Model Matrix

Where this stops working: model names, pricing and capabilities turn over every few months. This page carries a verification date; past it, re-check before relying on anything — especially prices, which should not be used for cost modelling straight from here.

Verified: 2026-07-31 This page is the single source of truth for model ids across the whole book. Every other chapter describes capability tiers and links back here for the actual model names.

Tool prices in this section were checked in 2026-08. SaaS pricing moves often — verify on the vendor’s own site before you commit.

Why chapters don’t name specific models

The methodology in this knowledge base has a half-life measured in years. Model ids have a half-life measured in months. Two examples from the same week this page was verified: OpenAI cut the Luna tier by 80% and the Terra tier by 20% on 2026-07-30; Google shipped three new Flash models on 2026-07-21. Any chapter that hard-codes a model id becomes wrong within a quarter.

So the convention here is:

• Chapters say “use a frontier-tier model to cluster competitor complaints”, not “use model X to cluster competitor complaints”
• Model ids, prices, and context limits — the perishable facts — live only on this page, with a verification date
• On a new generation you edit one file, once per language, instead of touching 60+ chapters

The one exception is The Evolution of AI. That chapter is about history; GPT-3 and Claude 2 are its subject matter, not a recommendation, so those names stay.

The four capability tiers

Tier definitions are stable. You can build a tech-selection process on them without revisiting it every time a vendor ships.

Rule of thumb: prove the workflow on T1 and generate a batch of “gold answers,” then move the volume to T2/T3 and use those gold answers as your acceptance test. Most e-commerce tasks settle on T2. Starting your prompt iteration on T3 is a common waste of time — you can’t tell whether a bad result means a bad prompt or an underpowered model.

Current models by tier

Models and prices below are current as of 2026-07-31. Prices are API dollars per million tokens (input/output) and are here for order-of-magnitude comparison only — the vendor page is authoritative.

Cloud APIs

Worth knowing:

• All three GPT-5.6 tiers share roughly a 1.05M-token context window, 128K max output, and a 2026-02-16 knowledge cutoff. “Feed every competitor listing in the category at once” no longer requires chunking.
• OpenAI cut Luna (−80%) and Terra (−20%) on 2026-07-30. If your cost model is a few months old, rerun it.
• Google’s version numbers don’t line up across tiers — Flash is on 3.6 while Pro’s 3.1 is still Preview, and 3.5 Pro is in partner testing. Pick by tier, not by which number is bigger.
• The ChatGPT web/app model names are not the API names (the app currently mixes GPT-5.3 Instant as default, GPT-5.4 Thinking/Pro, GPT-5.5, and GPT-5.6). Keep the distinction when writing SOPs — your operations team is on the app.

T4 Local (open weights)

To run: install Ollama or LM Studio, then ollama run qwen3:8b pulls a quantized GGUF and exposes an OpenAI-compatible local endpoint — meaning every OpenAI-SDK example in this book works locally by changing one base_url. Use vLLM for high-concurrency production, llama.cpp for CPU/edge. Short on VRAM? Drop the quantization (Q4 and below), at some cost to quality.

Full deployment walkthrough in B5 Local Model Deployment.

Video generation

The single most important rule for e-commerce: for the product itself, always go image-to-video (start from a real product photo). Never text-to-video. Text-to-video re-imagines your product — details, proportions, and logo will drift, and using that as ad creative is a compliance risk.

OpenAI’s Sora 2 is no longer a recommendation — the consumer experience ended in April 2026 and the API is scheduled to shut down on 2026-09-24. If your pipeline still depends on it, plan a migration.

Image generation

There is no longer a single “best” image model. The working combination is: volume through the FLUX.2 Pro or Nano Banana Pro API, hero images curated by hand out of Midjourney, precise-text work in Ideogram. See A7 Visual Content and B9 AI Image Pipeline.

How to re-verify this page

This table will go stale. Refresh it quarterly — it takes about ten minutes:

1. Official pricing pages (the only trustworthy source; third-party comparison sites lag badly)
   • Anthropic: https://www.anthropic.com/pricing
   • OpenAI: https://openai.com/api/pricing/
   • Google: https://ai.google.dev/gemini-api/docs/pricing
2. Official model lists — check for new tiers and for anything marked deprecated
   • https://docs.claude.com/en/docs/about-claude/models
   • https://developers.openai.com/api/docs/changelog
   • https://ai.google.dev/gemini-api/docs/models
3. Open-weight leaderboards for T4 turnover: the Hugging Face open LLM leaderboard, LMArena
4. Update the verification date at the top of this page and add a row to the changelog below

The test for “does the prose need to change too” is simple: if the tier boundaries still hold, no chapter needs a single edit. Only a genuinely new shape that the four tiers can’t hold — say, a vendor shipping a commerce-specific vertical model — justifies revisiting the chapter text.

Changelog