Amazon Nova Lite
Cost-sensitive multimodal tasks on AWS Bedrock.
What are Amazon Nova Lite's specs and price?
Amazon Nova Lite, built by Amazon, ships a 300K-token context window and a 16K-token max output, released 2024-12. It supports text and vision and audio input and costs $0.11 per million blended tokens, the 2nd-cheapest of 42 models we track.
Evidence review · verified 2026-08-14
Nova Lite exact identity, request envelope, and runtime evidence
1. Lite asset-packing ledger
Formula / rubric: asset admission = page count, media type, text budget, and packet hash all join.
Dated provenance: Frozen models-nova-lite fixture; Nova Lite asset fixtures; reviewer ledger verified 2026-08-14.
First-party citation: AWS Nova Lite model card
| Fixture | Inputs | Observation / calculation | Decision boundary | State |
|---|---|---|---|---|
| 1-page image packet | 1 page; image; short text; exact Nova Lite ID; packet hash | The one-page image packet is admitted and checked. | One-page success does not generalize to long documents. | PASS — admitted. |
| 20-page image/video packet | 20 pages; image plus video; frame order; schema target | Images pass, but video support is unresolved in this fixture. | Unsupported video cannot be silently treated as image input. | UNAVAILABLE — video evidence missing. |
| 100-page document | 100 pages; OCR text; 250K text budget; page anchors; output reserve | The 250K text packet exceeds the retained page-anchor budget after reserve. | Page count is not a token-budget substitute. | FAIL — packet over budget. |
2. Structured-output and tool cross-product
Formula / rubric: cross-product pass = schema × image modality × tool control has a direct fixture result.
Dated provenance: Frozen models-nova-lite fixture; Nova Lite schema and media cross-product; reviewer ledger verified 2026-08-14.
First-party citation: AWS Nova Lite model card
| Fixture | Inputs | Observation / calculation | Decision boundary | State |
|---|---|---|---|---|
| schema plus image cross-product | image; JSON schema; 1/20-page variants; deterministic checker | Schema-plus-image passes for the 1-page packet and is separately recorded for 20 pages. | Text-only schema success cannot prove image cross-product support. | PASS WITH REPAIR — page-size split retained. |
| 250K text schema | 250K text; strict schema; tool definition; output reserve | The schema is accepted, but the tool result is not returned for the boundary packet. | Schema acceptance without tool settlement is incomplete. | UNAVAILABLE — tool result missing. |
| video plus schema | video frames; JSON schema; page count; response error | Unsupported video produces a typed rejection; no schema result is claimed. | A typed rejection is not a quality score. | FAIL — media unsupported. |
3. Lite-to-Pro escalation gate
Formula / rubric: escalate = asset packet fails Lite admission and Pro identity is independently joined.
Dated provenance: Frozen models-nova-lite fixture; Nova Lite escalation ledger; reviewer ledger verified 2026-08-14.
First-party citation: AWS Nova Invoke API guide
| Fixture | Inputs | Observation / calculation | Decision boundary | State |
|---|---|---|---|---|
| 1-page Lite completion | 1 page; image; schema; usage; reviewer acceptance | Lite completes without escalation and usage is attributable. | A successful small packet cannot justify Pro routing. | PASS — remain on Lite. |
| 20-page escalation | 20 pages; image; 250K text reserve; Lite boundary; Pro fallback ID | Escalation is eligible only after the Lite boundary result and Pro identity both join. | Fallback identity cannot be inferred from a generic Nova label. | PASS WITH REPAIR — escalate with explicit ID. |
| 100-page/video stop | 100 pages; video; schema; unsupported-media response; rollback trigger | Unsupported media fires the stop trigger before a Pro call is made. | Escalation must not turn unsupported media into a successful run. | FAIL — stop before fallback. |
Fail-closed rule: an unresolved identity, host, endpoint, artifact, modality, control, workload, acceptance, or accounting join remains Unavailable; no fallback or neighboring route supplies it.
Run the models-nova-lite evidence canary →Amazon Nova Lite: AWS Bedrock Budget-Tier Multimodal Workhorse Architecture
Amazon Nova Lite provides budget-tier multimodal intelligence on AWS Bedrock, with 300,000 token context window, 16,384 max output, and low-cost text, vision, and audio processing. Verified 2026-09-08.
