Choosing a Claude Model: Don’t Ask Jensen Huang to Fetch Coffee

The most expensive model is not always the cheapest choice

When people start vibe coding, one thought comes naturally: “This is coding, so I should always use the best model.” But using a frontier model to rename one file or find one line in a long log can spend more tokens and time without materially improving the result.

Think of it as asking Jensen Huang, NVIDIA’s founder and CEO, to fetch coffee. The point is not that he cannot do it. The point is that his judgment and time are better reserved for work that needs them. AI models work the same way. Choosing the model that fits the current job matters more than choosing the biggest model every time.

This post covers Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5, and Claude Code’s effort setting as of Oct 10, 2026. Model availability, prices, and defaults can change, so check Claude Code’s /model screen and Anthropic’s official documentation before relying on them.

Think of the four models as a company team

Instead of memorizing which model is “better,” it is easier to think about the job each one is suited to.

• Fable 5.1: the executive for a long-running project. It fits large, ambiguous work that runs for hours, architecture spanning several codebases, and root-cause investigations.

• Opus 5.5: the team lead for complex judgment. It fits system design, difficult bugs, and reviews where a mistake would be expensive.

• Sonnet 5.5: the experienced operator for most day-to-day work. It balances speed and cost on well-scoped features, ordinary bug fixes, tests, and documentation.

• Haiku 5.5: the fast operator for repetitive work. It fits search, classification, summaries, data lookups, and small edits.

One practical arrangement is to have Opus plan, Sonnet and Haiku execute, and Fable advise on hard problems. It is not an absolute ranking. Anthropic describes Fable 5.1 as its most capable generally available model for the hardest and longest-running work. If the whole project is large enough, Fable may be the right model to lead it from the start.

Claude Code does not automatically split every task among these four models, either. When you switch with /model, subagents that inherit the main session normally use that model too. To keep a particular subagent on a smaller model, you need to set its model explicitly or configure an agent team. For a beginner, changing /model at clear stage boundaries is easier to understand than building an automatic routing system first.

Adjust effort before reaching for a bigger model

Effort controls how deeply a model reasons at each step. Lower effort is faster and uses fewer tokens. Higher effort checks more edge cases and gives the model more room to make decisions on its own.

• low: brainstorming, renaming, short questions, and work you will review immediately

• medium: well-scoped everyday coding; the Claude Code default for Opus 5.5, Sonnet 5.5, and Haiku 5.5

• high: fixing bugs in an existing codebase, migrations, and pre-deployment work where verification matters

• xhigh and max: hard problems such as security investigations or long autonomous work; test them on your own tasks because extra thinking can have diminishing returns

Starting at medium and moving up only when the result is insufficient or the cost of a mistake is high is more practical than locking every session to high or max. If the model itself is oversized for the task, however, switching from Opus to Sonnet or Haiku can save much more than lowering Opus’s effort alone.

How to switch in Claude Code

The simplest method is to use commands in the conversation:

/model
/model haiku
/model sonnet
/model opus
/model fable
/effort medium

/model opens the model picker. Pressing Enter selects the model and saves it as the default for later sessions. Pressing s on a model row applies it only to the current session. You can also choose the model and effort when launching Claude Code:

claude --model sonnet --effort medium

The exact versions behind the opus, sonnet, and haiku aliases can change with your provider and Claude Code version. On the Anthropic API as of Oct 10, 2026, they resolve to Opus 5.5, Sonnet 5.5, and Haiku 5.5. Unless you have a specific compatibility reason to pin an older model ID, the aliases are easier to maintain.

Measure total cost, not just token price

If you use the API directly, the price differences are large. At Anthropic’s standard prices on Oct 10, 2026, input/output prices per million tokens are $10/$50 for Fable 5.1, $4/$20 for Opus 5.5, $2/$10 for Sonnet 5.5, and $0.10/$0.50 for Haiku 5.5 when the Haiku prompt is 100K tokens or less. A Claude subscription may not bill each request at these API prices, but using more tokens can still consume the plan’s usage allowance faster.

The cheapest model per token is not always the cheapest way to finish a task. If Haiku needs several attempts to solve a complex bug that Sonnet or Opus would fix once, the total cost can become higher. Do not measure only the price of one token. Measure how many attempts, corrections, and human review cycles it took to reach a correct result.

A beginner-friendly selection ladder

You can reduce waste without building a sophisticated router:

• Step 1: define the task in one sentence. Replace “build a login page” with something like “add CSRF protection and a one-hour inactivity timeout to this PHP login page without breaking its existing tests.”

• Step 2: start classification, summaries, search, and small repetitive edits on Haiku at low or medium effort.

• Step 3: start well-scoped features and everyday bug fixes on Sonnet at medium effort.

• Step 4: move to Opus at medium or high effort when the cause is unclear or the work needs architectural judgment and deeper verification.

• Step 5: use Fable at low or medium effort for work spanning many systems and files that must run autonomously for a long time, then raise effort only when needed.

Do not jump to the biggest model simply because one attempt failed. First check whether the request was vague, whether a required file was missing, or whether you forgot to define what “done” means. A bad instruction given to a more expensive model can simply produce a more expensive wrong answer.

The model choice is part of vibe coding

The answer is neither “always Fable” nor “always Haiku.” Send repetitive work to a light model, well-scoped features to Sonnet, complex judgment to Opus, and ambitious long-running projects to Fable. Then adjust effort from low to medium or high to balance quality and token use.

The goal is not to select the highest tier. It is to use only the level of intelligence the job needs, at the moment it needs it. That is how you avoid asking Jensen Huang to fetch coffee while also getting better results from Claude Code with less wasted usage.