi-stage training process. Pre-training used a mixture of large-scale multilingual and multimodal data drawn from licensed data, proprietary data, open-source datasets, and publicly available data. The training data covered over 200 languages, with particular emphasis on 15 languages: Arabic, Dutch, English, French, German, Hebrew, Hindi, Italian, Japanese, Korean, Portuguese, Russian, Simplified Chinese, Spanish, and Turkish. For the multimodal models (Lite, Pro, and Premier), the training data also included images and video alongside text.
After pre-training, each model went through iterative fine-tuning stages. The first stage was Supervised Fine-Tuning (SFT) on instruction-demonstration pairs, including multimodal examples for Lite and Pro. Next came reward model training from human preference data, where human annotators compared model outputs and indicated which responses were better. Finally, the models underwent preference learning using Direct Preference Optimization (DPO) and Proximal Policy Optimization (PPO). These stages were applied iteratively, meaning the models went through multiple rounds of SFT, reward modeling, and preference optimization to progressively improve output quality.
Training infrastructure relied on Amazon's custom Trainium chips (Trn1 instances), as well as NVIDIA A100 (P4d instances) and H100 (P5 instances) accelerators. Distributed training was conducted on AWS SageMaker-managed Elastic Kubernetes Service (EKS) clusters, using Amazon FSx and Amazon S3 for data storage and checkpoint I/O. The use of Amazon's own Trainium chips alongside NVIDIA GPUs reflects Amazon's broader strategy of developing custom silicon to reduce dependency on third-party chip suppliers and lower training costs.
Amazon reported benchmark results for the Nova family in its technical report and launch blog posts. The following table summarizes selected results from Amazon's own evaluations.
| Benchmark | Nova Micro | Nova Lite | Nova Pro | Nova Premier |
|---|---|---|---|---|
| MT-Bench (median) | 8.0 | 8.0 | 8.5 | 8.6 |
| MMLU | - | - | 85.9 | 87.4 |
| MMMU | - | - | - | 87.4 |
| SimpleQA | - | - | - | 86.3 |
| IFEval | - | - | - | 91.5% |
| SWE-bench Verified | - | - | - | 42.4% |
| HumanEval (pass@1) | - | - | 89.0 | - |
| Context window | 128K | 300K | 300K | 1M |
Note: Dashes indicate that specific scores were not disclosed by Amazon for that model-benchmark combination.
On MT-Bench, Amazon reported that Nova Premier had the most stable performance across evaluation categories, with a narrow min-max margin of 1.5 points, while Nova Pro showed greater variability with a min-max margin of 2.7 points.
The following table compares Amazon Nova models against selected competing models based on publicly available data from Amazon's reports and third-party evaluations. Prices reflect on-demand rates as of 2025.
| Model | Provider | Modality | Context window | Input price (per 1M tokens) | Output price (per 1M tokens) | MMLU | SWE-bench Verified |
|---|---|---|---|---|---|---|---|
| Nova Micro | Amazon | Text | 128K | $0.035 | $0.14 | - | - |
| Nova Lite | Amazon | Multimodal | 300K | $0.06 | $0.24 | - | - |
| Nova Pro | Amazon | Multimodal | 300K | $0.80 | $3.20 | 85.9 | - |
| Nova Premier | Amazon | Multimodal | 1M | $2.50 | $12.50 | 87.4 | 42.4% |
| GPT-4o | OpenAI | Multimodal | 128K | $2.50 | $10.00 | 88.7 | - |
| Claude 3.5 Sonnet | Anthropic | Multimodal | 200K | $3.00 | $15.00 | 88.7 | 49.0% |
| Gemini 1.5 Pro | Multimodal | 2M | $1.25 | $5.00 | 85.9 | - |
Nova Pro and Nova Lite are notably cheaper than most competing models in their capability tiers. However, on harder benchmarks (GPQA Diamond, AIME 2025, SWE-bench Verified), Nova Premier trails the top models from Google, Anthropic, and OpenAI. The Nova family's main competitive advantage lies in cost efficiency rather than raw intelligence on the most difficult evaluations.
All Amazon Nova models are billed on a pay-as-you-go basis through Amazon Bedrock. Pricing is calculated per 1,000 tokens for text models. Amazon offers three service tiers: Standard (regular rates), Priority (premium service with preferential compute allocation), and Flex (discounted pricing for workloads that tolerate delayed processing). Batch processing is available at a 50% discount from on-demand rates.
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Nova Micro | $0.035 | $0.14 |
| Nova Lite | $0.06 | $0.24 |
| Nova Pro | $0.80 | $3.20 |
| Nova Premier | $2.50 | $12.50 |
For cached prompts (prompt caching), Nova Micro input tokens cost $0.00875 per million tokens, a 75% discount for repeated context. This is useful for applications that reuse the same system prompt or reference documents across many requests.
