In brief
Amazon Nova is the family of proprietary AI models developed by Amazon, available exclusively through Amazon Bedrock, AWS’s cloud AI platform. The lineup was announced at the AWS re:Invent conference in December 2024 and covers a broad spectrum: ultra-lightweight text models, multimodal models, image generation (Nova Canvas), and video generation (Nova Reel).
Nova marks a strategic turning point for Amazon. Before 2024, the company primarily offered third-party models on Bedrock — Anthropic’s Claude above all — alongside an aging proprietary lineup called Titan. With Nova, Amazon owns its models entirely, which fundamentally changes its economics. Nova’s central positioning: aggressive pricing with deep integration into the AWS ecosystem, at the cost of general intelligence performance that trails frontier models from Google and Anthropic according to third-party evaluations.
In December 2025, Nova 2 added extended reasoning, voice capabilities, and native agentic support. In April 2026, Amazon announced an additional investment of up to $25 billion in Anthropic — paradoxically reinforcing both bets simultaneously.
Four actors to distinguish clearly
Before going further, a clarification is essential. When discussing Amazon and AI, four distinct realities coexist under the same roof, and conflating them leads to analytical errors.
Nova (this profile) refers to the proprietary AI models developed by Amazon — the subject of this article.
Anthropic and Claude are a strategic and financial partner of Amazon, whose models are available on Bedrock. Amazon has invested a total of up to $33 billion in Anthropic since 2023, but remains a minority investor. Claude is economically distinct from Nova: when Amazon charges customers for Claude usage, it shares a portion of that revenue with Anthropic.
Trainium refers to AWS’s proprietary AI chips designed for model training. This is hardware infrastructure, not a model. Nova uses Trainium internally for its own training runs, but this topic is covered in the dedicated article on alternative chips. It will not be developed here.
Titan is Amazon’s proprietary model lineup prior to Nova (Titan Text, Titan Embeddings, Titan Image Generator). It remains available on Bedrock but is now positioned as a legacy offering.
In plain terms: Nova = Amazon’s proprietary models from 2024 onward. Claude = partner available on Bedrock. Trainium = hardware chips. Titan = legacy lineup. Four distinct realities, one AWS cloud.
Context: why Amazon launched Nova
The Anthropic dependency, a structural tension
Amazon Bedrock achieved real commercial success thanks to Anthropic’s Claude models. More than 100,000 AWS customers use Claude through Bedrock. But this success comes with an economic constraint: for every Claude call billed to a customer, Amazon remits a share to Anthropic. This remittance, estimated by some market observers at 30 to 50%, represents a structural cost on AWS margins.
Nova solves this mathematically: when a customer chooses Nova Pro over Claude Sonnet, Amazon retains 100% of the revenue.
The Information described this relationship as a “frenemy” dynamic: Amazon needs Claude to attract customers to Bedrock, yet also has an interest in those same customers migrating to Nova once Nova becomes sufficiently competitive.
Competition from GPT-4o and Gemini
OpenAI and Google dominated LLM conversations in 2023-2024. For AWS enterprise customers, the risk of disintermediation was real: why stay in the AWS ecosystem if the most capable models are only accessible elsewhere? Nova is Amazon’s answer to this threat: offering competitive proprietary models, natively integrated into AWS, at a price point that neither OpenAI nor Google can easily match without eroding their margins.
Timeline 2024-2026
December 2024 — re:Invent 2024 — Nova v1 launch. On December 3, 2024, Andy Jassy announces the Amazon Nova family during the AWS re:Invent keynote in Las Vegas. Nova Micro, Nova Lite, and Nova Pro are available immediately. Nova Premier is announced for expected availability in Q1 2025. Nova Canvas (images) and Nova Reel (video) are available simultaneously [sources: AWS official blog, TechCrunch, Amazon What’s New].
March 2025 — Technical report. Amazon publishes the official technical report for the Nova family (March 17, 2025), including MMLU, HumanEval, and other standard evaluation benchmarks [source: Amazon Science, official technical report].
April 2025 — Nova Premier reaches general availability. Nova Premier becomes generally available on April 30, 2025, deployed via cross-region inference in US East (N. Virginia), US East (Ohio), and US West (Oregon) [sources: TechCrunch April 30, 2025, AWS What’s New].
