Amazon Nova 2 Pro

Amazon's flagship reasoning model for complex agentic workflows, with a 1M-token context window.

Nova 2 Pro Review: 1M Context & 70% SWE-bench (2026)

Amazon Nova 2 Pro: 1M-token context, 70.0% SWE-bench Verified, built-in code interpreter and web grounding. Preview on AWS Bedrock since December 2025.

Amazon Nova 2 Pro is AWS's flagship Nova 2 model, released December 2, 2025 in Bedrock preview, with a 1,000,000-token context window and a 70.0% SWE-bench Verified score. It ships built-in code interpreter and web grounding tools and integrates natively with Bedrock AgentCore for production agent deployment.

Amazon Nova 2 Pro, released December 2, 2025 on AWS Bedrock, is a multimodal reasoning model with a 1,000,000-token context window and a 70.0% SWE-bench Verified score. It adds a built-in code interpreter, web grounding with citations, and native MCP tool support for agentic workflows.

Provider: Amazon · Family: Nova 2

Context window: 1,000,000 tokens · Max output: 64,000

Input modalities: text, image, pdf, video, tool-calls · Output: text, tool-calls

About Amazon Nova 2 Pro

Amazon Nova 2 Pro is a multimodal foundation model built by Amazon, announced at AWS re:Invent in December 2025 and released into Bedrock preview on December 2, 2025 (model ID amazon.nova-2-pro-preview-20251202-v1:0). It is the flagship of the Nova 2 family, sitting above Nova 2 Lite and alongside the speech-focused Nova 2 Sonic and multimodal-generation Nova 2 Omni. Amazon has not disclosed the exact architecture (dense versus mixture-of-experts) or parameter count, consistent with its practice for prior Nova generations. Nova 2 Pro was designed to handle highly complex, multistep workloads such as multi-document analysis, video reasoning, and software migrations, positioning it as AWS's answer to frontier reasoning models from OpenAI, Anthropic, and Google rather than a budget option. On independent third-party evaluation, Nova 2 Pro Preview scores 70.0% on SWE-bench Verified, Amazon's stated benchmark for autonomous software engineering capability. Artificial Analysis places its composite Intelligence Index at 14 out of a field of 78 tracked models, below the frontier tier occupied by GPT-5-series and Claude Opus models, but the model is explicitly optimized for cost and speed rather than topping every leaderboard. Amazon has not published standalone GPQA Diamond, AIME, or MMLU-Pro subscores separate from the composite index, and Nova system cards historically report fewer individual benchmark breakdowns than Anthropic or OpenAI system cards. Nova 2 Pro's headline architectural feature is its 1,000,000-token context window, a roughly 4x increase over the original Nova Pro's 300K limit, paired with a 64,000-token maximum output. This puts it ahead of GPT-5's and Claude's standard context tiers on raw window size, though Amazon has not published a needle-in-haystack recall benchmark at the full 1M depth, so retrieval quality at maximum context is unverified beyond Amazon's own claims. The model accepts multimodal input, text, images, PDFs, and video, and generates text output. It ships with built-in agentic tooling: a code interpreter that runs and evaluates Python within the same request, web grounding that retrieves live public information with citations, and support for remote MCP (Model Context Protocol) tools, letting it call external services directly. Function calling and structured output are supported through Bedrock's standard tool-use API. As of mid-2026 there is no confirmed native audio input/output on Nova 2 Pro itself, that capability lives in the separate Nova 2 Sonic model. Pricing is Amazon Bedrock's standard per-token model. Third-party trackers report Nova 2 Pro Preview at approximately $1.25 per 1M input tokens and $10.00 per 1M output tokens on the standard on-demand tier (Artificial Analysis), while a separate CloudPrice snapshot lists $2.19 input and $17.50 output with a $0.547 cached-read rate, likely reflecting a different regional or provisioned-throughput listing captured on a different date. Because Nova 2 Pro is still in preview, buyers should treat both figures as directional and confirm the live rate on the AWS Bedrock pricing console before committing production spend. A representative workload, summarizing a 500K-token document set at the lower quoted rate, costs roughly $0.63 in input tokens alone. Nova 2 Pro is available through Amazon Bedrock's standard API and console, with early access also granted to Amazon Nova Forge customers, AWS's service for pre-training custom Nova variants on proprietary data. It integrates natively with Amazon Bedrock AgentCore for production agent deployment, including AgentCore Policy for blocking unauthorized agent actions, and with Amazon CloudWatch for observability. As a proprietary Amazon model, there are no open weights, no Hugging Face listing, and no self-hosting path; access is exclusively through AWS-managed infrastructure and inherits Bedrock's regional availability and IAM-based authentication. Amazon's safety approach for Nova models runs through Bedrock's operational controls rather than a published responsible-scaling framework: Bedrock Guardrails for content filtering, AgentCore Policy for agent action governance, and standard Amazon content restrictions on CSAM, weapons, and malware generation. Amazon has not disclosed named third-party red-team partners for the Nova 2 family in the way Anthropic names METR or Apollo Research, and no public system card equivalent to a frontier-lab release has been located for Nova 2 Pro specifically as of mid-2026; the closest public documentation is Amazon Science's technical report for the Nova 2 family and AWS's Bedrock model card page. Nova 2 Pro is best suited to AWS-native enterprise teams already inside the Bedrock ecosystem who need long-context document and video analysis, agentic coding, or multi-step planning without leaving AWS's compliance boundary, and who value price-performance over topping every reasoning leaderboard. Teams chasing the absolute highest benchmark scores on SWE-bench, GPQA, or AIME should look at Claude Opus or GPT-5-series instead, both of which outscore Nova 2 Pro on independently verified reasoning evals. Teams needing self-hosted or open-weight deployment should also look elsewhere, since Nova 2 Pro has no downloadable weights. Training data cutoff for Nova 2 Pro has not been publicly disclosed by Amazon in a dated system card; Amazon's general practice trains on a mix of licensed data, public web text, and synthetic reasoning traces, and by default Bedrock does not use customer API inputs or outputs to train the underlying foundation models. Enterprise customers can rely on AWS's broad compliance footprint, SOC 2, ISO 27001, and HIPAA-eligibility apply to Bedrock as an AWS service, though Nova-specific compliance disclosures are not separately published from the parent AWS compliance program. As a preview model, Nova 2 Pro represents Amazon's move from Nova 1.0's cost-first positioning toward a more capable, still price-competitive reasoning tier. The prior generation, Nova Pro (1.0), priced at $0.80 input / $3.20 output per 1M tokens with a 300K context window, remains available on Bedrock for workloads that don't need Nova 2 Pro's expanded context or agentic tooling. Amazon has signaled that general availability and stabilized pricing for the full Nova 2 lineup will follow the December 2025 preview launch, with Nova 2 Pro reaching broader GA access being the next expected milestone.

