← Agent Frameworks 🕐 16 min read
Agent Frameworks

Amazon Nova Model Family — Cost Positioning and FinOps Patterns (2026)

Executive Summary · March 2026

Amazon Nova is AWS's native model family on Bedrock, positioned to undercut third-party models (Claude, GPT-4o, Llama via Bedrock) on price while staying within the same capability tier.

Amazon Nova is AWS’s native model family on Bedrock, positioned to undercut third-party models (Claude, GPT-4o, Llama via Bedrock) on price while staying within the same capability tier. The family now spans two generations (Nova 1 launched December 2024, Nova 2 launched December 2025) and five capability tiers from text-only Micro at $0.035/1M input tokens up to multimodal Premier at $2.50/1M input tokens. For FinOps teams the critical distinction is that Nova charges under AmazonBedrock product codes while Anthropic Claude routes through AWS Marketplace — a billing split that breaks naive cost queries if not accounted for.

1. Nova Generation Overview

Nova 1 (December 2024 GA)

Model Context Input modalities Output modalities
Nova Micro 128K Text only Text
Nova Lite 300K Text, Image, Video Text
Nova Pro 300K Text, Image, Video Text
Nova Premier 1M Text, Image, Video Text

Nova 2 (December 2025 GA)

Model Context Max output Notes
Nova 2 Lite 1M 64K Reasoning mode (low/medium/high/off), 1 FPS video sampling
Nova 2 Pro (Preview) 1M 64K Early access via Nova Forge; best distillation teacher

Nova 2 Lite is the primary upgrade recommendation for production pipelines already using Nova 1 Lite — same pricing tier, 3× context window, native reasoning toggle, and measurable throughput gains (Securonix reported 33% faster response times in internal benchmarks).

Model IDs (Bedrock canonical format):

amazon.nova-micro-v1:0
amazon.nova-lite-v1:0
amazon.nova-pro-v1:0
amazon.nova-premier-v1:0
amazon.nova-2-lite-v1:0
amazon.nova-2-pro-v1:0        # preview; Nova Forge access required

Cross-region inference profile IDs use a geographic prefix:

us.amazon.nova-lite-v1:0      # routes within US regions
eu.amazon.nova-lite-v1:0      # routes within EU regions

2. Pricing — Full Detail

All prices are on-demand Standard tier, per 1M tokens unless noted. Verified April–June 2026.

2.1 Text Token Pricing

Model Input Output Cache read Cache write Batch (50% off)
Nova Micro $0.035 $0.140 $0.00875 — $0.0175 / $0.070
Nova Lite (v1) $0.060 $0.240 — — $0.030 / $0.120
Nova 2 Lite $0.300 $2.500 — — $0.150 / $1.250
Nova Pro $0.800 $3.200 $0.200 — $0.400 / $1.600
Nova Premier $2.500 $12.500 — — $1.250 / $6.250

Notes:

  • Nova Micro prompt caching is priced at $0.00875/1M (75% discount vs input). Cache window is 5 minutes.
  • Nova Pro cache read is $0.20/1M (75% discount, consistent with Bedrock caching standard).
  • Nova 2 Lite pricing reflects its reasoning model positioning — output is 8.3× more expensive than input at standard tier, similar to Claude’s reasoning tier ratio. Route high-output reasoning workloads through Flex tier for 50% savings when latency allows.

2.2 Service Tier Multipliers

Bedrock now supports four service tiers for Nova models:

Tier Relative cost Latency SLA Workload fit
Standard 1× Best-effort Interactive apps
Priority ~1.3× (estimated) Higher priority during congestion SLA-sensitive production
Flex 0.5× (50% discount) Higher latency; lower priority Batch jobs, evals, fine-tune prep
Batch 0.5× (50% discount) Async, non-interactive Bulk document processing

Flex tier was introduced November 2025. Use it for: model evaluations, content summarization pipelines, data labeling, multi-step agentic workflows that tolerate queue depth. Batch tier requires an async submit/retrieve pattern via the Bedrock Batch API.