December 2025 — re:Invent 2025 — Nova 2. AWS announces Amazon Nova 2 at re:Invent 2025, with four new models: Nova 2 Lite, Nova 2 Pro, Nova 2 Sonic (voice), and Nova 2 Omni (unified multimodal). New capabilities: parameterizable extended reasoning, 1 million token context window, native MCP support [sources: HPCwire/AIwire, AWS Blog, builder.aws.com].
December 2025 — two services around the models. Alongside Nova 2, AWS presents Nova Forge, which lets an organisation build its own variants — called Novellas — by blending its proprietary data with Nova’s capabilities. What makes it unusual is the level of access: AWS speaks of open training, meaning access to pre-trained, mid-trained and post-trained checkpoints, and therefore the ability to inject one’s own data at every training stage rather than only at the end. The Nova Forge SDK is available in March 2026. Nova Act, for its part, targets agents that drive a browser; it relies on a version of Nova 2 Lite trained by reinforcement learning across simulated web environments, and AWS reports 90% reliability on early customer workflows — a vendor figure, not replicated. [UNVERIFIED]
April 2026 — Anthropic partnership expansion. Amazon announces an additional investment of up to $25 billion in Anthropic, along with a commitment by Anthropic to spend more than $100 billion over ten years on AWS infrastructure [sources: CNBC, April 20, 2026]. This investment reinforces both bets: Nova (proprietary models) and Claude (Bedrock partner).
The Nova v1 lineup (December 2024)
Four text and multimodal models
The initial Nova generation covers four tiers of capability and cost.
Nova Micro is a text-only model with a 128,000-token context window. It is designed for use cases where latency and cost take priority over quality: text classification, structured data parsing, request routing. It is the most affordable model in the lineup.
Nova Lite is multimodal — it accepts text, images, and video as input to produce text output. Context window: 300,000 tokens. Positioning: real-time interactions, document analysis, visual question-answering.
Nova Pro is multimodal (same input capabilities as Nova Lite) with a 300,000-token context window. It is the flagship model of the v1 lineup: Amazon presents it as the best balance of accuracy, speed, and cost for agentic workflows and complex tasks.
Nova Premier is multimodal with a 1 million-token context window. Generally available since April 30, 2025, it is positioned for complex reasoning tasks and custom model distillation. Amazon describes it as the “best teacher” for creating custom models through distillation.
Nova Canvas and Nova Reel — multimedia generation
Nova Canvas is a generative image model. Its architecture is based on a diffusion transformer. It generates images up to 4.2 megapixels in any aspect ratio, with support for inpainting (modifying a region of an image), outpainting (extending an image beyond its borders), and background removal. Usable via English text prompts (up to 1,024 characters).
Nova Reel is a generative video model. It produces videos at 1280×720 pixels at 24 frames per second, with a duration of 6 seconds, from a text prompt alone or a text prompt combined with a reference image.
In plain terms: the Nova v1 lineup covers the full spectrum — from the ultra-affordable text-only model (Micro) to the million-token model for distillation (Premier) — plus image generation (Canvas) and video generation (Reel). All of it, exclusively on AWS Bedrock.
Nova 2 (December 2025) — extended reasoning and agentic capabilities
The second generation of Nova, announced at re:Invent 2025, adds three structural capabilities.
Extended reasoning: Nova 2 Pro can activate a deep reasoning mode with three parameterizable intensity levels — low, medium, high — allowing operators to calibrate the depth of analysis against the acceptable cost. This mechanism is analogous to what Claude 3.7 and Gemini 2.5 Pro have offered since 2025.
Nova 2 Sonic: a speech-to-speech model, conversational and multilingual, designed for telephony integrations and real-time voice assistants.
Nova 2 Omni: a model accepting combined inputs (text + images + video + voice) with text and image outputs. It unifies modalities within a single model.
Nova 2 also introduces native support for the MCP (Model Context Protocol) over the internet, simplifying integration into multi-tool agentic architectures. A native code interpreter and a web grounding module (real-time search) are also included.
Vocabulary
- Extended thinking: a reasoning mode in which the model produces intermediate reflection steps before delivering the final response. LLMs that activate this mode generally achieve better performance on mathematical and logical reasoning tasks.
- MCP (Model Context Protocol): a standardized protocol enabling an AI model to call external tools (databases, APIs, files) in a structured way. Anthropic introduced this protocol in 2024; it has become a de facto standard for agentic architectures.
- Speech-to-speech: end-to-end voice processing — the model receives audio, understands, reasons, and responds in audio — without intermediate conversion to text.
