Which model to use for which task?

Every month a new model comes out claiming to be "the best". The question that actually matters is not that one — it is where the data needs to live and how long the task is.

The commercial landscape today

Claude is already on its fifth generation: Haiku 4.5 for the fast and cheap, Sonnet 5 as the quality/cost balance point, Opus 5 for complex agentic code, and Fable 5 at the top. OpenAI unified its code stack under GPT-5.6 — Codex is no longer a separate product, it's the same stack with a specialized mode (GPT-5.3-Codex, with Max variants for long work and Mini for cost). Gemini plays the context card: its 3.1 Pro and 3 Deep Think variants reach windows of up to 1M tokens, useful when you need to feed an entire document in one sitting.

And the local ones

This is where this site takes a firm position: if the task doesn't require absolute frontier reasoning, a local model handles most of the daily work at zero marginal cost. Qwen3.6 27B runs on one or two consumer GPUs and has 262K of native context — that's what runs this very site.

The real matrix

• Sensitive data (clients, contracts, health): local, no debate.
• Large refactor or long-format agentic coding: Claude Opus or GPT-5.3-Codex-Max.
• Huge documents in one sitting: Gemini, for the context window.
• Everything else — chat, summaries, day-to-day tasks: the cheapest local model that solves it well.

The most common trap is using the most expensive model 'just in case' for trivial tasks. Most of what you do every day is handled just as well by an 8 to 27B model running on your own hardware.