Module 1: Genai In Production Vs A Notebook
6. Bedrock, 2026 landscape: models, families, and why the cheap path stays the default
Description
extract-shipment-manifest-fields is going to invoke "a Bedrock model" — but Bedrock isn't a model, it's a catalog of models from several different providers, served through a single API. This lesson does the work lesson 3 deliberately left pending: mapping which model families exist in Bedrock, verified against official AWS documentation as of the time of writing (August 2026), not against unchecked training knowledge. And it closes by revisiting the thesis you already installed in lesson 3, one level down: the cheap path stays the default even within the catalog of available models — choosing the cheapest model that solves the problem is the same discipline as choosing the deterministic parser before any model.
Connection to the module
Lesson 5 confirmed the real starting point of the inherited infrastructure. This lesson maps the side you haven't touched yet: what's on the other side of bedrock:InvokeModel. Lesson 7, immediately after, tries to actually touch it against this $0 lab — and honestly documents exactly where that attempt stops.
Analogy: a translation provider's catalog, not a single translator
Think of Bedrock as the catalog of an agency that offers freelance translators of various specialties, all hirable through the same platform, with the same consolidated monthly bill. Some translators are fast, cheap generalists, good for a high volume of simple texts. Others are expensive specialists, better for difficult nuance, legal documents, real ambiguity. Hiring "a translator" without specifying which one is an incomplete question — the right answer depends on how difficult the text is and how many texts need translating per month. Bedrock is exactly that agency: an infrastructure provider that exposes, through a single API, models from several different labs, each with a different cost, speed, and capability profile. Choosing the right model for extract-shipment-manifest-fields isn't a question of "which model is the best?" — it's a question of "which is the cheapest translator that can reliably read an email describing a poorly worded shipment?"
How this lesson was verified
Every claim in this lesson about which models exist in Bedrock today was verified against official AWS documentation, checked at the time of writing this guide (August 2026) — not assumed from training knowledge. The exact sources are in the Resources section, at the close. Bedrock's model catalog changes frequently — AWS adds and retires models regularly, and some stay marked as preview before general availability —, so this lesson focuses on families with a stable position in the catalog, not an exhaustive inventory that would go stale within months. If you need the exact catalog current on the day you're reading this, the authoritative source is always docs.aws.amazon.com/bedrock/latest/userguide/model-cards.html — the official documentation's "Models at a glance" page, cited below.
The model families, verified
Amazon Nova — AWS's native family
Nova is Amazon's own line of foundation models, designed specifically for Bedrock, with several sizes covering different cost and latency points:
| Model | Profile | Typical use |
|---|---|---|
| Nova Micro | The fastest and cheapest, text only | Simple classification, field extraction, high-volume tasks |
| Nova Lite | Multimodal (text, image, video), fast | General-purpose tasks with good cost/capability ratio |
| Nova Pro | Multimodal, higher reasoning capability | More complex tasks, still notably cheaper than other providers' top-tier models |
| Nova Premier | The highest-capability model in the original line | Tasks demanding maximum reasoning within the Nova family |
| Nova 2 Lite | Next generation, with extended step-by-step reasoning, generally available | Replaces Nova Lite for tasks where explicit reasoning helps |
| Nova 2 Pro | Next generation, in preview at time of writing | Complex agentic tasks — multi-document analysis, extended reasoning |
Amazon announced Nova 2 in December 2025: Nova 2 Lite immediately available in Bedrock, Nova 2 Pro in preview, with early access for Amazon Nova Forge customers — both with three reasoning-intensity levels (low, medium, high) and context of up to one million tokens (Source #1, below). For a task like Andes Cargo's — extracting a handful of structured fields from a short text —, Nova Micro or Nova Lite are, typically, the most reasonable starting point: the task doesn't demand the multi-step reasoning that would justify a more expensive model.
Anthropic Claude — available in Bedrock, with an extra enablement step
Claude is available in Bedrock as one of the longest-standing third-party providers in the catalog, with several sizes active simultaneously (the full line includes variants optimized for speed versus variants optimized for deep reasoning — the exact naming and current versions change over time, so the authoritative source is always Bedrock's model page at the time you check it, not a number fixed in this lesson). One specific, verified fact that's worth fixing here, because it doesn't change with every new version: since the simplification of Bedrock model access, most serverless models get enabled automatically by default on a new account — but Anthropic's models remain the explicit exception, and require a one-time use-case form before first use, both via API and via console (Source #2, below). It's an honest fact worth anticipating: if you ever complete this guide's M3-M4 against a real AWS account with a Claude model, that form is an extra step Nova or Llama don't require.
