Amazon Web Services announced Nova Pro in November 2024 as its first proprietary large language model, but the company has refused to publish the performance benchmarks, pricing details, or technical documentation that typically justify enterprise AI investments. Instead, AWS is asking customers to adopt the model based entirely on brand reputation and infrastructure integration, a strategy that fundamentally challenges how the AI industry has operated since the release of GPT-3.

Why Is Amazon Hiding Nova Pro's Performance Data?

Every major language model released in the past three years has published benchmark scores on standardized tests like MMLU (a knowledge assessment), HumanEval (coding ability), and GPQA Diamond (graduate-level science reasoning). These scores let enterprises compare models directly and predict real-world performance on their specific tasks.

Nova Pro has published none of these. AWS disclosed that the model offers a 300,000-token context window, meaning it can process roughly 300,000 words at once, and includes multimodal capabilities (processing text and images, at minimum). But the company has not revealed the model's parameter count, training data sources, vision capabilities compared to competitors, or whether it supports function calling and JSON mode, features that developers rely on for building applications.

The silence is particularly striking because AWS has published performance benchmarks for decades across its compute, storage, and database services. The decision to launch Nova Pro without equivalent AI performance data suggests either the benchmarks are not competitive or AWS believes its enterprise relationships can drive adoption without them. As one analysis noted, if AWS thought Nova Pro's scores were impressive, they would publish them. The silence implies the numbers either do not exist yet or do not tell the story AWS wants to tell.

What Are the Practical Barriers to Evaluating Nova Pro?

Enterprise AI procurement typically follows a structured evaluation process. Teams run pilots, compare vendors, and build scorecards based on performance metrics and cost projections. Nova Pro eliminates the data needed for this process.

• Pricing Unknown: Claude 3.5 Sonnet on Bedrock costs $3 per million input tokens and $15 per million output tokens. GPT-4o costs $2.50 input and $10 output. Without Nova Pro's rates, enterprises cannot build budgets, compare total cost of ownership, or determine whether the model is positioned as a premium or cost-competitive option.
• Performance Unverified: The 300,000-token context window is claimed but unverified. GPT-4 Turbo claims 128,000 tokens but performs poorly beyond 64,000 in practice. Until independent testing confirms Nova Pro maintains coherence across its full context window, the specification should be treated as a marketing number, not a deployment guarantee.
• Capability Gaps: AWS has not disclosed whether Nova Pro supports streaming responses, fine-tuning, function calling, or JSON mode. These features are standard in competing models and essential for many enterprise applications.

The absence of pricing is the most immediate barrier. Without rates, you cannot estimate whether this model is positioned as a premium option or a cost-competitive alternative to Claude or GPT-4o. You cannot calculate return on investment. You cannot even begin the procurement conversation with finance teams.

How Is AWS Competing Without Performance Data?

AWS is not competing on benchmarks because it is competing on infrastructure. The pitch is straightforward: if you are already running on AWS, Nova Pro eliminates the friction of third-party AI services. There is no need to send data outside your AWS environment. There is no compliance friction from external model providers. There is no third-party API dependency.

This represents a significant strategic shift. AWS Bedrock launched in September 2023 as a managed service hosting third-party models like Claude, Llama, and Stable Diffusion. Nova Pro marks AWS's transition from model aggregator to model maker. The company is extending its core business strategy, which has built a $90 billion cloud business by being the reliable, boring choice, into the AI market. EC2 instances do not publish benchmark wars against Azure virtual machines. RDS does not compete on database leaderboards. AWS sells infrastructure trust.

The question is whether enterprise AI procurement works the same way as compute and storage procurement. For decades, enterprises have accepted AWS's infrastructure decisions based on reliability and integration. The company is betting that AI adoption will follow the same pattern.

How to Evaluate Nova Pro Against Competing Models

• Request Independent Benchmarks: If you are considering Nova Pro, ask AWS for performance data on MMLU-Pro, HumanEval, GPQA Diamond, and any benchmarks relevant to your use case. If the company cannot provide this data, treat the model as unproven and require a longer pilot period before committing to production deployment.
• Conduct Cost Modeling: Without published pricing, request a detailed cost estimate from AWS for your expected token volume. Compare this estimate against Claude 3.5 Sonnet and GPT-4o pricing on Bedrock to determine whether Nova Pro offers cost advantages or premium positioning.
• Test Context Window Performance: If the 300,000-token context window is critical to your use case, run tests with documents and conversations at 200,000 tokens and above. Verify that the model maintains coherence and retrieval accuracy at full capacity, not just theoretical maximum.
• Verify API Compatibility: Confirm whether Nova Pro supports the specific features your applications require, including streaming, function calling, JSON mode, and fine-tuning. Do not assume feature parity with other Bedrock models.

What Does This Mean for AWS Customers?

For enterprises already committed to AWS, Nova Pro offers genuine advantages. Native integration eliminates data residency concerns. No third-party API calls mean lower latency and reduced compliance friction. The 300,000-token context window, if it performs as claimed, would be competitive with or exceed most alternatives.

But those advantages come with a cost: you are making a seven-figure AI infrastructure investment based on AWS's reputation, not this model's demonstrated performance. You are accepting that AWS's track record in compute and storage justifies trust in AI. You are betting that the company would not launch a proprietary model without confidence in its capabilities, even if that confidence is not publicly documented.

Meanwhile, AWS is simultaneously expanding its relationship with OpenAI. In April 2026, Amazon announced that AWS Bedrock now offers OpenAI's latest models, its Codex code-writing service, and a new Bedrock Managed Agents product specifically designed to use OpenAI's reasoning models. This move came after Microsoft's exclusive rights to OpenAI products were lifted, allowing AWS to offer OpenAI's services directly.

"This is the beginning of a deeper collaboration between AWS and OpenAI."

Amazon Web Services, official blog announcement

This creates an interesting dynamic: AWS is simultaneously launching its own proprietary model with no public performance data while deepening its partnership with OpenAI, a competitor with fully transparent benchmarks and pricing. Enterprises now have a choice between trusting AWS's infrastructure reputation with Nova Pro or using OpenAI's proven models through the same Bedrock service. The decision hinges on whether you value integration and data residency over performance transparency.

For the AI industry, Nova Pro represents a test case. If AWS successfully drives enterprise adoption without publishing benchmarks, it signals that infrastructure trust can overcome the performance transparency that has defined AI model competition since 2022. If enterprises demand benchmarks before committing, it reinforces that AI procurement requires the same rigor as any other technology investment. The outcome will shape how other cloud providers approach proprietary AI models in 2026 and beyond.