Nova Canvas image generation costs $0.04 per standard-quality image (up to 1024x1024) and $0.06 per premium-quality image. Higher-resolution images (2048x2048) cost $0.06 (standard) and $0.08 (premium).
Nova Reel video generation costs approximately $0.08 per second of generated video at 1280x720 resolution. A full 2-minute video generated with Reel 1.1 would therefore cost around $9.60.
Nova Canvas is an image generation model that creates images from text prompts and image inputs. It accepts English-language text prompts of up to 1,024 characters and produces images at resolutions up to 4.2 megapixels in any aspect ratio. The model competes with image generators such as DALL-E 3, Stable Diffusion, and Midjourney.
Canvas supports several editing operations beyond basic text-to-image generation. Users can replace objects or backgrounds in existing images using text prompts, and a background removal feature is built in. The model can also be fine-tuned on proprietary data to generate images that match specific style guidelines or brand requirements. Amazon positions Canvas for use in advertising, marketing, and entertainment content creation.
For safety and traceability, Nova Canvas applies an invisible watermark to every generated image. Amazon developed this watermark to be resistant to alterations such as rotation, resizing, color inversion, and flipping. In addition, Canvas attaches C2PA (Coalition for Content Provenance and Authenticity) content credentials to generated images. This metadata records the model, platform, and task type used, allowing downstream verification of image provenance. According to Amazon, the model's safety controls block 98.8% of potentially harmful prompts and 98.1% of toxic output content. AWS also offers IP indemnification for Nova Canvas outputs, meaning Amazon assumes liability for intellectual property claims related to images generated by the model when used through Bedrock.
Nova Reel is a video generation model that produces short videos from text and image inputs. At its initial launch in December 2024, Reel generated clips of up to 6 seconds at 1280x720 resolution and 24 frames per second. It competes with video generation models from Runway, Pika, Sora, and Kling.
In April 2025, Amazon released Nova Reel 1.1, which extended the maximum video length to 2 minutes. Videos are still composed of 6-second segments, but Reel 1.1 chains them together with style consistency across shots. The update introduced two generation modes:
Nova Reel also supports an image-to-video feature, where a reference image guides the generation of the video. Camera motion can be controlled through natural-language instructions in the prompt (for example, "slow pan left" or "zoom in on the subject"). Like Canvas, Reel includes built-in watermarking, content moderation, and C2PA content credentials. Nova Reel 1.1 is available in the US East (N. Virginia) AWS Region through Amazon Bedrock.
Nova Sonic is a speech-to-speech foundation model announced in April 2025. Unlike traditional voice systems that chain together separate speech-to-text, language model, and text-to-speech components, Sonic is a unified model that processes speech input and generates speech output natively. This reduces latency and enables more natural conversational dynamics, including real-time turn-taking. At launch, Sonic supported American and British English across various speaking styles and acoustic conditions.
At AWS re:Invent 2025 in December, Amazon announced Nova 2 Sonic with significant upgrades. The updated model expanded language support to English, French, Italian, German, Spanish, Portuguese, and Hindi. Nova 2 Sonic introduced "polyglot voices," where a single voice (such as the Tiffany voice) can switch between all supported languages within one conversation without changing voice identity.
The model also improved its handling of alphanumeric inputs, short utterances, and telephony-quality (8 kHz) audio. Asynchronous tool calling was added, allowing the model to continue responding to user input while tools execute in the background. This is particularly useful for call-center scenarios where the agent needs to look up account information while keeping the conversation going. Nova 2 Sonic integrates with Amazon Connect and third-party telephony providers including Vonage, Twilio, and AudioCodes, as well as open-source frameworks like LiveKit and Pipecat.
Pricing for Nova Sonic speech processing is $0.003 per 1,000 input tokens and $0.012 per 1,000 output tokens.
Nova Act is an AI model and SDK designed for browser automation. It was first announced as a research preview in March 2025 through nova.amazon.com, and reached general availability at AWS re:Invent in December 2025. The underlying model is a custom variant of Nova 2 Lite, specifically trained for browser control.
Nova Act is trained to perform actions within a web browser. The SDK allows developers to break complex workflows into atomic commands (for example: search for a product, complete a checkout, or answer questions about the current screen). Developers can also add detailed instructions to those commands, such as "don't accept the insurance upsell." The SDK is installed with a single command and agents are written in Python, combining natural-language instructions with code. The SDK supports interleaving Python logic (tests, breakpoints, asserts, thread pools for parallelization) with browser actions. For sensitive operations like password entry, developers can use direct Playwright-based browser manipulation instead of the AI model.
Nova Act was trained with reinforcement learning and extensive in-domain browser interaction data. Amazon reports that the latest version demonstrates over 90% reliability across early enterprise workflows, including automated quality assurance, complex form handling, and process execution.
At general availability, Nova Act integrates with AWS IAM for secure credentialing, Amazon S3 for data storage, and the Amazon Bedrock AgentCore Browser Tool for scalable cloud-based browser execution. Agents can be deployed via Docker containers to Amazon ECR with automatic infrastructure setup. The service is priced at $4.75 per agent hour, billed for real-world elapsed time while agents work. Human-in-the-loop wait time is excluded from billing. Parallel agents each generate separate charges.
