TheRouter.aiTheRouter.ai
Back to Models

Amazon Nova 2 Lite

amazonamazon/nova-2-lite

Nova 2 Lite is an advanced multimodal reasoning model with 1M context. Dynamically adjusts reasoning depth. Extended thinking on complex problems.

AWS announcement β†—Nova product page β†—Nova 2 technical report (PDF) β†—

Amazon Nova 2 Lite is the small, agentic reasoning model in Amazon's Nova 2 family, announced in early December 2025 and served through Amazon Bedrock and SageMaker AI. It accepts text, images, documents, and video, with a 1 million-token context window β€” large enough to take a full codebase or a long-form video transcript as input in a single call.

From a TheRouter operator's perspective, Nova 2 Lite is what you reach for when you need a million-token context at mini-tier price. Extended thinking is off by default β€” Amazon ships it with three opt-in budget levels (low / medium / high) so you choose the speed-cost-quality trade-off explicitly. Web grounding and a code interpreter are first-class runtime tools, useful when you want a single model call to actually do the work, not just suggest it.

Best for
  • β€’ Million-token context jobs on a budget β€” full codebase Q&A, long-form video summarisation, multi-document RAG without per-chunk routing
  • β€’ Agentic workflows that need web grounding or a code interpreter without wiring external tools
  • β€’ AWS-bound enterprise workloads that already use Bedrock β€” VPC isolation, SageMaker fine-tuning, and AWS-grade SLAs come for free
  • β€’ Multimodal extraction over PDFs and short video clips where Nova Pro is overkill
Reach for something else if
  • β€’ Frontier reasoning or coding benchmarks β€” route to Claude Opus 4.7 or DeepSeek V3.2 if you need top-of-leaderboard performance
  • β€’ Native audio in/out workflows β€” Nova 2 Lite does not expose audio as a native modality; use a dedicated speech model
  • β€’ Multi-region routing latency demands β€” Bedrock pricing carries a ~10% cross-region surcharge when using global endpoints
Context Length
1M
Max Output
65K
Input Priceper 1M tokens
$0.180/ 1M tokens
Output Priceper 1M tokens
$0.720/ 1M tokens

Modalities

textimage→text

Pricing Breakdown

TypeRate
Input$0.180 / 1M tokens
Output$0.720 / 1M tokens

Supported Parameters

temperaturemax_tokenstop_ptoolstool_choiceresponse_formatreasoningstop

Specifications

Release windowEarly December 2025 (Nova 2 family launch)aws.amazon.com β†—verified
Bedrock model idglobal.amazon.nova-2-lite-v1:0aws.amazon.com β†—verified
Training cutoffNot publicly disclosedunknown
Extended thinking budget levelsoff (default), low, medium, highaws.amazon.com β†—verified
Built-in toolsWeb grounding (live HTTP with citations), code interpreteraws.amazon.com β†—verified
CustomizationFine-tuning available on Amazon Bedrock and Amazon SageMaker AIaws.amazon.com β†—verified
Bedrock tiersStandard, Priority (+75% premium), Flex (-50% discount)aws.amazon.com β†—to verify
LicenseAWS Service Terms (proprietary, API-only via Bedrock/SageMaker)verified

Benchmarks

BenchmarkDistributionScoreSource
MMLU-Pro
Per third-party analysis; AWS has not published an official MMLU-Pro number for Nova 2 Lite.
~80.9medium.com β†—
GPQA Diamond
Third-party reported; AWS has not published an official GPQA Diamond number.
~79.6medium.com β†—
Artificial Analysis Intelligence Index
Composite score across MMLU-Pro, GPQA Diamond, HLE, LiveCodeBench, SciCode, AIME 2025, IFBench, AA-LCR, Terminal-Bench Hard, τ²-Bench Telecom.
Comparable to gpt-oss-120B and Qwen3 235Bmedium.com β†—
Long-context benchmarks (LongCodeBench 1M, OpenAI-MRCR)
Amazon's Nova 2 technical report says it evaluates long-context on LongCodeBench 1M and OpenAI-MRCR, but per-model numeric scores for Nova 2 Lite were not surfaced in the public release.
β€”Not publicly disclosedβ€”

API Usage Examples

Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.

cURL
curl https://api.therouter.ai/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $THE_ROUTER_API_KEY"   -d '{
    "model": "amazon/nova-2-lite",
    "messages": [
      {"role": "user", "content": "Summarize the key points from this input."}
    ]
  }'

API guide

Full API reference β†’

Chat completion

Call Nova 2 Lite through TheRouter's OpenAI-compatible surface. TheRouter handles the Bedrock InvokeModel / Converse translation upstream β€” your client code stays OpenAI-shaped.

cURL
curl https://api.therouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "amazon/nova-2-lite",
    "messages": [{"role": "user", "content": "Summarise this 200-page PDF in 5 bullets."}]
  }'

More from amazon

amazon/nova-proNova flagship β€” higher accuracy, higher cost; route here when Lite hits its ceiling.amazon/nova-premierTop of the Nova family β€” for the most demanding agentic and reasoning workloads on Bedrock.amazon/nova-litePrevious-generation Lite β€” smaller context and no extended-thinking budgets; useful for cost A/B.amazon/nova-microCheapest, text-only Nova β€” for high-volume classification, routing, and short-form completion.