2.3 Multimodal Pricing — Images

Nova Lite and Pro handle images at a flat billing rate regardless of resolution:

  • 230 tokens per image — billed at the model’s standard input token rate
  • This is reflected in the API’s inputTokens response count

Effective image cost by model:

Model Tokens per image Input $/1M Cost per image
Nova Lite 230 $0.060 ~$0.0000138
Nova Pro 230 $0.800 ~$0.000184
Nova Premier 230 $2.500 ~$0.000575

Images are auto-rescaled server-side: shortest side ≥ 896px, up to 8000×8000. Aspect ratio is preserved. For document OCR pipelines, high-resolution scans do not increase the token count — the flat 230 token rate is resolution-invariant.

Up to 5 images per request (inline). Up to 1,000 images per request via S3 URI reference (combined ≤ 2 GB).

2.4 Multimodal Pricing — Video

Video billing is token-based, calculated from actual frames sampled:

  • Sampling rate: 1 frame per second (FPS) for videos ≤ 16 minutes
  • For videos > 16 minutes: sampling rate decreases to maintain max 960 frames
  • Each frame = approximately 288 tokens

Approximate token cost by video length (Nova 2 Lite):

Video length Frames sampled ~Tokens Cost at $0.30/1M
10 seconds 10 2,880 $0.000864
30 seconds 30 8,640 $0.002592
5 minutes 300 86,400 $0.025920
16 minutes 960 276,480 $0.082944
30 minutes 1,800 518,400 $0.155520

At Nova Pro rates ($0.80/1M), multiply by ~2.67×. At Nova Premier ($2.50/1M), multiply by ~8.3×.

Key caveat: Audio is NOT processed. Nova video understanding is purely visual frame analysis. Speech-to-text tasks should route to Nova Sonic ($0.0034/1K speech input tokens) rather than Nova Pro video.

PDF document pricing: ~2,560 input tokens per standard 8.5×11 page. A 50-page contract ≈ 128,000 tokens ≈ $0.10 at Nova Pro rates.

3. Nova vs. Claude vs. Llama — Cost-Per-Capability Comparison

3.1 Pricing Matrix (June 2026 On-Demand Standard)

Model Input $/1M Output $/1M Context Multimodal
Nova Micro $0.035 $0.14 128K Text only
Nova Lite $0.060 $0.24 300K Image, video
Nova 2 Lite $0.300 $2.50 1M Image, video
Nova Pro $0.800 $3.20 300K Image, video
Nova Premier $2.50 $12.50 1M Image, video
Claude Haiku 3.5 $0.80 $4.00 200K Image
Claude Sonnet 4.x $3.00 $15.00 200K Image
GPT-4o $2.50 $10.00 128K Image
GPT-4o mini $0.15 $0.60 128K Image
Gemini 2.5 Flash $0.15 $0.60 1M Image, video
Gemini 2.5 Pro $2.50 $15.00 1M Image, video
Llama 4 Scout (Bedrock) ~$0.17 ~$0.65 128K Image

3.2 Task-Type Routing Decision Matrix

Use Nova Micro when:

  • Classification, routing, intent detection
  • Structured extraction from short inputs (< 4K tokens)
  • High-volume triage (millions of requests/day)
  • Cost ceiling: < $0.035/1M input is the floor on Bedrock for text tasks
  • Nova Micro consistently outperforms Llama 3.1 8B and Gemini 1.5 Flash-8B at equivalent cost

Use Nova Lite when:

  • Q&A on documents with visual content (charts, tables)
  • Multimodal inputs where video frame analysis is needed
  • Long-context retrieval up to 300K (or 1M with Nova 2 Lite)
  • Alternatives (GPT-4o mini at $0.15/1M, Gemini 2.5 Flash at $0.15/1M) are cheaper for text-only tasks but lack Bedrock-native video support

