Amazon Nova Micro 1.0: A Text-Only Model Priced for Volume

Not every model call needs intelligence. A large share of production traffic is routing, tagging, short summaries and format conversion, where the right model is the cheapest one that clears an accuracy bar. Nova Micro exists for that traffic.

It is text only, it has the smallest window in the first Nova generation, and Amazon positions it on latency and cost rather than capability. That is a narrow brief, and judged against it the model is easy to evaluate.

The job Micro is built for

Amazon announced Nova Micro on 3 December 2024 alongside Nova Lite and Nova Pro, and the Bedrock model card dates its launch to 5 December 2024. AWS describes it as its fastest text-only model, aimed at summarisation, translation and classification.

The specification from AWS documentation:

• Context window: 128K tokens.
• Maximum output: 10K tokens. Some catalogues show 5K; AWS lists 10K.
• Inputs: text only. No images, no video, and unlike Lite and Pro, no document attachments.
• Knowledge cutoff: October 2024.
• Languages: over 200 supported, with 15 that AWS says it optimised for, including English, German, Spanish, Japanese, Arabic and Hindi.

Client-side tool calling is supported. Structured outputs are listed as unsupported on Bedrock, which matters for a model whose natural job is producing labels and small JSON objects.

The arithmetic at a million calls

OpenRouter lists Nova Micro at $0.035 per million input tokens and $0.14 per million output at the time of writing. Per call those numbers are meaningless, so put them against a real workload.

Take a support-ticket classifier: one million tickets a month, about 500 input tokens each including the instructions, and a ten-token label back.

• Input: 500 million tokens at $0.035 per million is $17.50.
• Output: 10 million tokens at $0.14 per million is $1.40.
• Total: about $19 a month.

The same job on Nova Lite 1.0, at $0.06 and $0.24, comes to about $32. On Nova 2 Lite with reasoning off, at $0.30 and $2.50, it is about $175. None of these is large, but the ratio is what matters once volume grows or once the classifier is one stage of many.

Note where the money goes: input is over 90% of the bill. Shortening the instructions saves more than switching models would. LLM classification covers why a tighter taxonomy usually beats a longer prompt.

Speed: the 210 tokens per second claim

Amazon's launch announcement gave Nova Micro an output speed of 210 tokens per second. Third-party trackers have reported different figures since, some higher, which is normal: throughput depends on region, load and request shape.

For a classifier, output speed barely matters, because the output is ten tokens. What you feel is time to first token and network round trip. If latency is the reason you are looking at Micro, measure end-to-end time from your own region with your real prompt length, not tokens per second from a chart.

Micro or Lite

The two first-generation siblings split cleanly. Pick Micro when every input is text, fits in 128K and the task is short. Pick Nova Lite when any input is an image, a video or an attached document, or when you need up to 300K of context. The Nova Lite guide covers that model's video limits.

Amazon's own benchmark claims place Micro against small models of its era: equal or better than Llama 3.1 8B on all 11 applicable benchmarks and Gemini 1.5 Flash-8B on all 12, where "equal" means overlapping confidence intervals. That is the right comparison class. Do not expect it to reason like a large model.

Where it falls over

Small models fail in recognisable ways, and Micro is no exception. Expect trouble with:

• Multi-step reasoning, where an early mistake carries through to a confident wrong answer.
• Long instructions with many conditions, where later rules get dropped.
• Ambiguous categories that a human would also argue about.
• Strict output formats without validation, since there is no structured output mode to lean on.

The fix is architectural rather than a better prompt. Send everything to Micro first, check the result cheaply, and escalate the failures to a larger model. Model routing and fallbacks shows how to set the escalation threshold so the cheap model handles the bulk without silently degrading quality.

Fine-tuning instead of prompting

AWS lists fine-tuning support for Nova Micro on Bedrock, and the Nova family was designed to be a distillation target, with larger Nova models generating training data for smaller ones. For a narrow, stable task with a few thousand labelled examples, a tuned small model can match a much larger prompted one at a fraction of the running cost.

Whether that is worth the effort depends on volume and on how often the task changes. A classifier whose categories change monthly is a poor candidate; one that has been stable for a year is a good one. Fine-tuning vs prompting vs RAG lays out the trade.

The takeaway: Micro is a component, not an assistant. Use it where the task is narrow, the volume is high and you can check its answers. Measure accuracy on a few hundred labelled examples, and if it clears your bar, you have found the cheapest reliable stage in your pipeline.