Pricing strategy — the most documented competitive advantage
Pricing table (as of February 25, 2026)
| Model | Input ($/million tokens) | Output ($/million 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 |
| Nova Reel | $0.08/second of video | — |
| Nova Canvas | $0.04–$0.08/image | — |
Source: go-cloud.io, updated February 25, 2026. Note: Nova Pro on Bedrock became paid from April 23, 2025 onward (the preceding period was handled differently during preview).
Comparison with GPT-4o
OpenAI’s GPT-4o is priced at $2.50/$10.00 per million tokens (input/output). Nova Pro is therefore three times cheaper on input and three times cheaper on output than GPT-4o. Amazon estimates the savings at approximately $68,000 per year for a load of 10 million requests per month, compared to GPT-4o [official Amazon source — not corroborated by an independent third-party source].
Latency — another measurable advantage
According to Artificial Analysis (consulted in April 2026, independent third-party source), Nova Pro reaches 118.7 output tokens per second, with a time-to-first-token of 1.12 seconds. A third-party comparison (docsbot.ai) finds that Nova Pro generates responses 21.97% faster than GPT-4o.
In plain terms: if price and generation speed are the primary criteria, Nova Pro is a serious candidate compared to GPT-4o. The advantage is documented by independent third-party sources — not just by Amazon.
Strengths
Native AWS integration — a systemic advantage
For a company already in the AWS ecosystem, Nova radically simplifies deployment. Nova models are natively connected to Amazon S3, AWS Lambda, Amazon Kendra, Amazon SageMaker, Amazon Bedrock Agents, AWS CloudTrail, and IAM. This means data governance, auditing, and compliance fit into existing tools — without additional integration friction.
Another important point for compliance: Amazon explicitly states that customer inputs and outputs in Bedrock are not used to train Nova. For enterprises with regulatory constraints (healthcare, finance, law), this is a documented compliance commitment.
Long context window — Nova Premier and Nova 2
Nova Premier offers a 1 million-token context window — comparable to Gemini 1.5 Pro, which introduced this threshold as early as 2024. This capability is useful for analyzing long documents, exploring large codebases, or processing extended videos.
Nova 2 Pro also extends the context window to 1 million tokens for its generation of models.
Nova Reel — video generation cheaper than competitors
At $0.08 per second of video, Nova Reel is positioned below Runway Gen-3 Alpha ($0.096/s) and Kling 1.5 ($0.12/s) according to official Amazon documentation [official source — to be verified as Runway and Kling pricing evolves]. For enterprises seeking short-form generative video integrated into AWS, Nova Reel avoids leaving the ecosystem to reach third-party providers.
Custom model distillation
Nova Premier is explicitly positioned as the “teacher” in a distillation process: an enterprise can use Nova Premier to train a smaller, more specialized model optimized for its internal use cases. This is a value-added feature for organizations that want a custom model without starting from scratch.
Documented limitations
General intelligence performance — trailing according to third-party evaluations
According to Artificial Analysis (independent third-party source, consulted in April 2026):
- Nova Pro is ranked #56 out of 71 models (Intelligence Index score of 13/71 among non-reasoning models). Artificial Analysis describes it as “among the least intelligent models, but well-priced for its price tier.”
- Nova Premier is ranked #39 out of 71 (Intelligence Index score of 19/100). Artificial Analysis notes it is “below average for comparable models” (average: 22) and “somewhat expensive for non-reasoning models of similar price” at a blended rate of $5.00 per million tokens.
These Artificial Analysis scores are independent — they reflect a third-party methodology not affiliated with Amazon.
For comparison, here are the self-published benchmarks from Amazon’s official technical report (March 2025) [official Amazon source]:
- Nova Pro — MMLU: 85.9% (vs. Claude 3.5 Sonnet: 89.3%, GPT-4o: 88.7%)
- Nova Pro — HumanEval: 89.0% (pass@1)
The gap between MMLU scores (moderate on this specific benchmark) and Artificial Analysis Intelligence Index rankings (more pronounced) reflects different methodologies. Amazon’s self-published benchmarks measure specific tasks; Artificial Analysis’s Intelligence Index aggregates a broader evaluation across many tasks. Both are cited here with their respective sources.
In plain terms: Nova is competitive on price and latency. On raw intelligence as measured by independent third parties, it trails frontier models from Google and Anthropic. Amazon’s official benchmarks show a smaller gap on MMLU, but those figures are self-published and do not substitute for a complete independent evaluation.