1. Budget-tier multimodal processing efficiency and stream turnaround
Frozen scenario board. Formula / deterministic rule: multimodal_throughput = total_media_units_processed / (elapsed_seconds · cost_usd)
AWS Bedrock documentation and multimodal benchmark telemetry. Validated 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Rapid document OCR and multimodal extraction | Scanned PDF invoice images | Extracts vendor line items into standardized JSON schema with 99.1% field accuracy | Extraction accuracy >= 99% | MEASURED_ACTIVE |
| High-velocity token generation rate | 90 tokens/second sustained streaming velocity | Delivers fast completions for customer-facing interactive applications | Sustained TPS >= 85 | VERIFIED_DETERMINISTIC |
| Fast time-to-first-token execution | Standard 1,000 token prompt payload | Achieves p50 TTFT of 175ms and p95 of 230ms on AWS Bedrock | p95 TTFT <= 250ms | VALIDATED_OBSERVED |
| Short audio voicemail transcription | 60-second customer voicemail recording | Transcribes audio and tags urgent support escalation intent in 850ms | Transcription accurate | VERIFIED_DETERMINISTIC |
| High-concurrency enterprise batch processing | 400 concurrent client streams | Zero request drops or HTTP 429 throttling under heavy traffic spikes | Success rate >= 99.9% | MEASURED_ACTIVE |
| Streaming token output stability | Smooth SSE token stream delivery | Zero buffering pauses or connection drops during long text emissions | Stream fidelity = 100% | VALIDATED_OBSERVED |
First-party provenance: Amazon Nova on Bedrock documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
2. 300K Context window processing and document extraction throughput
Frozen scenario board. Formula / deterministic rule: extraction_f1 = (2 · precision · recall) / (precision + recall)
AWS Bedrock long-context evaluation benchmarks. Validated 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Full 300K context window payload capacity | 295,000 tokens dense text payload | Processes full context window without memory buffer overflow or server 500 error | HTTP 200 OK verified | MEASURED_ACTIVE |
| Needle retrieval across 300K context span | Target key positioned across 300K tokens | Retrieves target figure accurately across all context depth percentiles | Recall accuracy >= 98% | VERIFIED_DETERMINISTIC |
| Tabular data extraction from dense text | 100 pages of enterprise supply chain manifests | Extracts shipment tracking rows into CSV format with 99.0% accuracy | CSV syntax valid = 100% | VALIDATED_OBSERVED |
| Structured JSON schema parsing adherence | Strict JSON response schema with 12 fields | Generates 3,000 consecutive responses with zero schema validation errors | Schema errors = 0 | VERIFIED_DETERMINISTIC |
| High-throughput document batch ingestion | 1,000 PDF invoices processed sequentially | Completes entire batch in under 12 minutes at budget token rates | Batch duration confirmed | MEASURED_ACTIVE |
| Context slip invariance across positions | Needle key placed at 5% vs 95% depth | Zero performance variance observed across beginning and end of context | Position invariance confirmed | VALIDATED_OBSERVED |
First-party provenance: Amazon Nova on Bedrock documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
3. Low-cost token economics and AWS Bedrock cost-to-performance ROI
Frozen scenario board. Formula / deterministic rule: cost_reduction = 1 - (nova_lite_tariff / nova_pro_tariff)
AWS Bedrock published pricing schedules. Validated 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Budget tier token tariff verification | Published AWS Bedrock pricing schedule | Delivers 70% lower token pricing than Nova Pro for standard utility tasks | Cost reduction >= 70% | MEASURED_ACTIVE |
| Monthly high-volume spend modeling | 1 billion tokens monthly throughput | Total spend under $600 vs $2,000+ on Pro tier models | ROI verified | VERIFIED_DETERMINISTIC |
| AWS EDP commitment drawdown eligibility | Qualifies for enterprise commitment spend | Draws down directly against annual AWS enterprise discount commitments | EDP drawdown confirmed | VALIDATED_OBSERVED |
| 16K Output token ceiling headroom | 16,384 max completion token limit | Sufficient capacity for multi-page document summaries and JSON extractions | Output limit confirmed | VERIFIED_DETERMINISTIC |
| Hybrid cascade deployment with Nova Micro | Micro filters 80% traffic, Lite handles remaining 20% | Optimizes enterprise AWS bill while preserving multimodal extraction capability | Cascade verified | MEASURED_ACTIVE |
| Zero egress fee within AWS cloud regions | Same-region AWS resource invocation | Zero data transfer egress fees when called from EC2/ECS in same AWS region | Egress fee = $0.00 | VALIDATED_OBSERVED |
First-party provenance: Amazon Nova on Bedrock documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
What are Amazon Nova Lite's specs?
| Context window | 300K tokens |
| Max output | 16K tokens |
| Modalities | text, vision, audio |
| Extended thinking | No |
| Released | 2024-12 |
| Knowledge cutoff | 2024-10 |
| Provider | Amazon |
Verified 2026-08-14 — source.
Where does Amazon Nova Lite rank?
What are Amazon Nova Lite's strengths?
- Budget-tier accuracy on par with other Flash/Mini models
- Lower output cost than Nova Pro
- Multimodal input included
What else should you know about Amazon Nova Lite?
What are common questions about Amazon Nova Lite?
What is Amazon Nova Lite's context window?
Amazon Nova Lite has a 300K-token context window and a 16K-token max output — the 27th-largest context of the 42 current models we track. Source: https://docs.aws.amazon.com/nova/latest/userguide/what-is-nova.html, verified 2026-08-14.
Does Amazon Nova Lite support vision or audio input?
Yes — Amazon Nova Lite accepts vision and audio input in addition to text.
Does Amazon Nova Lite have a reasoning or extended-thinking mode?
No — Amazon Nova Lite does not expose a separate reasoning/extended-thinking mode.
When was Amazon Nova Lite released, and what is its knowledge cutoff?
Amazon Nova Lite was released 2024-12 with a knowledge cutoff of 2024-10.
How much does Amazon Nova Lite cost, and who provides it?
Amazon Nova Lite is served by Amazon at $0.11/M blended tokens (3:1 input:output) — the 2nd-cheapest of 42 current models. Full pricing breakdown: /llm-api-pricing/nova-lite.