Meta Llama — the open-weight line
Meta Llama is available in Bedrock as a managed service, with several active generations: the 3.x line (3, 3.1, 3.2, 3.3, in various parameter sizes) and the 4 line, with the Maverick and Scout models (Source #3, below). Unlike Nova and Claude, Llama is an open-weight model — Meta publishes the model's parameters, and Bedrock serves it as managed infrastructure on top of those public weights. For Andes Cargo's case this doesn't change the billing mechanics (it's still pay-per-token, managed the same as any other Bedrock model), but it is a real business-model difference worth knowing: open-weight models, in general, can also be self-hosted outside Bedrock — an option this guide names, without building, because it falls outside the $0 scope (it would imply dedicated GPU, see the boundary with kubernetes-and-eks-in-production-guide in lesson 8).
Mistral AI — the European provider, with smaller, efficient models
Mistral AI is available in Bedrock with a line of models ranging from compact, highly efficient sizes (Ministral, with a few billion parameters) to large general-purpose models (Mistral Large), including mixture-of-experts models (Mixtral) and multimodal models (Pixtral) (Source #4, below). It's, along with Nova Micro, one of the catalog's most economical profiles — a reasonable candidate, just like Nova, for a low-semantic-risk field extraction task.
The catalog, summarized
BEDROCK — MAIN FAMILIES, AUGUST 2026
Amazon Nova Native AWS. Micro/Lite/Pro/Premier + Nova 2 Lite (GA)/Pro (preview).
The cheapest starting point for simple extraction tasks.
Anthropic Claude Several versions active simultaneously. Requires a one-time
use-case form before first use -- the only requirement of this
kind among the families listed here.
Meta Llama 3.x line + 4 line (Maverick, Scout). Open weight -- also
self-hostable outside Bedrock, an option not built in this guide.
Mistral AI Ministral (compact) to Mistral Large (general-purpose),
Mixtral (mixture of experts), Pixtral (multimodal).The thesis, revisited one level down: cheap within cheap
Lesson 3 installed this guide's central architecture decision: the deterministic parser is the default, the LLM is the escalation path. This catalog reveals that same discipline applies one level down, within the decision of which model to use once you've already decided to escalate. extract-shipment-manifest-fields doesn't need the catalog's most capable model — it needs the cheapest model that reliably solves a specific, bounded task: reading a short text and returning five structured fields. Choosing Nova Premier or Claude's highest-capability model for this task would be the same antipattern, under a different name, as invoking any Bedrock model for a manifest the deterministic parser can already read on its own: paying more than the problem demands, out of habit instead of criteria.
This isn't yet the final decision of which model Andes Cargo uses — that decision, with the real cost numbers behind it, belongs to this guide's M2. What this lesson leaves installed is the criterion: within a catalog with dozens of models available, the right question is never "which one is best?" — it's "which is the cheapest one my specific task needs?", the same question, at a smaller scale, you already answered in lesson 3 when deciding when to escalate at all.
Common mistakes
Assuming "more expensive" always means "more reliable" for this specific task (intuition mistake). What happens: someone, looking at the family table, concludes the catalog's largest, most expensive model is, by default, the safest choice for extract-shipment-manifest-fields. How to spot it: if your reasoning is "better to overpay and not risk it." How to fix it: for a low-semantic-risk structured field extraction task, a larger model doesn't necessarily extract Andes Cargo's five fields any better — it may simply cost more for the same quality of result. The correct decision, which M2 formalizes with numbers, is empirical: test the cheapest candidate model first, and scale up in size only if the real success rate falls short of the acceptable threshold.
Memorizing this lesson's exact version names as if they were fixed forever (expectation mistake). What happens: someone writes down "Claude such-and-such version" or "Nova such-and-such version" from this lesson as if it were Bedrock's definitive catalog. How to spot it: if, months after reading this guide, you're still citing this lesson as the source for the current catalog. How to fix it: this lesson states explicitly that the catalog changes frequently, and that's exactly why it focuses on stable families, not a version inventory. The correct source, always, is Bedrock documentation's official "Models at a glance" page — not this lesson, nor any guide written at a fixed point in time.