Use Nova Pro when:

  • Complex reasoning over long documents
  • Code generation, multi-step tool use, agentic workflows inside Bedrock
  • Priced identically to Claude Haiku 3.5 ($0.80 input) but with 300K context vs. 200K
  • Nova Pro runs at ~200 tokens/second throughput vs. Claude’s ~85–90 tokens/second — meaningful for latency-sensitive agentic loops
  • 44% cheaper than GPT-4o for comparable summarization tasks in documented benchmarks

Use Nova Premier when:

  • 1M-token context is required (very long contract review, multi-document analysis)
  • Priced identically to Gemini 2.5 Pro on input ($2.50/1M) but 16% cheaper on output ($12.50 vs. $15.00)
  • Inferior to Claude Sonnet / Opus on complex reasoning benchmarks — Artificial Analysis Intelligence Index scores Nova Premier at 13/100 vs. ~17+ for comparable models; use only when the 1M context window is the actual requirement

Avoid Nova Premier for:

  • Tasks that need strong coding or reasoning — Claude Sonnet 4.x or Gemini 2.5 Pro outperform it meaningfully on MMLU, HumanEval, and SWE-bench at similar price points
  • Nova Premier is slow (32.7 tokens/second; ~59th percentile) — latency-sensitive pipelines should use Nova Pro or Nova 2 Lite instead

3.3 Nova vs. GPT-4o-mini / Gemini Flash for Enterprise Evaluation

For the common enterprise evaluation brief: “find the cheapest model that can handle customer-facing Q&A with occasional image inputs”:

Dimension Nova Lite GPT-4o mini Gemini 2.5 Flash
Input $/1M $0.060 $0.150 $0.150
Output $/1M $0.240 $0.600 $0.600
Context 300K 128K 1M
Video input Yes No Yes
Bedrock-native Yes No (via API GW) No (via API GW)
CUR product code AmazonBedrock Marketplace Marketplace

Nova Lite wins on price for text-heavy, image-occasionally workloads when the team is already on AWS. Gemini 2.5 Flash wins on context and equals Nova Lite on benchmark performance but requires a separate GCP billing relationship or Vertex API routing layer.

4. Nova Micro — Use Cases Where It Outperforms Its Cost Tier

Nova Micro at $0.035/$0.14 per 1M tokens is the cheapest text model available on Bedrock. It is text-only with a 128K context window.

Documented outperformance cases:

  1. Intent classification / request routing — At sub-millisecond-equivalent cost per request, Micro can act as the router in a cascade (Micro classifies → Lite/Pro handles complex cases). A 10M request/day router costs $350/day at Nova Micro vs. $600/day at Nova Lite.

  2. Structured extraction from dense tables — Micro handles well-formatted tabular input (CSV, JSON schema-prompted) with accuracy competitive with Llama 3.1 8B. For invoice line-item extraction where inputs are predictably structured, Micro is the default choice.

  3. Prompt caching anchor — Micro is the only Nova model with documented cache pricing ($0.00875/1M cache reads). For systems with stable system prompts and high query volume, the effective cost with caching approaches $0.01/1M on repeat reads.

  4. Guardrails + moderation — Pair Micro with Bedrock Guardrails for content moderation pipelines. Guardrails adds $0.15 per 1,000 text units; at Micro’s base cost this overhead is still cheaper than Nova Lite baseline for the same task.

  5. Agent orchestration metadata — In multi-step agentic workflows, internal “thinking” steps (decide next action, summarize tool output, select next tool) can route to Micro while Nova Pro handles the reasoning-heavy steps, cutting total agent token cost by 30–40%.

