Models: which brain for which task

Different tasks need different brains — and the price of using one brain for everything is either quality (fast model on a hard problem) or usage limits (frontier model on trivia).

The principle

Powerful models (the frontier tier) — architecture decisions, tricky debugging, code review, planning in an unfamiliar codebase. Anything where reasoning is the bottleneck.
Fast models — mechanical edits, renames, boilerplate, formatting, simple lookups. Anything where the task is clear and the model just executes.

A good default: run routine work on a fast model, escalate deliberately when you hit a reasoning-heavy step — and drop back down after.

Anthropic's lineup (as of September 2026)

Anthropic ships Claude as a small ladder of families. From top to bottom:

Claude Fable 5.1 — the most capable generally available model: demanding reasoning, long-running agent work, hard debugging, architecture. 1M-token context window. This is the "escalate to" model — and the priciest, so use it deliberately.
Claude Opus 5 — the strong everyday model for complex agentic coding. In Claude Code it's the recommended default. Most of your serious work happens here.
Claude Sonnet 5 — the balance of speed and capability: solid coding and tool use at lower cost. Good for well-defined tasks, refactors, tests.
Claude Haiku 4.5 — the fast, cheap one: quick answers, mechanical edits, lookups, anything you'd feel silly waiting on a frontier model for.

Two names you'll run into and should be able to place:

Claude Mythos 5.1 — the same underlying model as Fable 5.1 with fewer safeguards, available only to a restricted set of organizations (Project Glasswing, mostly security research). You won't see it in your picker; if someone mentions "Mythos-class," they mean the Fable/Mythos tier above Opus.
Opus 4.8, Sonnet 4.6 and other 4.x names — the previous generation. They still run, but they're superseded; if you see them in an old tutorial, mentally map them to Opus 5 / Sonnet 5.

Lineups change every few months. The names above are correct as of September 2026 — the picker is the source of truth, not this page.

How to switch

Desktop: the model picker next to the input (model name + effort level).
CLI: the /model command inside a session. Your pick becomes the default for new sessions; press s to apply it to the current session only.

The same pickers expose finer grain than just the model: an effort level and a /fast toggle — same principle, smaller dial. And read the picker's own labels: "most capable for your hardest tasks," "efficient for routine," "fastest for quick answers" — the lineup describes the powerful-vs-fast principle itself.

Subagents can pin their own model in frontmatter (see the subagents page) — so your reviewer can always run on a powerful model while the main thread stays fast.

Models from other companies

This track runs on Claude Code, so you'll work with Claude models — but you'll meet the others constantly in articles, tools, and job descriptions. Worth knowing the map:

OpenAI — GPT-6 Astra (released September 3, 2026) is the new frontier model, sitting above the GPT-5.6 family (Sol / Terra / Luna: flagship / balanced / cheap). Astra is the direct counterpart to Fable 5.1: 1M-token context, priced at the same top tier, and — same story as Fable/Mythos — shipped with its most sensitive capabilities (cybersecurity) gated behind a vetted-access program. OpenAI's coding agent is Codex.
Google — Gemini 3.x (Pro for reasoning and huge context, Flash for speed and price). Known for very large context windows and multimodal input.
xAI — Grok 4.x — competitive on agentic and tool-use tasks, priced aggressively.
Open-weight models you can self-host: Meta Llama 4, DeepSeek V4, Alibaba Qwen3-Coder, Mistral Codestral (autocomplete specialist). Near-frontier coding at a fraction of the price, if you're willing to run the infrastructure.

Notice the pattern: every lab ships the same ladder — a fast/cheap tier (Haiku, Flash, Luna), an everyday tier (Sonnet, Terra), and a frontier tier (Fable, Astra, Gemini Pro). The names differ; the principle of this page doesn't. Rankings between labs leapfrog every few months, so treat any "best model" article as a snapshot, not a verdict.

The habit

Before a task, one-second question: is this reasoning or execution? Reasoning → powerful. Execution → fast. Mixed → split the task (Week 3's breakdown skill) so each part runs on the right brain.

One click next to the input: the model, the effort level, more models. The habit: reasoning → powerful, execution → fast.

— and notice the labels: "most capable for your hardest tasks," "efficient for routine," "fastest for quick answers." The picker itself is the principle of this page.

/model — and notice the labels: "most capable for your hardest tasks," "efficient for routine," "fastest for quick answers." The picker itself is the principle of this page.

The HackYourFuture curriculum is licensed under CC BY-NC-SA 4.0 *https://hackyourfuture.net/*

CC BY-NC-SA 4.0 Icons

Built with ❤️ by the HackYourFuture community · Thank you, contributors

Found a mistake or have a suggestion? Let us know in the feedback form.

Week 4