Skip to main content

Part II — The 2026 AI Stack

Once you can ship, the next question is: which of the noise actually matters?

This part has two layers. The model tiers and tool picks are opinionated — concrete recommendations, ranked by whether you should adopt them now, learn them later, or skip them entirely. The trends page is directional — broad shifts in how 2026 AI is built.

Read this part as a shipping engineer

Part II assumes you've shipped at least one production-style AI feature (Stage 9 in Part I, or its equivalent). If you haven't, the trade-offs here won't land — adoption questions only matter when you have something to adopt them into. Build first; then come back here for the next layer.

What's in this part​

Model tiers (start here)​

  • Frontier tier — GPT-5, Claude Opus 4.x, Gemini 2.x Ultra. The "use when nothing else passes the eval" tier.
  • Workhorse tier — Claude Sonnet 4.x, GPT-5 family, Gemini Pro. The default-for-real-work tier.
  • Cheap tier — Claude Haiku 4.5, GPT-5-mini, Gemini Flash. The "start here, climb only when forced" tier.
  • Embedding tier — text-embedding-3-small/large, Voyage, Cohere v3, open-source picks.

Infrastructure picks​

  • Vector DB pick — pgvector by default; when to reach for Qdrant / Pinecone / Weaviate.
  • Framework pick — LangChain, LlamaIndex, Vercel AI SDK, OpenAI Agents SDK, Pydantic AI, raw — which when.
  • Eval tool pick — Braintrust, Langfuse, Promptfoo, DeepEval, Ragas, OpenAI Evals.
  • Observability pick — Langfuse, Helicone, LangSmith, Phoenix, Braintrust.
  • Gateway pick — Cloudflare AI Gateway, LiteLLM, Helicone, OpenRouter — and whether you need one.
  • Trends — six 2026 directional shifts: context engineering, agents-as-product, multimodal-default, on-device inference, MCP everywhere, eval-as-CI.

The tier rule (read first)​

The tier list isn't "use Tier 1 things, ignore Tier 3." It's a decision-cost ladder:

  • Tier 1 = adopt without much thought; mature, well-supported, the boring-tech answer.
  • Tier 2 = worth knowing; reach for when Tier 1 fails a specific named constraint.
  • Tier 3 = skip or defer; either too new to be safe, too niche to be relevant, or a fashion item.

The cardinal sin is treating Tier 1 as a checklist to adopt all at once. Pick the one item that fixes a pain you have right now in the next project. Six rewrites in parallel is how the project stalls.

How this pairs with Chapter 4 (Stack)​

Chapter 4 decodes what each tool does and why it exists — the reference. This part tells you which ones to actually use right now — the opinion.

If you want the wide view of every option, go to Chapter 4. If you want the short list, stay here.

→ Start with Cheap tier — most production AI runs there, contrary to the hype.