Reference

Which Claude model should you use in Claude Code

Updated Jul 10, 2026 4 min

I get asked "which model should I use" more than almost anything else, so here is the short version of how I pick. There is no single best model. The right one depends on how hard the task is, how fast you want it, and what you are paying. Here is the whole lineup in one place, and the rules I actually use.

The lineup at a glance

Model Alias Best for Relative cost Speed
Fable 5 fable The hardest reasoning and longest autonomous runs Highest Slowest
Opus 4.8 opus Complex coding and agentic work High Medium
Sonnet 5 sonnet The everyday balance of quality and speed Medium Fast
Haiku 4.5 haiku Quick, well-scoped tasks and cheap sub-agents Lowest Fastest

Capability and cost move together. Fable 5 is the most capable and the most expensive; Haiku 4.5 is the fastest and cheapest. Sonnet 5 sits in the middle and is where most of my work happens.

What each one actually costs

If you use Claude Code through a Pro or Max subscription, you are not paying per token, you are spending against your plan's usage and rate limits. Picking a lighter model on a heavy task stretches those limits further and comes back faster. If you connect an API key instead, you pay real money per token, and the gap is large:

Model Input ($/million tokens) Output ($/million tokens)
Fable 5 $10 $50
Opus 4.8 $5 $25
Sonnet 5 $3 (introductory $2 through 2026-08-31) $15 (introductory $10)
Haiku 4.5 $1 $5

Those are the list prices per million tokens on the Claude API. The headline number: Opus costs five times what Haiku does, and Fable ten times. That ratio is the whole reason the orchestrator pattern works, where a capable model plans and cheap models do the bulk of the reading and searching.

Which one I reach for

  • Sonnet 5 for most work. It is fast, it is strong at coding, and it keeps a session responsive. Unless a task is clearly beyond it, I start here.
  • Opus 4.8 or Fable 5 when the task is genuinely hard. A tricky refactor, a long autonomous run, a problem where getting it right the first time matters more than speed. This is where the extra capability earns its cost.
  • Haiku 4.5 for narrow, well-defined jobs. Renaming across files, a quick lookup, a mechanical edit. It is also my default model for sub-agents that just read and report back, because they burn the most tokens and need the least brainpower.

The mistake I see most is leaving everything on the most capable model out of habit. Half of what you ask an agent to do is mechanical, and running mechanical work on Opus or Fable is the single easiest way to run up your usage for no gain in quality.

What Claude Code starts you on

Your default depends on your plan. Max, Team, Enterprise, and API accounts start on Opus 4.8; Pro accounts start on Sonnet 5. Either way you can change it any time.

How to switch models

  • In a session: run /model, or /model sonnet to jump straight to one. Your choice is saved as the default for next time.

    /model haiku

  • At launch: pass the flag.

    claude --model sonnet

  • As an environment variable (handy for scripts and per-project shells):

    export ANTHROPIC_MODEL=claude-sonnet-5

  • In settings.json: set the model field so a project always opens on the model you want.

The short aliases (opus, sonnet, haiku, fable) are version pointers that follow the latest release. If you need a build to stay fixed, use the full model ID (claude-opus-4-8, claude-sonnet-5, claude-haiku-4-5, claude-fable-5) instead.

The 1M-token context option

Opus and Sonnet also offer a one-million-token context window through the opus[1m] and sonnet[1m] aliases. It is not gated behind a beta and, on the current models, it does not cost a premium per token: a request that fills 900k tokens is billed at the same rate as a small one. That said, a bigger window is not free of downsides. More tokens in context means more to process each turn and a real risk of context rot, so reach for the 1M window when a task truly needs it, not as a default.

Where to go next