Pricing

Artificial Analysis reports $1.25 per 1M input tokens and $10.00 per 1M output tokens for Nova 2 Pro Preview on standard on-demand Bedrock pricing. A separate CloudPrice snapshot lists $2.19 input / $17.50 output with $0.547 cached-read, likely a different tier or capture date. Because the model is still in preview, verify live rates on the AWS Bedrock pricing console before production use.

Key Features

Pros

Cons

Benchmarks

Frequently Asked Questions

What is Amazon Nova 2 Pro and who built it?

Amazon Nova 2 Pro is a multimodal foundation model built by Amazon, released into AWS Bedrock preview on December 2, 2025 (model ID amazon.nova-2-pro-preview-20251202-v1:0) alongside the wider Nova 2 family. It is the flagship reasoning model in the Nova 2 lineup, sitting above Nova 2 Lite and alongside the speech-focused Nova 2 Sonic and generation-focused Nova 2 Omni. Amazon has not disclosed the model's exact architecture or parameter count, consistent with prior Nova releases. It scores 70.0% on SWE-bench Verified and carries an Artificial Analysis Intelligence Index of 14. The model was designed for highly complex, multistep tasks such as multi-document analysis, video reasoning, and software migrations. It positions Amazon in the frontier reasoning-model race against OpenAI's GPT-5-series and Anthropic's Claude Opus, competing primarily on price-performance and context length rather than topping every benchmark. It costs roughly $1.25 per 1M input tokens and offers a 1,000,000-token context window.

How much does Amazon Nova 2 Pro cost per 1M tokens?

Artificial Analysis reports Nova 2 Pro Preview pricing at $1.25 per 1M input tokens and $10.00 per 1M output tokens on Amazon Bedrock's standard on-demand tier. A separate CloudPrice snapshot lists a different rate, $2.19 per 1M input tokens, $17.50 per 1M output, and $0.547 per 1M cached-read tokens, which likely reflects a different region, tier, or capture date during the preview window. Because Nova 2 Pro has not reached general availability, buyers should verify the live rate on the AWS Bedrock pricing console before committing to production budgets. At the lower quoted rate, summarizing a 500K-token document set costs roughly $0.63 in input tokens alone, and a daily 2M-input / 300K-output batch workload runs about $5.50. There is no confirmed batch-API discount for Nova 2 Pro as of mid-2026. The model cannot be self-hosted, so there is no infrastructure-cost alternative to the per-token API rate.

What is Amazon Nova 2 Pro's context window and max output?

Nova 2 Pro has a 1,000,000-token context window, roughly four times the 300K-token limit of the original Nova Pro (1.0) model, and a maximum output of 64,000 tokens. This puts its raw context size ahead of many competitors' standard tiers. Amazon has not published a needle-in-haystack or long-context recall benchmark verifying retrieval accuracy at the full 1M-token depth, so reliability that deep is currently unverified beyond Amazon's own product claims. The model handles multi-document sets and long video inputs within this window, which is the core use case it was built for. There is no separate extended-context tier beyond the standard 1M window. For document-heavy workloads, Amazon recommends chunking or summarization strategies if reliability at the extreme end of the window is a concern, since this has not been independently benchmarked.

