Amazon Nova 2 Lite: The Nova Model AWS Kept Building

The name suggests a small refresh of Nova Lite. It is not. Nova 2 Lite is a different kind of model: a reasoning model with a 1M-token window and built-in tools, priced well above the first-generation Lite and well below the first-generation Premier it is meant to replace for a lot of work.

It also matters for a reason that has nothing to do with benchmarks. In July 2026 Amazon scaled back several Nova models, including Premier, while Nova 2 Lite was among those reported as still in active development. If you are choosing a Nova model to build on today, this is the one with a future attached.

What changed from the first Nova generation

AWS launched Nova 2 Lite on 2 December 2025 at re:Invent, and its Bedrock model card lists it as generally available. The headline differences from Nova Lite 1.0 are all on the spec sheet:

• Context rises from 300K tokens to 1M.
• Maximum output rises from 10K tokens to roughly 64K.
• Extended thinking exists at all. Nova 1 had none.
• Built-in tools for web grounding and code execution, which Nova 1 also lacked.
• Knowledge cutoff moves from October 2024 to October 2025.

Inputs stay the same shape: text, images, video and documents in, text out. That is why AWS describes the move from Nova Lite 1.0 as a direct upgrade. For the full picture of the older model, see the Nova Lite 1.0 guide.

AWS also publishes a handful of scores in its migration guide: 80.9% on MMLU Pro, 70.8% on IF-Bench and 76.0% on tau2-bench Telecom. More striking is the claim that Nova 2 Lite beats Nova Premier on multi-step problem solving at about seven times lower cost and up to five times faster, citing the Nova 2 technical report. Treat those as vendor numbers on vendor harnesses until your own tests agree.

Reasoning is off until you turn it on

Extended thinking is disabled by default. Out of the box, Nova 2 Lite answers like a conventional chat model, which is the right default for classification, extraction and short replies.

You enable it per request through additionalModelRequestFields on the Converse API:

"additionalModelRequestFields": {
  "reasoningConfig": {
    "type": "enabled",
    "maxReasoningEffort": "medium"
  }
}

Effort takes low, medium or high. Two details catch people out. The reasoning text comes back redacted, so you cannot read the chain of thought, but the reasoning tokens are still billed. And the cost is invisible in the response body, so you need the usage figures to see it. The token cost of reasoning models covers why hidden output can dominate a bill.

The parameter trap at high effort

This is the change most likely to break a working integration. With maxReasoningEffort set to high, AWS documents that requests carrying maxTokens, temperature, topP or topK return an error. The fix is to drop the inferenceConfig block entirely for those requests.

If your client library always sends a temperature, high effort will fail until you special-case it. Low and medium do not carry the same restriction in AWS's guidance, which makes them the easier place to start.

Timeouts are the other surprise. AWS says extended thinking requests can run for up to 60 minutes and recommends a read timeout of 3,600 seconds, against the 60-second default in Boto3. A request that times out client-side has still consumed tokens upstream.

Built-in tools, and where they work

Nova 2 Lite ships with two server-side tools. nova_grounding searches public web content and returns citations; AWS lists it as US Regions only, with charges beyond normal inference. nova_code_interpreter runs Python in a sandbox and is available in US East (N. Virginia), US West (Oregon) and Asia Pacific (Tokyo).

Both need the InvokeTool permission added to your IAM policy, because the default Bedrock role does not include it. Ordinary client-side tool calling works as usual. Structured outputs, however, are listed as unsupported on the Bedrock model card, so schema-constrained JSON still means validating and retrying yourself.

Price: five times Nova Lite, and why that can still be fine

At the time of writing OpenRouter lists Nova 2 Lite at $0.30 per million input tokens and $2.50 per million output. Nova Lite 1.0 lists at $0.06 and $0.24. That is five times the input rate and roughly ten times the output rate.

Against Nova Pro 1.0, at $0.80 and $3.20 on the same listing, Nova 2 Lite is cheaper per token in both directions while offering more context and more output. That inversion is why AWS recommends moving Nova Pro workloads here too, even though it reads like a step down in tier.

The honest caveat is reasoning. Every request with thinking enabled spends output tokens you never see, at the output rate. A model that is cheaper per token can still cost more per task if high effort is left on for work that did not need it.

How to test it against what you run now

The migration is mostly a model ID change, so the expensive part is evaluation rather than code. A short plan:

1. Take twenty real requests from production logs, covering your common cases and a few hard ones.
2. Run them on your current Nova model and on Nova 2 Lite with reasoning off, then with low and medium effort.
3. Record cost per request from the usage figures, not the list price, so the hidden reasoning tokens are counted.
4. Only try high effort on the requests that failed at medium, and only after you have removed the sampling parameters.

If reasoning-off quality matches what you have, you get a bigger window for a modest price change. If you need medium effort to match, the per-task cost is the number to compare. How to benchmark LLMs on your own work has the harness, and if you are coming from the top of the old range, the Nova Premier guide explains why that model is no longer a safe target.