5. High-Throughput Workloads: Quotas, Cross-Region, Flex Tier

5.1 Default Service Quotas (On-Demand Standard Tier)

Default quotas are per-region, per-account, adjustable via Service Quotas console:

Model Requests/min (in-region) Tokens/min (in-region)
Nova Micro ~1,000 ~2,000,000
Nova Lite ~1,000 ~2,000,000
Nova Pro ~600 ~1,200,000
Nova Premier ~200 ~600,000

Note: Quotas have been subject to reductions in specific regions (confirmed Nova Lite reports from re:Post of sudden drops to 100 RPM). Always check current limits in Service Quotas console before sizing a production deployment.

5.2 Cross-Region Inference

Cross-region inference profiles route requests across multiple AWS regions within a geography (US, EU, APAC), effectively pooling quota:

  • Throughput benefit: Up to 2× the in-region quota by aggregating capacity across multiple regions
  • Cost: Same per-token price as in-region Standard tier (no cross-region surcharge for Nova models via geographic profiles)
  • CUR distinction: Cross-region requests appear with suffix -cross-region-global or -cross-region-{geo} in line_item_usage_type

Profile ID format:

us.amazon.nova-pro-v1:0        # US geographic profile
eu.amazon.nova-lite-v1:0       # EU geographic profile

Knowledge Bases supports cross-region inference for Nova models natively (enabled September 2024).

5.3 Flex Tier for High-Volume Non-Interactive Jobs

Flex tier provides 50% cost reduction in exchange for lower priority. Configuration is a parameter on the API call, not a different model ID:

response = bedrock_runtime.invoke_model(
    modelId="amazon.nova-pro-v1:0",
    body=json.dumps({"messages": [...]}),
    performanceConfig={"latency": "optimized"}   # Standard
    # performanceConfig={"latency": "standard"}  # Flex equivalent
)

Use Flex for: batch document processing, fine-tune dataset preparation, offline evals, RAG index generation, content summarization queues where SLA is hours not seconds.

6. Nova Fine-Tuning — Cheapest Native SFT on Bedrock

6.1 Pricing

Nova models support supervised fine-tuning (SFT) natively on Bedrock via Nova Forge. Fine-tuning token pricing as of Q1 2026:

Phase Nova Pro SFT cost
Training tokens $0.008 per 1,000 tokens ($8.00/1M)
Custom model storage $1.95/month per model
Inference (post-SFT) Requires Provisioned Throughput; cannot run on-demand

Practical cost examples:

  • 1M token training run: $8.00
  • 5M token training run: $40.00
  • 10M token fine-tune (quality baseline for most domain adaptation): $80.00

6.2 Comparison with Alternatives

Provider Model SFT cost/1M train tokens On-demand inference post-SFT
Amazon Bedrock Nova Pro $8.00 No (Provisioned Throughput required)
Amazon Bedrock Llama 3.x $8.00 No
OpenAI GPT-4o mini $25.00 Yes (on-demand)
Google Vertex Gemini Flash $2.00 Yes (on-demand)
Together.ai Llama 3.1 70B ~$3.00 Yes

Nova SFT is 3× cheaper than OpenAI’s fine-tuning but more expensive than Vertex/Together options. The critical operational constraint: all Bedrock fine-tuned models require Provisioned Throughput, which bills hourly (model unit hours) regardless of usage. This makes Bedrock SFT economically viable only when inference volume is high enough to amortize the PT commitment. At moderate volume (< 100K requests/day), Together.ai or Vertex fine-tuned inference will typically be cheaper end-to-end.

6.3 Nova Forge Customization Capabilities

Nova Forge (the Nova-specific fine-tuning SDK, distinct from Bedrock’s generic fine-tuning UI) adds:

  • Data mixing — combine multiple task-specific datasets with configurable blend ratios
  • Distillation — use Nova 2 Pro as teacher model to distill into smaller Nova variants
  • Thinking controls — fine-tune reasoning behavior (Lite only)
  • Annual subscription pricing (contact AWS); not pay-per-token like Bedrock SFT

7. Nova in Knowledge Bases — Model Selection Tradeoffs

7.1 Default Model Behavior

Amazon Bedrock Knowledge Bases requires two model choices:

  1. Embedding model — converts chunks to vectors (default: Amazon Titan Embeddings V2)
  2. Foundation model — generates the RAG response from retrieved context

The console default for the FM is not Nova — it historically defaulted to older Claude variants. Teams migrating to Nova for cost reasons must explicitly select a Nova model at KB creation or via the RetrieveAndGenerate API parameter.

