Amazon Nova Lite 1.0: Cheap Video and Image Input, Honestly Assessed

Most cheap models are cheap because they do less. Nova Lite 1.0 is an exception worth knowing about: it accepts images and video as well as text, and it lists at a price usually reserved for small text-only models.

It is also nearly two years old, and its successor has shipped. So the useful question is not whether Nova Lite is good in the abstract. It is whether the workloads it still suits are the ones you actually run.

What you get for six cents

Amazon announced Nova Lite at re:Invent on 3 December 2024, and the Bedrock model card gives a launch date of 5 December 2024. The specification, from that card:

• Context window: 300K tokens, which Amazon also expresses as about 30 minutes of video.
• Maximum output: 10K tokens.
• Inputs: text, images and video, plus documents such as PDF, CSV, DOCX and spreadsheets. Output is text only.
• Knowledge cutoff: October 2024.
• Lifecycle: listed as Active on Bedrock at the time of writing.

Some catalogues list Nova Lite as text and image only with a 5K output cap. AWS's own documentation says video in and 10K out, and that is what this guide uses.

On price, OpenRouter lists it at $0.06 per million input tokens and $0.24 per million output at the time of writing. For a model that reads video, that is unusually low. It is the main reason to consider it at all.

Video input, and its limits

Video understanding was rare in December 2024. At launch, Simon Willison noted that Gemini and Nova were the only model families accepting video, and called Nova's version neat but limited compared with Gemini. One limit is structural: Nova models do not process audio, so the soundtrack of a video is ignored.

That shapes what Nova Lite is good for. Visual questions about footage work: what is on screen, whether a product appears, roughly what happens in a clip. Anything that depends on speech, such as summarising a recorded meeting, needs a separate transcription step first.

For still images and document pages the picture is better. Hands-on reports at launch described its image handling as solid, and document Q&A is one of the jobs AWS names for it directly. Vision-language models compared covers what to test before trusting any model with images.

What "equal or better" means in Amazon's comparison

Amazon's launch claims for Nova Lite were specific: equal or better than GPT-4o mini on 17 of 19 benchmarks, than Gemini 1.5 Flash-8B on 17 of 21, and than Claude 3.5 Haiku on 10 of 12.

Read the definition before reading the counts. Amazon counted two models as equal when the 95% confidence intervals of their measured accuracy overlapped. That is a defensible statistical choice, but it means "equal" can include cases where Nova Lite scored lower and the gap was within noise. The claim is that it is competitive with that tier, not that it beats it.

The comparison set is also from late 2024. Those were the cheap models of their day; they are not today's cheap models. Benchmark counts against retired competitors tell you little about where Nova Lite sits now.

Nova Lite or Nova 2 Lite

This is the decision most readers actually face. Nova 2 Lite, launched in December 2025, keeps the same input types but adds a 1M window, around 64K of output, optional reasoning and built-in tools. It is covered in the Nova 2 Lite guide.

It also costs more. OpenRouter lists Nova 2 Lite at $0.30 in and $2.50 out, five times Nova Lite on input and roughly ten times on output. A rule that holds up:

• Stay on Nova Lite when the task is short, perception-heavy and high volume: tagging images, checking a frame for an object, pulling fields from a scanned form. Reasoning adds cost and little else here.
• Move to Nova 2 Lite when inputs exceed 300K tokens, when the answer needs multi-step reasoning, or when you need long outputs.
• Drop to Nova Micro when there are no images at all. The Nova Micro guide covers that tier.

A worked example: tagging a product catalogue

Say you need category and colour tags for 200,000 product photos. Each request carries an image plus a short instruction, and the answer is a few words of JSON.

Image token counts depend on resolution, so measure a sample first rather than guessing. Then multiply: tokens per request, times requests, times the input rate, plus a small output term. With output this short, the input rate dominates, which is exactly where Nova Lite's pricing is strongest.

Two practical points. Bedrock lists structured outputs as unsupported for Nova Lite, so validate every JSON reply and retry the ones that fail to parse. And Bedrock supports batch inference for the Nova 1 models, which suits a job like this that nobody is waiting on.

Before you commit

Nova Lite is a reasonable choice for cheap visual perception at scale, and a weak one for anything needing judgement across a long document. Check three things on your own data: accuracy on fifty labelled images, the parse failure rate of its JSON, and the real token count of your typical image. If those hold up, when a cheap model is enough is the argument for not paying more.