Similar models

Cross-provider sibling models
anthropic/claude-haiku-4.5
Comparable mini-tier reasoning model on Bedrock with similar price profile; benchmarks reportedly close.
google/gemini-2.5-flash
Google's fast, cheap, long-context option β€” direct competitor on the cost / context-length axis.
openai/gpt-5-mini
OpenAI's mini tier β€” different tool surface and pricing structure; useful as the third corner of a routing triangle.
deepseek/deepseek-v3.2
Open-weight alternative for teams that want a self-host fallback alongside Bedrock-hosted Nova.

News & changes

2025-12-02

AWS introduces Nova 2 Lite β€” 1M context, extended thinking, web grounding built in

AWS announced Nova 2 Lite as the mini-tier of the Nova 2 family, available through Bedrock and SageMaker AI. For TheRouter operators the practical hooks are the 1M context window at mini-tier pricing, opt-in extended thinking budgets (low / medium / high), and first-class web grounding plus code interpreter as runtime tools β€” useful when one model call needs to do the work, not just describe it.

re-authored by TheRouteraws.amazon.com/blogs/aws β†—
2025-12-02

Amazon publishes Nova 2 technical report

Amazon's Nova 2 technical report covers evaluation methodology (MMLU-Pro, GPQA-Diamond, AIME 2025, IF-Bench, MultiChallenge, LongCodeBench 1M, OpenAI-MRCR) plus safety evaluation under Amazon's Frontier Model Safety Framework. Per-model numeric scores for Nova 2 Lite are sparser than community trackers would like β€” treat third-party benchmarks (low confidence) accordingly.

re-authored by TheRouterassets.amazon.science β†—

Frequently asked

What is the context window of Nova 2 Lite?

1 million tokens. Large enough for whole-codebase Q&A, long-form video transcripts, or multi-document RAG without chunked routing.

aws.amazon.com β†—
Does Nova 2 Lite support extended reasoning?

Yes, but off by default. Enable via the reasoning field and pick a budget β€” low, medium, or high β€” to trade speed and cost for depth. AWS recommends keeping it off until a task actually needs it.

aws.amazon.com β†—
What tools come built in?

Two: web grounding (live HTTP queries with citations) and a code interpreter. Both are managed by Bedrock as runtime tools, separate from any function-calling tools you declare in the request.

aws.amazon.com β†—
Can I fine-tune Nova 2 Lite?

Yes β€” fine-tuning is offered on both Amazon Bedrock and Amazon SageMaker AI. Useful when prompt engineering hits a ceiling on a narrow vertical workflow.

aws.amazon.com β†—
How does Nova 2 Lite pricing compare across Bedrock tiers?

Bedrock offers three tiers: Standard (baseline), Priority (+75% premium for preferential compute), Flex (-50% discount for workloads that tolerate higher queue time). TheRouter routes Standard-tier traffic by default; ask if you want a Flex-only or Priority-only routing policy.

aws.amazon.com β†—
Fact ledger β€” every claim on this page traces here
sourceURLretrieved
Release windowaws.amazon.com β†—2026-05-22verified
Bedrock model idaws.amazon.com β†—2026-05-22verified
Training cutoffβ€”β€”unknown
Extended thinking budget levelsaws.amazon.com β†—2026-05-22verified
Built-in toolsaws.amazon.com β†—2026-05-22verified
Customizationaws.amazon.com β†—2026-05-22verified
Bedrock tiersaws.amazon.com β†—2026-05-22to verify
Licenseβ€”β€”verified
MMLU-Promedium.com β†—2026-05-22single source
GPQA Diamondmedium.com β†—2026-05-22single source
Artificial Analysis Intelligence Indexmedium.com β†—2026-05-22single source
Long-context benchmarks (LongCodeBench 1M, OpenAI-MRCR)assets.amazon.science β†—2026-05-22unknown
AWS introduces Nova 2 Lite β€” 1M context, extended thinking, web grounding built inaws.amazon.com/blogs/aws β†—2026-05-22verified
Amazon publishes Nova 2 technical reportassets.amazon.science β†—2026-05-22verified
What is the context window of Nova 2 Lite?aws.amazon.com β†—2026-05-22to verify
Does Nova 2 Lite support extended reasoning?aws.amazon.com β†—2026-05-22to verify
What tools come built in?aws.amazon.com β†—2026-05-22to verify
Can I fine-tune Nova 2 Lite?aws.amazon.com β†—2026-05-22to verify
How does Nova 2 Lite pricing compare across Bedrock tiers?aws.amazon.com β†—2026-05-22to verify
Customer Support