7.2 Nova vs. Claude in RAG Pipelines

Dimension Nova Lite Nova Pro Claude Haiku 3.5
Input $/1M $0.060 $0.800 $0.800
Context for retrieved chunks 300K 300K 200K
Instruction-following for citations Good Strong Strong
Hallucination rate on grounded tasks Comparable Comparable Lower
Cost for 10K RAG queries (avg 2K tokens in/out) $1.20 $16.00 $16.00

For Knowledge Base workloads where the retrieval step keeps context under 16K tokens and the task is factual Q&A (not creative generation), Nova Lite offers ~13× cost reduction vs. Nova Pro or Claude Haiku with comparable quality. The Nova Pro advantage over Nova Lite in RAG is primarily in multi-document synthesis tasks where reasoning over conflicting retrieved passages matters.

Recommended KB model selection defaults by use case:

Use case Recommended model
Simple FAQ, policy lookup Nova Micro (text only) or Nova Lite
Document Q&A with images/charts Nova Lite
Multi-document synthesis, audit evidence Nova Pro
Contract analysis with 500K+ token context Nova Premier

8. Amazon Nova Premier — Positioning vs. Nova Pro

Nova Premier launched April 30, 2025. Updated Nova Premier 1.0 shipped October 31, 2025.

Attribute Nova Pro Nova Premier
Input $/1M $0.800 $2.500
Output $/1M $3.200 $12.500
Context window 300K 1M
Speed (tokens/sec) ~200 ~33
AI Analysis Intelligence Index 8/100 13/100
Tool use / function calling Yes Yes
Vision Yes Yes

The 3.1× price premium for Premier over Pro is justified only by the 1M context window. On raw benchmark scores, Premier does not outperform Pro (both rank below average for their respective price tiers on Artificial Analysis). Teams choosing Premier for reasoning quality over Pro will be disappointed — the correct comparison is Premier vs. Claude Sonnet or Gemini 2.5 Pro for frontier-quality tasks, where those external models win.

Pragmatic guideline: Use Premier as a context-length override, not a quality upgrade. If your workflow requires > 300K tokens in a single call (full codebase analysis, all-hands transcript synthesis, multi-chapter document review), Premier is the cheapest Bedrock-native option at 1M context. For everything else, Pro or Nova 2 Lite is the better choice.

9. Nova vs. GPT-4o-mini / Gemini Flash — Enterprise Evaluation Benchmark

This section provides a compact decision framework for enterprise teams benchmarking cheapest-viable options.

9.1 Cost Benchmark — 1M Input / 200K Output Workload

Model Input cost Output cost Total Context
Nova Micro $0.035 $0.028 $0.063 128K
Nova Lite $0.060 $0.048 $0.108 300K
GPT-4o mini $0.150 $0.120 $0.270 128K
Gemini 2.5 Flash $0.150 $0.120 $0.270 1M
Nova Pro $0.800 $0.640 $1.440 300K
Claude Haiku 3.5 $0.800 $0.800 $1.600 200K
Nova 2 Lite $0.300 $0.500 $0.800 1M

9.2 Benchmark Score Reference (Artificial Analysis, June 2026)

Model Intelligence Index Output speed (tok/s) Pricing tier
Nova Micro ~5 High Cheapest
Nova Lite ~8 High Budget
Nova Pro 8 ~200 Mid
Nova 2 Lite 11.8 226.6 Mid
Nova Premier 13 32.7 Premium
GPT-4o mini ~40 High Budget
Gemini 2.5 Flash ~55 High Budget
Claude Haiku 3.5 ~45 High Mid