Amazon provides several customization options for Nova models through Amazon Bedrock.
Nova Micro, Lite, and Pro all support fine-tuning with proprietary data. Developers can fine-tune with text data (all three models) or multimodal data (Pro and Lite). Two fine-tuning approaches are available:
| Method | Description | Best for |
|---|---|---|
| Parameter-Efficient Fine-Tuning (PEFT) | Updates only a small subset of model parameters | Lightweight adaptation with limited training data |
| Full fine-tuning | Updates all model parameters | Extensive training datasets requiring maximum customization |
On-demand inference pricing for fine-tuned Nova models is the same as for the base models. Nova Canvas also supports fine-tuning for custom image styles.
Amazon Bedrock supports model distillation, where a larger "teacher" model transfers its knowledge to a smaller "student" model. Nova Premier was designed specifically to serve as a teacher. The process works as follows: the developer selects a teacher model (such as Premier) and a student model (such as Pro or Lite), provides use-case-specific prompts, and Bedrock generates responses from the teacher model for those prompts. It then fine-tunes the student model on the teacher's responses using automated data synthesis techniques.
Amazon reported that distilling Nova Pro using Nova Premier as the teacher model achieved 20% higher accuracy on tool selection and API calling tasks, while maintaining the speed and cost advantages of the smaller model. This approach lets organizations get close to Premier-level quality for specific tasks at Pro-level pricing.
Announced at AWS re:Invent 2025, Nova Forge is an "open training" service that lets organizations build custom model variants starting from early Nova checkpoints. Unlike standard fine-tuning, which adapts a finished model, Forge allows customers to influence the model's training from earlier stages. It supports custom data mixing (blending proprietary data with training data) and custom reward functions evaluated in proprietary environments. Nova Forge targets specialized industries such as manufacturing, research and development, media production, and other domains requiring deep customization beyond what fine-tuning can achieve. The service has a $100,000 annual minimum subscription.
At AWS re:Invent 2025 on December 2, 2025, Amazon announced the Nova 2 model family. The second generation includes four models.
| Model | Status | Context window | Key features |
|---|---|---|---|
| Nova 2 Lite | Generally available | 1M tokens | Cost-effective reasoning, extended thinking, built-in web grounding and code interpreter |
| Nova 2 Pro | Preview | 1M tokens | Complex multi-step tasks, multi-document analysis, video reasoning, software migrations |
| Nova 2 Sonic | Generally available | - | Multilingual speech-to-speech, polyglot voices, telephony integration |
| Nova 2 Omni | Preview | 1M tokens | Text + image + video + speech input; text + image output; multi-speaker transcription |
A significant upgrade from the first generation is the expansion of context windows. Nova 2 Lite now supports 1 million tokens, up from 300,000 in the original Nova Lite. All Nova 2 text and multimodal models support extended thinking with step-by-step reasoning and task decomposition. Three thinking intensity levels (low, medium, and high) give developers control over the trade-off between speed, intelligence, and cost. The models also include built-in tools such as a code interpreter and web grounding (the ability to search the web and incorporate results into responses), and they support remote MCP (Model Context Protocol) tools.
Nova 2 Omni is particularly notable as what Amazon describes as the first reasoning model that accepts text, images, video, and speech inputs while generating both text and image outputs in a single model. It supports over 200 languages and features character consistency and text rendering in generated images, along with multi-speaker transcription and native translation capabilities.
| Model | Input (per 1K tokens) | Output (per 1K tokens) |
|---|---|---|
| Nova 2 Lite | $0.00125 | $0.0025 |
| Nova 2 Pro | $0.0003 (text/image/video/audio) | $0.01 (text) |
| Nova 2 Sonic (speech) | $0.003 | $0.012 |
| Nova 2 Sonic (text) | $0.00033 | $0.00275 |
| Nova 2 Omni (text/image/video input) | $0.0003 | $0.0025 (text), $0.04 (image) |
All Amazon Nova models are accessed exclusively through Amazon Bedrock, Amazon's fully managed service for foundation models. Bedrock provides a unified API for invoking Nova models alongside third-party models from Anthropic, Meta, Mistral, Cohere, and others.
Through Bedrock, Nova models support several deployment and usage modes:
Nova Micro, Lite, and Pro became generally available in Amazon Bedrock on December 3, 2024, initially in the US East (N. Virginia) AWS Region. Nova Premier launched on April 30, 2025. Nova Reel 1.1 became available in April 2025. Nova Sonic launched in April 2025, with Nova 2 Sonic following in December 2025. Nova 2 Lite reached general availability in December 2025, while Nova 2 Pro and Nova 2 Omni remain in preview as of early 2026.
Amazon has expanded regional availability over time, with additional AWS Regions gaining access based on demand.