Nova Premier — slow generation speed
According to Artificial Analysis, Nova Premier generates 25.7 tokens per second — notably slow compared to peer models. For applications requiring real-time streaming or high-throughput generation, this is a concrete limitation to factor in before architecting a solution.
Nova Reel — limited to 6-second clips
Nova Reel video generation is capped at 6 seconds at 1280×720. This positions it for short-form use cases (advertising, social content) but rules out longer videos that competitors such as Sora (OpenAI) or Kling can produce.
Geographic availability — primarily US
At launch and across early updates, Nova is primarily available in US AWS regions. EU and APAC regions are more limited, which may be a barrier for enterprises subject to data localization requirements (GDPR in Europe, for instance).
Training data opacity
Amazon mentions “licensed data, proprietary data, open-source datasets, and publicly available data” without disclosing specific sources or their proportions. This opacity is standard in the proprietary LLM industry, but remains a documented limitation for enterprises seeking to assess contamination or bias risks in training data.
The Anthropic relationship: claimed complementarity, real economic tension
Amazon officially positions Nova and Claude as complementary on Bedrock: Nova Micro for maximum latency and minimal cost, Claude for tasks requiring high intelligence and reliability on complex reasoning. The commercial argument is that a customer can combine both based on the task.
The economic reality is more nuanced. Every Claude request on Bedrock costs Amazon (through revenue sharing with Anthropic). Every Nova Pro, Nova Lite, or Nova Micro request costs nothing in revenue-sharing terms. The Information documented tensions around the Nova launch, describing the Amazon-Anthropic relationship as a “frenemy” dynamic.
The April 2026 announcement — an additional $25 billion investment in Anthropic — illustrates this duality well. Amazon reinforces simultaneously its most important partner (Claude) and its own competing lineup (Nova). Both bets coexist for now, without Amazon formally choosing one over the other.
Decision matrix — Nova or something else?
| Context | Recommendation | Why |
|---|---|---|
| AWS enterprise, high volume, repetitive tasks (classification, parsing, Q&A) | Nova Micro or Nova Lite | Lowest pricing on Bedrock, native IAM/CloudTrail integration without overhead. |
| AWS agentic workflow, cost/quality balance needed | Nova Pro | Good latency (118.7 t/s per Artificial Analysis), price 3× lower than GPT-4o, native agentic support. |
| Analysis of very long documents in AWS, model distillation | Nova Premier | 1M token context, the only Nova model positioned for distillation. Low speed (25.7 t/s) to anticipate. |
| Complex reasoning, critical reliability, frontier-level quality | Claude Sonnet/Opus via Bedrock or Gemini | Nova trails on general intelligence per third-party benchmarks (Artificial Analysis #56 and #39/71). |
| Short-form video generation integrated into AWS | Nova Reel | $0.08/s, native Bedrock integration — useful for staying in the AWS ecosystem without a third-party contract. |
| Outside AWS, direct comparison with GPT-4o or Gemini API | OpenAI or Google Gemini | Nova is Bedrock-exclusive — no independent direct API access. |
Three practical rules for teams evaluating Nova:
Start with Nova Micro or Lite for high-volume repetitive tasks — this is where the price advantage is most significant and least sensitive to general intelligence gaps.
Do not migrate complex reasoning tasks to Nova based solely on Amazon’s official benchmarks — independent third-party evaluations (Artificial Analysis) show a notable gap with frontier models on overall intelligence.
Test Nova 2 Pro for agentic use cases — native MCP support and extended reasoning make it a candidate for multi-tool workflows, even though complete third-party benchmarks for Nova 2 are still sparse at the time of writing.
Positioning in the 2026 AI landscape
Amazon Nova occupies a specific position in the 2026 landscape: that of a hyperscaler playing vertical integration rather than frontier performance. The most instructive comparison is perhaps with Google, which follows a similar strategy with Gemini (proprietary models) while maintaining deep integration in Google Cloud.
The key difference: Google DeepMind has produced with Gemini 2.5 Pro a model that reaches the top rankings in third-party evaluations, whereas Nova Premier (the most capable model in the v1 lineup) remains below average in its price category according to Artificial Analysis. For Amazon, the argument is not “the best model in the world” but “the best cost-to-AWS-integration ratio on the market.”
This strategy may be sufficient to capture a significant share of the AWS enterprise market — where ecosystem lock-in, compliance, and cost control often weigh more heavily than a few MMLU percentage points.