Honest assessment: GPT-4o mini and Gemini 2.5 Flash significantly outperform Nova Micro and Nova Lite on intelligence benchmarks at 2.5× the price. The Nova budget tier advantage is real-money cost savings on AWS-native workloads where Bedrock consolidation, IAM-native auth, and CUR simplicity outweigh benchmark gaps. For teams that are Bedrock-committed, Nova is the right default. For teams with flexibility, GPT-4o mini or Gemini 2.5 Flash deliver better reasoning per dollar in the budget tier.

10. LiteLLM Configuration for Nova Models

10.1 Version Requirement

Nova models require LiteLLM v1.53.5 or higher. Earlier versions lack Nova-specific routing and will fail silently or miscalculate costs.

10.2 Model ID Formats

Nova models use the bedrock/converse/ route prefix in LiteLLM. The converse route is required for tool use and structured output:

bedrock/converse/amazon.nova-micro-v1:0
bedrock/converse/amazon.nova-lite-v1:0
bedrock/converse/amazon.nova-pro-v1:0
bedrock/converse/amazon.nova-premier-v1:0
bedrock/converse/amazon.nova-2-lite-v1:0

Cross-region inference profiles:

bedrock/converse/us.amazon.nova-pro-v1:0
bedrock/converse/eu.amazon.nova-lite-v1:0

10.3 LiteLLM Proxy config.yaml — Nova Fleet

model_list:
  # Cheapest tier — classification and routing
  - model_name: nova-micro
    litellm_params:
      model: bedrock/converse/amazon.nova-micro-v1:0
      aws_access_key_id: os.environ/AWS_ACCESS_KEY_ID
      aws_secret_access_key: os.environ/AWS_SECRET_ACCESS_KEY
      aws_region_name: us-east-1

  # Budget multimodal — Q&A, document extraction
  - model_name: nova-lite
    litellm_params:
      model: bedrock/converse/amazon.nova-lite-v1:0
      aws_region_name: us-east-1

  # Cross-region for throughput scaling
  - model_name: nova-lite-cross-region
    litellm_params:
      model: bedrock/converse/us.amazon.nova-lite-v1:0
      aws_region_name: us-east-1

  # Mid-tier reasoning — agentic workflows
  - model_name: nova-pro
    litellm_params:
      model: bedrock/converse/amazon.nova-pro-v1:0
      aws_region_name: us-east-1

  # Reasoning with 1M context
  - model_name: nova-2-lite
    litellm_params:
      model: bedrock/converse/amazon.nova-2-lite-v1:0
      aws_region_name: us-east-1

  # 1M context long-document tasks
  - model_name: nova-premier
    litellm_params:
      model: bedrock/converse/amazon.nova-premier-v1:0
      aws_region_name: us-east-1

router_settings:
  routing_strategy: least-busy
  fallbacks:
    - nova-lite:
        - nova-micro    # fallback to Micro if Lite quota exceeded
    - nova-pro:
        - nova-lite     # fallback to Lite for graceful degradation

10.4 Known LiteLLM Issues with Nova

  1. Streaming on converse route — Nova models work with streaming via bedrock/converse/, but do not use the non-converse bedrock/ prefix for streaming — it lacks tool use support.
  2. Cost tracking for multimodal — LiteLLM’s token counter does not apply the flat 230-token image billing rate; it may undercount image costs. Validate with CUR rather than LiteLLM cost estimates for multimodal workloads.
  3. Nova 2 Lite reasoning tokens — When thinking is enabled, reasoning tokens appear in the usage response as a separate field. LiteLLM v1.53.5+ passes these through correctly; earlier versions drop them from cost calculations.