Thinking Llama, being open-weight, is "free" on Bedrock (business-model confusion mistake). What happens: someone concludes that, since Meta publishes Llama's parameters, using it through Bedrock has no inference cost. How to spot it: if your expectation is "Llama on Bedrock should be free, because the model itself is open." How to fix it: open weight describes who can self-host the model outside Bedrock, not how it's billed inside Bedrock. A Llama model served by Bedrock is billed exactly the same as any other model in the catalog — pay per token, under the same on-demand/Provisioned/Batch pricing model this guide's M2 explains thoroughly.
Exercises
Exercise 1 — Choose an initial candidate for extract-shipment-manifest-fields, and justify it using this lesson's criterion. Based solely on this lesson's catalog (not on M2, which you haven't read yet), which model family and size would you test first for Andes Cargo's extraction task? Justify it without using the word "best."
See solution
There's no single correct answer, but a solid justification sounds like this: "I'd start with Nova Micro or Nova Lite, or a compact Mistral model (Ministral) — both are the catalog's cheapest profiles, and the task (extracting five structured fields from a short text, without deep semantic ambiguity) doesn't seem to demand a larger model's advanced reasoning. The criterion isn't 'which is the catalog's best model,' it's 'which is the cheapest model that probably solves this specific task' — if the real success rate, measured with M7's smoke test harness, turns out insufficient, that would be the moment to move up in size, not before."
Exercise 2 — Explain the difference between Nova and Llama in terms of who owns the model. A colleague asks why this lesson mentions Llama is "open-weight" but doesn't say the same about Nova. Explain the difference.
See solution
Amazon Nova is Amazon's own property — the model's parameters aren't public, and the only way to use it is through a managed service like Bedrock (or, in some cases, other proprietary AWS surfaces). Meta Llama is open-weight — Meta publishes the model's parameters, which means, in principle, anyone with enough compute capacity could self-host it outside Bedrock, without depending on Amazon as an intermediary. Bedrock, for Llama, offers the convenience of not having to manage that infrastructure yourself — Amazon hosts the model, charges per token, and the user never sees a server —, but that convenience is a choice, not the only way to use Llama, unlike Nova.
Exercise 3 — Explain why this lesson doesn't give a price figure. A colleague asks why, if this lesson compares several model families as "cheaper" or "more expensive" relative to each other, it doesn't include a single dollar figure. Why is that omission intentional?
See solution
Because the exact price per million tokens, input and output, is a figure that changes more often than the model families themselves, and because mixing it in here would compete with the specific work of this guide's M2, which exists exactly for that: building the complete cost model, with figures cited from Bedrock's official pricing page, and the custom calculator that converts them into a real monthly projection given a volume assumption. This lesson deliberately stays at the level of "cheaper/more expensive in relative terms" so as not to duplicate, with worse upkeep, what M2 does with the rigor and exact citation the topic deserves.
Summary and next step
In this lesson you mapped the main model families available in Bedrock, verified against official documentation from August 2026: Amazon Nova (native, the cheapest starting point), Anthropic Claude (with its extra enablement form), Meta Llama (open-weight, served the same as any other), and Mistral AI (compact and efficient). And you revisited this guide's central thesis one level down: the same criterion that decides when to escalate from the deterministic parser to the LLM — the cheapest path that solves the problem — also decides which model, within the catalog, is the right starting point.
Before moving on you should be able to: name this lesson's four families and one key difference for each; explain why Claude requires an enablement step the other families don't; and explain, without using the word "best," why the catalog's most expensive model isn't automatically the right choice for extract-shipment-manifest-fields.
Lesson 7 does what this lesson still hasn't: try to actually touch Bedrock, against this $0 lab, with the exact command and the exact reason — cited, not assumed — for what happens when you try.
Resources
- AWS — Announcing Amazon Nova 2 foundation models now available in Amazon Bedrock — Source #1, the December 2025 announcement of Nova 2 Lite (general availability) and Nova 2 Pro (preview).
- AWS — Simplified model access in Amazon Bedrock — Source #2, the default automatic enablement and Anthropic's use-case-form exception.
- AWS — Amazon Bedrock — Source #3, general overview of available model providers, including Meta's Llama line.
- AWS Docs — Models at a glance (Amazon Bedrock) — Source #4, the official, always-current page for the complete catalog of providers and models, including Mistral AI — the authoritative source to verify the current catalog any day you're reading this.
- AWS — Amazon Bedrock Pricing — reference for this guide's M2, where the per-token price for each family is cited with the detail this lesson deliberately doesn't include.