How does Amazon Nova 2 Pro compare on benchmarks vs GPT-5 and Claude?

Nova 2 Pro scores 70.0% on SWE-bench Verified, a respectable agentic coding result, but its Artificial Analysis Intelligence Index of 14 places it well below frontier models like GPT-5-series and Claude Opus 4.8, which score in the 60s on the same composite index. Amazon has not published standalone GPQA Diamond, AIME 2025, or MMLU-Pro subscores for Nova 2 Pro separately from the composite index, making direct axis-by-axis comparison difficult. In practice, a meaningful benchmark gap of this size typically translates to less reliable multi-step reasoning and a higher rate of subtle errors on hard problems compared to Claude Opus or GPT-5. Nova 2 Pro's advantage is not raw intelligence but context length (1M tokens vs typically 200K-400K for rivals) and price: at roughly $1.25/$10.00 per 1M tokens it undercuts most frontier competitors on cost. Teams needing the highest possible reasoning accuracy should choose Claude Opus or GPT-5-series; teams needing the largest context window at the lowest cost should consider Nova 2 Pro.

Is Amazon Nova 2 Pro open source or proprietary?

Amazon Nova 2 Pro is fully proprietary. There are no downloadable weights, no Hugging Face listing, and no license permitting self-hosting or fine-tuning of the base model. Access is exclusively through Amazon Bedrock's managed API, authenticated via AWS IAM, with early access also extended to Amazon Nova Forge customers who pre-train custom variants on proprietary data through Amazon SageMaker AI rather than modifying the base weights directly. The model is governed by AWS's standard Service Terms, not a separate open-weights license like Llama's Community License or Apache 2.0. This means there is no VRAM or quantization consideration for end users, since inference always runs on AWS infrastructure rather than customer hardware. Commercial use is permitted under standard Bedrock consumption pricing with no separate commercial-use restriction beyond AWS's general service terms.

What modalities does Amazon Nova 2 Pro support?

Nova 2 Pro accepts multimodal input across text, images, PDFs, and video, and generates text output. It does not support native audio input or output; voice-focused use cases require pairing it with the sibling Nova 2 Sonic model, which handles unified speech-to-speech processing. Function calling and structured output are supported through Bedrock's standard tool-use API, and the model ships with three built-in agentic tools: a code interpreter that runs and evaluates Python within the same request, web grounding that retrieves live public information with citations, and native support for remote MCP (Model Context Protocol) tools that let it call external services directly. Parallel tool calls and computer-use-style automation are handled through Amazon's separate Nova Act service rather than being built into Nova 2 Pro itself. Compared to GPT-5 or Claude, which support native audio in their flagship multimodal tiers, Nova 2 Pro's modality set is narrower on the audio side but stronger on video input.

Does Amazon Nova 2 Pro train on user data?

No. By default, Amazon Bedrock does not use customer API inputs or outputs to train the underlying Nova foundation models, including Nova 2 Pro. Standard AWS Service Terms and Bedrock's data handling policy govern retention and use. AWS holds SOC 2 Type II, ISO 27001, and HIPAA-eligible status for Bedrock as a broader AWS service, with GDPR compliance support through Standard Contractual Clauses and, for EU customers, the AWS European Sovereign Cloud region that became generally available in January 2026. Amazon has not published a Nova-2-specific data governance disclosure separate from the parent AWS compliance program, so enterprise buyers rely on AWS's general trust center documentation (aws.amazon.com/compliance/) rather than a dedicated Nova system card for data policy details. Nova Forge customers who pre-train custom variants on proprietary data retain ownership of their fine-tuned model artifacts, distinct from the shared base Nova 2 Pro model.

Who is Amazon Nova 2 Pro best for and who should avoid it?

Nova 2 Pro is best for AWS-native enterprise teams building agentic coding workflows, multi-document or video analysis pipelines, and teams that want built-in code execution and web grounding without assembling custom agent scaffolding, especially if they're already inside the Bedrock ecosystem for compliance or billing reasons. It's also a strong fit for Nova Forge customers pre-training custom variants on proprietary data. Teams chasing the absolute highest reasoning benchmark scores should avoid it and use Claude Opus or GPT-5-series instead, both of which outscore Nova 2 Pro on the Artificial Analysis Intelligence Index and on independently verified coding evals. Voice-first product teams needing native audio input and output should avoid it and use Nova 2 Sonic or a dedicated speech model. Teams requiring self-hosted, air-gapped, or open-weight deployment should avoid it entirely, since it has no downloadable weights and runs exclusively on AWS infrastructure. Budget-conscious teams should also note it remains in preview with unsettled, inconsistently reported pricing as of mid-2026.

Visit Amazon Nova 2 Pro Official Page