11. CUR Line Item Patterns for Nova vs. Anthropic — Athena Segmentation

11.1 Critical Billing Split

Amazon Nova and Anthropic Claude are billed under different AWS product codes:

Model family line_item_product_code Notes
Nova (all tiers) AmazonBedrock Native AWS models
Anthropic Claude marketplace Routes through AWS Marketplace
Llama (via Bedrock) marketplace Same as Claude
Bedrock Guardrails AmazonBedrock Native
Bedrock Knowledge Bases AmazonBedrock Native

A query that only scans product_code = 'AmazonBedrock' will capture all Nova costs but miss all Claude/Llama costs. A query that only scans product_code = 'marketplace' captures Claude/Llama but misses Nova, Guardrails, and Knowledge Bases overhead.

11.2 line_item_usage_type Format

The line_item_usage_type field encodes: region + model + token type + tier + routing.

Format: {region}-{model}-{token-type}[-{tier}][-cross-region-{scope}]

Nova examples:

USE1-NovaMicro-input-tokens                          # Nova Micro, Standard, in-region
USE1-NovaLite-output-tokens                          # Nova Lite, Standard, in-region
USE1-NovaPro-input-tokens-flex                       # Nova Pro, Flex tier
USE1-Nova2.0Lite-input-tokens-flex                   # Nova 2 Lite, Flex tier
USE1-NovaPro-cache-read-input-token-count            # Nova Pro, cache read
USE1-NovaPro-output-tokens-cross-region-global       # Nova Pro, cross-region

Anthropic examples (Marketplace):

USE1-Claude4.6Sonnet-input-tokens                    # Claude Sonnet 4.6, Standard
USE1-Claude4.6Sonnet-cache-read-input-token-count    # Claude Sonnet 4.6, cache read
USE1-Claude4.6Sonnet-output-tokens-cross-region-global

Important: Cache read and cache write tokens are separate line_item_usage_type entries. Summing only input + output tokens will undercount total cost for any workload using prompt caching. This is the most common CUR reconciliation error.

11.3 Athena Queries — Model Cost Segmentation

Query 1: Total Bedrock spend by model family (month)

SELECT
  bill_billing_period_start_date,
  CASE
    WHEN line_item_product_code = 'AmazonBedrock'
         AND line_item_usage_type LIKE '%Nova%'
      THEN 'Amazon Nova'
    WHEN line_item_product_code = 'marketplace'
         AND line_item_usage_type LIKE '%Claude%'
      THEN 'Anthropic Claude'
    WHEN line_item_product_code = 'marketplace'
         AND line_item_usage_type LIKE '%Llama%'
      THEN 'Meta Llama'
    WHEN line_item_product_code = 'AmazonBedrock'
      THEN 'Bedrock (Other Native)'
    ELSE 'Other'
  END AS model_family,
  SUM(line_item_blended_cost) AS total_cost_usd
FROM cur_database.cur_table
WHERE
  line_item_usage_start_date >= DATE_TRUNC('month', CURRENT_DATE - INTERVAL '1' MONTH)
  AND line_item_usage_start_date < DATE_TRUNC('month', CURRENT_DATE)
  AND (
    line_item_product_code = 'AmazonBedrock'
    OR line_item_product_code = 'marketplace'
  )
GROUP BY 1, 2
ORDER BY 1, 3 DESC;

Query 2: Nova token type breakdown — catch cache reconciliation gaps

SELECT
  line_item_usage_type,
  CASE
    WHEN line_item_usage_type LIKE '%-input-tokens%'
         AND line_item_usage_type NOT LIKE '%-cache-%'
      THEN 'input'
    WHEN line_item_usage_type LIKE '%-output-tokens%'
      THEN 'output'
    WHEN line_item_usage_type LIKE '%-cache-read-%'
      THEN 'cache_read'
    WHEN line_item_usage_type LIKE '%-cache-write-%'
      THEN 'cache_write'
    ELSE 'other'
  END AS token_type,
  SUM(line_item_usage_amount) AS total_tokens,
  SUM(line_item_blended_cost) AS total_cost_usd
FROM cur_database.cur_table
WHERE
  line_item_product_code = 'AmazonBedrock'
  AND line_item_usage_type LIKE '%Nova%'
  AND line_item_usage_start_date >= DATE_TRUNC('month', CURRENT_DATE - INTERVAL '1' MONTH)
GROUP BY 1, 2
ORDER BY 4 DESC;

