Amazon Nova Pro 1.0: Should You Still Run It in 2026?

Nova Pro has an awkward position in 2026. It was launched as the balanced middle of Amazon's first Nova generation, and for a year it was the obvious Nova default. Then Nova 2 Lite arrived with a bigger window, longer output and a lower list price, and AWS's own migration guide began recommending the move.

That does not make Pro useless. It does mean the burden of proof has flipped: the question is no longer whether Pro is good enough, but what specifically keeps you on it.

Pro is now priced above its successor

Start with the numbers, because they drive the rest. At the time of writing OpenRouter lists Nova Pro 1.0 at $0.80 per million input tokens and $3.20 per million output. Nova 2 Lite lists at $0.30 and $2.50.

On paper the successor wins on every axis that a spec sheet shows. Per the Bedrock model cards:

• Nova Pro 1.0: 300K context, 10K maximum output, text, image and video input, knowledge cutoff October 2024, launched 5 December 2024.
• Nova 2 Lite: 1M context, about 64K maximum output, the same input types, knowledge cutoff October 2025, launched 2 December 2025.

AWS acknowledges the move from Pro to a model called Lite looks like a step down in tier, and recommends it anyway. The Nova 2 Lite guide covers what that model adds and what its reasoning settings cost.

Reading the 17-of-20 claim

Amazon's December 2024 launch announcement compared Nova Pro with the frontier models of the time: equal or better than GPT-4o on 17 of 20 benchmarks, than Gemini 1.5 Pro on 16 of 21, and than Claude 3.5 Sonnet v2 on 9 of 20.

The third number is the informative one. Against the strongest comparison model, Pro matched or beat it on fewer than half the benchmarks. And Amazon counted overlapping 95% confidence intervals as equal, so some of those matches may be narrow losses. Read together, the claims say Pro was a capable model just below the top of its era, not a peer of it.

Hands-on reviews at launch were consistent with that. Simon Willison found Pro's code output spot on for a simple task, but noted that a long PDF conversion stopped around page nine and missed some tables. Long-document extraction is worth testing specifically if that is your use.

Reasons to keep running it

There are real ones, and none of them is raw capability.

• No reasoning variance. Pro has no extended thinking, so cost per request is predictable from input and visible output. Nova 2 Lite with reasoning on spends billed tokens you cannot see.
• Existing customisation. AWS lists fine-tuning and Provisioned Throughput for Nova Pro. A tuned Pro deployment that already passes your evaluation is not something to throw away casually.
• Service tiers. The Bedrock card lists Standard, Priority and Flex tiers for Pro. Flex in particular suits workloads that tolerate delay in exchange for a lower rate.
• Regional coverage. Pro is offered in-Region in several places, including GovCloud, which can matter more than any benchmark for regulated work.

If none of those applies to you, the per-token price argument for staying is gone.

Pro as a distillation student

One of the more interesting uses of Nova Pro was never as a general model. When AWS launched Nova Premier in April 2025, it described using Premier as a teacher to distil a custom Nova Pro for tool selection and API calling, and reported that the distilled Pro had 20% higher accuracy on API invocations than the base model.

The pattern is sound regardless of the specific models: a large model generates high-quality examples for one narrow task, and a cheaper model is trained on them. Premier itself has since passed its announced end-of-life date on Bedrock, as the Nova Premier guide explains, but a Pro model distilled before that still works. Model distillation explained covers when the technique pays off.

A migration checklist to Nova 2 Lite

AWS says the core API structure is unchanged, so most of the code change is the model ID. The risk is behavioural, not syntactic. Before switching:

1. Pin what you have. Bedrock IDs carry a version suffix, such as amazon.nova-pro-v1:0. Record the exact ID and inference profile so you can roll back.
2. Build a regression set. Fifty to a hundred real requests with outputs you have already accepted.
3. Run Nova 2 Lite with reasoning off first. That is the closest like-for-like comparison with Pro.
4. Compare cost per request from usage figures, then repeat with low effort if quality falls short.
5. Check tooling assumptions. High reasoning effort rejects temperature and token-limit parameters, and long reasoning requests need a much longer client timeout.

Regression testing a model upgrade has the full method. If Nova 2 Lite with reasoning off matches Pro on your set, you are paying more today for less context. If it only matches with reasoning on, do the per-task arithmetic before deciding; the cheaper list price may not survive it.