Choose your AI model
CDK Insights routes AI analysis through AWS Bedrock and lets you pick the model that fits each scan. More capable models cost more credits per resource. The default works well for most stacks — switch up only when you want deeper reasoning.
Quick reference
| Model | Alias | Credits / resource | Tier |
|---|---|---|---|
| Amazon Nova Lite Amazon | nova-lite | 0.5 | Free tier default |
| Mistral 14B Mistral AI | mistral-14b | 1 | Pro / Team default |
| Llama 3.3 70B Meta | llama-3-3-70b | 2 | Pro / Team |
| Claude Haiku 4.5 Anthropic | haiku-4-5 | 4 | Pro / Team |
| Claude Sonnet 4.6 Anthropic | sonnet-4-6 | 16 | Pro / Team |
When to pick which model
Amazon Nova Lite
0.5 credits / resourceFree tier defaultnova-litePick when: Quick scans, cost-conscious workflows. Default for the Free tier — your 500 credits stretch to ~1,000 resource analyses.
Strengths
- Lowest credit cost (2× more analyses per credit)
- Fast inference
- AWS-native model
Avoid when
Architecturally complex stacks where nuanced reasoning matters more than raw count.
Mistral 14B
1 credits / resourcePro / Team defaultmistral-14bPick when: Balanced reasoning for everyday infrastructure. The default for paid tiers — works well for the vast majority of CDK stacks.
Strengths
- Anchor cost: 1 credit per resource
- Solid code understanding
- Reliable for typical AWS patterns
Avoid when
Deeply layered constructs or complex permission graphs — Llama 3.3 70B or Haiku 4.5 will reason more thoroughly.
Llama 3.3 70B
2 credits / resourcePro / Teamllama-3-3-70bPick when: Deeper code understanding and pattern recognition than the default — without the premium price of Anthropic models.
Strengths
- Stronger reasoning than Mistral 14B
- Open-weight model with broad training
- Good middle ground on cost vs depth
Avoid when
Routine scans where Mistral 14B already produces clean output — pay for depth only when you need it.
Claude Haiku 4.5
4 credits / resourcePro / Teamhaiku-4-5Pick when: High-quality findings for security-critical infrastructure. Strong reasoning at speed — great for production reviews.
Strengths
- Anthropic-quality reasoning
- Faster than Sonnet for similar quality on most CDK code
- Excellent at explaining nuanced findings
Avoid when
Quick iteration loops where speed/cost matter more than depth.
Claude Sonnet 4.6
16 credits / resourcePro / Teamsonnet-4-6Pick when: Maximum-depth reasoning for architecturally complex stacks where every nuance counts. Reserve for high-stakes audits.
Strengths
- Top-tier reasoning across the registry
- Best at multi-resource architectural patterns
- Catches subtle interactions Mistral and Llama miss
Avoid when
Routine development scans — 16× the credit cost of Mistral means a 5,000-credit Pro plan covers ~312 Sonnet analyses vs 5,000 on the default.
How to switch models
Three ways to set the model, in order of precedence. The CLI walks them in the order shown — the first non-empty value wins.
1. Per-scan flag
Highest precedence. Use for one-off audits or when iterating on which model to settle on.
npx cdk-insights scan --model haiku-4-52. Project default in cdk.json context
Sets a default for the whole CDK project. Lives next to the rest of your CDK config — convenient when the team agrees on one model per project.
{
"context": {
"cdkInsights:aiModel": "mistral-14b"
}
}3. User default in .cdk-insights.json
Local user override — useful when you personally prefer a specific model across all your projects.
{
"ai": {
"model": "llama-3-3-70b"
}
}If none of the three is set, the resolver falls back to the tier-default model — Nova Lite for Free, Mistral 14B for Pro and Team.
Tier gating
Free
Restricted to nova-lite. If you set a higher-tier model on Free, the resolver falls back to Nova Lite with a warning rather than failing — so the scan still runs.
Pro
Full access to all 5 models. Default is Mistral 14B; pick any model per scan with --model.
Team
Same as Pro — full access to all 5 models. Each seat gets its own 10,000-credit allowance.
Credit math, worked examples
Each AI analysis on a resource spends credits at the model's rate. Cached results don't spend credits. Static scans are always free. Here's what each plan's monthly allowance buys at each model:
Free (500 credits/month)
Pro (5,000 credits/month)
Team (10,000 credits/seat/month)
Counts assume one analysis per resource. With ai.batchSize enabled, multiple resources share a single Bedrock call — the credit cost stays the same per resource, but Bedrock-side prompt tokens are amortised across the batch.
Ready to pick a model?
Start with the default, switch up with --model when you need it.