Query 3: Cross-region inference cost premium (should be zero for Nova, confirm)

SELECT
  CASE
    WHEN line_item_usage_type LIKE '%-cross-region-%' THEN 'cross-region'
    ELSE 'in-region'
  END AS routing,
  SUM(line_item_blended_cost) AS cost_usd,
  SUM(line_item_usage_amount) AS tokens
FROM cur_database.cur_table
WHERE
  line_item_usage_type LIKE '%Nova%'
  AND line_item_usage_start_date >= DATE_TRUNC('month', CURRENT_DATE - INTERVAL '1' MONTH)
GROUP BY 1;

Query 4: Per-IAM-principal Nova spend (requires CUR 2.0 + tag activation)

SELECT
  line_item_iam_principal,
  line_item_usage_type,
  SUM(line_item_blended_cost) AS cost_usd
FROM cur_database.cur_table
WHERE
  line_item_product_code = 'AmazonBedrock'
  AND line_item_usage_type LIKE '%Nova%'
  AND line_item_iam_principal IS NOT NULL
  AND line_item_usage_start_date >= DATE_TRUNC('month', CURRENT_DATE - INTERVAL '1' MONTH)
GROUP BY 1, 2
ORDER BY 3 DESC
LIMIT 50;

Note: line_item_iam_principal requires CUR 2.0 (AWS Data Exports). Legacy CUR format does not include this column.

11.4 CUR FinOps Gotchas — Nova-Specific

  1. Guardrails overhead is a separate line item — Bedrock Guardrails adds $0.15 per 1,000 text units under a distinct line_item_usage_type. It appears under AmazonBedrock product code. At high volume with Guardrails enabled, this can exceed the Nova token cost itself.

  2. Nova Act is time-billed, not token-billed — Nova Act bills at $4.75/agent-hour. It does not appear as a token usage type. Filter by line_item_operation = 'AgentHour' or the equivalent Nova Act product code.

  3. Batch API requests use async job IDs — Batch inference does not produce real-time CUR line items. Jobs settle within 24 hours; cost attribution appears as a bulk line item without per-request granularity.

  4. Flex tier requests are marked in usage type — The -flex suffix in line_item_usage_type is the only way to distinguish Flex from Standard billing retroactively. Applications that accidentally send Flex traffic to Standard tier will not show an error; they will simply incur 2× the expected cost. Audit for missing -flex suffixes when Flex tier is intended.

12. Operational Recommendations Summary

Priority Action
Cost baseline Route classification/intent/routing traffic to Nova Micro first; escalate to Lite/Pro only on failure
Throughput Enable cross-region inference profiles on Nova Lite/Pro; enables 2× throughput at no cost premium
Batch jobs All non-interactive pipelines (evals, summarization, indexing) to Flex or Batch tier for 50% savings
Long context Nova 2 Lite over Nova Premier for 1M-context tasks — better benchmark scores, faster, and identical price tier after Nova 2 pricing
Fine-tuning Use Bedrock SFT for Nova Pro only when provisioned throughput volume justifies the PT commitment; for smaller deployments, Vertex or Together is cheaper end-to-end
CUR Set up CUR 2.0 (AWS Data Exports); activate IAM principal tags; audit for cache token reconciliation gaps monthly
LiteLLM Pin to v1.53.5+; use bedrock/converse/ prefix; validate multimodal cost estimates against CUR independently

Sources