Two glowing doorways, a modest open one before a sealed radiant one, on a dark navy background
EP 2April 21, 20265 min read

The Second Best Model You Can Actually Buy

PodcastModel TieringStrategy
MW
Matt Wozniak
April 21, 2026 · 5 min read

My take on Anthropic shipping its best public model while admitting it's only their second best, and what frontier-model tiering means for anyone building. Companion notes to Human in the Loop Episode 2.

Anthropic shipped their best public model this week — and in the same breath admitted it's their second best. The best one is locked behind a door most of us will never walk through. That's the new frontier-lab playbook, and Oscar and I broke down what it means for anyone building in 2026.

Here's how I called the week.

Signal or Noise

Claude Opus 4.7 ships with a 1M context window at standard pricing — and Anthropic says on record it trails Mythos

The capability is great; the honesty about the tiering is the real story. If the best model is one you can't buy, plan your product around the best model you can. Signal.

OpenAI's updated Agents SDK makes sandboxed agents a primitive

Sandboxed agents as a first-class building block is genuinely useful. Python first, TypeScript later. Signal.

AI lands on the desktop — Gemini's native macOS app and Perplexity's Personal Computer

The interface is moving to where you actually work. Signal, lightly — worth watching how deep the OS integration goes.

Stanford's 2026 AI Index — only 10% of Americans are more excited than concerned, vs 56% of experts

That gap between builders and the public is going to shape regulation. Signal for the backdrop.

Allbirds pivots to AI, stock jumps 582% the same day

Noise for your stack, signal for the hype cycle. When a shoe company can add "AI" and jump 582%, you're watching a bubble breathe.

No Jargon Required

Three terms worth having straight:

  1. Context engineering — prompt engineering is writing a good email; context engineering is designing the whole information environment the model works inside. This is the skill that actually moves output quality.
  2. Agentic AI — what an agent actually is, and why Gartner expects a big share of enterprise agent projects to get cancelled. Most fail on scope, not capability.
  3. Frontier model tiering — what happens when "the best model" is one you probably can't buy, and how to build around the best one you can.

The through-line: the frontier is being tiered, and honesty about it is rare. Build on what you can actually rely on. Watch or listen below.