Choosing a Model and Effort in Claude: Which of Opus, Sonnet, Haiku and Fable for Which Job?

Version note: This post covers choosing a model and effort in the Claude apps as of 15 November 2026; the model list, prices and plan rules were checked the same day in the Claude Help Center, the models overview page on platform.claude.com and Anthropic's Claude Fable 5.2 announcement of 9 November 2026. This post is based on the official documentation. The options in the model picker can look different depending on your plan and your admin's settings.

Guess First

Emre is a developer on the team behind Bookstore, the online bookstore used as the shared example across this site's series. When he moved to the Max plan, he picked Fable 5.2 in the model picker, set effort to "Max" and did every job this way for a week: writing commit messages, translating an error message, explaining a SQL query, drawing up a refactoring plan for the ordering module. By Wednesday the Fable part of his weekly limit was used up.

Two questions: which setting burned more of the limit, the model or the effort? And what would be a sensible combination for small jobs like a commit message? We look for the answer in Anthropic's own advice; that advice says that when you change settings, you start with effort, not the model.

The Model Family (November 2026)

On 15 November 2026 there are four current Claude models. The models overview page on platform.claude.com (https://platform.claude.com/docs/en/about-claude/models/overview) describes each in one sentence; the prices are the Claude API price in dollars per million tokens (MTok). In the apps you do not pay per token, but these prices are a good sign of how "heavy" the models are relative to each other; a more expensive model also uses up your limit faster.

All four have a context window of 1 million tokens, a maximum output of 128 thousand tokens per response and a reliable knowledge cutoff of June 2026. So the difference between them is not the size of the context; it is reasoning power, speed and cost. The release history is fast too: Opus 5.5 came out on 22 September, Sonnet 5.5 on 28 September, Haiku 5.5 on 7 October and Fable 5.2 on 9 November 2026. Older models may also appear in the model picker under "More models"; for Fable, versions 5.2, 5.1 and 5 are listed separately.

The Model Picker and the Effort Setting

According to the Help Center article "Change the model, effort, and thinking settings" (https://support.claude.com/en/articles/8664678), all three settings are in the model menu next to the send button. Click the model name and choose "Effort" to see the levels: Low, Medium, High, Extra high and Max. Each model's recommended level is marked "Default" in the menu; Anthropic's Fable 5.2 announcement says the default in the apps is Medium. Extra high exists only on Opus 4.7 and newer models. The effort setting is offered on Fable 5.2, Opus 5.5, Sonnet 5.5, Haiku 5.5 and a range of models going back to the 4.6 generation.

Effort decides how much thinking Claude puts into each response. A higher level gives a more thorough answer but takes longer and uses more tokens, so you get closer to your limit faster. Thinking is a separate switch, inside the Effort submenu. On the current 5.x models thinking cannot be turned off; on Fable 5.2 and Opus 5.5 it is always on, so the switch does not appear at all. On these models the way to reduce thinking is to lower effort. By expanding the "Thinking" section above an answer you can see a summary of Claude's reasoning; according to the documentation, part of it may be withheld for safety reasons.

The model, effort and thinking settings can be changed in the middle of a conversation, and the change applies from the next response. This is a small but useful detail: in one conversation you can first scan a document quickly with Sonnet and then switch to Opus for the single question that needs a decision.

There is an admin side too. On Enterprise an admin can turn off certain models or effort levels for a role, and decide which model and effort new conversations start with. If an option you expect is missing from the menu, this is most likely why.

Effort First, Then the Model

Anthropic's Claude Academy guide "Choosing a Claude model for knowledge workers" (https://academy.claude.com/tutorials/choosing-the-right-claude-model) recommends starting with Opus for most coding and knowledge work; if your plan or organization starts you on Sonnet, it says to start there and raise effort first. When you change settings it recommends changing one thing at a time, starting with effort, so you can see what made the difference.

The usage argument is in the same guide: a more capable model costs more per token, but because it can finish a hard job in fewer turns, its cost per task can come out lower. If you want to protect your limit, the guide's advice is to lower effort first; this reduces spending while keeping the model's reasoning quality. Back to Emre: jobs like a commit message or translating an error message needed neither Fable nor Max effort. What burned the limit was both settings sitting at the top together; but if you have to start with one knob, effort is the first to come down.

Fable: The Strongest Model, Separate Rules

Fable models are counted differently from the others in the Claude apps. According to the Help Center article "Claude Fable models on your plan" (https://support.claude.com/en/articles/15424964):

• Fable is not available on the Free plan.
• On Pro and on a Team standard seat, Fable is not included in the plan limit; it is taken from usage credits from the first message. On Pro you turn usage credits on yourself; on Team the organization must have turned them on.
• On Max, Ultra and a Team premium seat, Fable can use at most half of the weekly limit. This does not add 50 percent to the limit; it draws from the same pool as the other models and uses it up faster.
• When you reach this cap you either continue with Fable on usage credits or switch to another model.

There is also a temporary offer that came with the launch on 9 November 2026: Pro and Team standard users can use Fable until 16 November, included up to half of the weekly limit; in addition, Pro and Team standard get one "limit reset", and Max, Ultra and premium seats get between one and three depending on how long they have been subscribed. A reset is used from "Limit resets" in Settings > Usage and expires on 8 March 2027. The default model does not change; you have to pick Fable in the picker yourself.

Two more details are worth knowing. According to Anthropic's announcement, Fable 5.2 flags some requests in cybersecurity and biology as a safeguard and routes them to a less capable model; flagged cybersecurity questions get an answer from Opus 4.8, biology questions from Opus 5. The same announcement also says that for customers without a zero data retention agreement, Fable models have 30-day retention for safety monitoring. We cover privacy settings in the last post of the series.

Job → Model Mapping

Turning the descriptions in the documentation into the Bookstore team's daily jobs gives the following starting point. It is not a rule; it is a first guess that you will correct as you try it on your own work:

• Translating an error message, writing a commit message, explaining a short SQL query: Haiku 5.5 or Sonnet 5.5, Low or Medium effort.
• A user story draft, an e-mail, explaining code, reading a medium-length document: Sonnet 5.5, default effort.
• Multi-step analysis, long architecture documents, weighing several files together and proposing a decision: Opus 5.5, High if needed.
• Work that lasts days, with high uncertainty and many steps, where Opus 5.5 falls short even at high effort: Fable 5.2.

The models overview page gives a similar order: if you are unsure, start with Opus 5.5, and use Fable 5.2 for demanding reasoning and long-horizon agentic work, or when Opus 5.5 falls short even at higher effort.

Bookstore: Trying the Return Policy Draft with Two Settings

Selin wants the first draft of Bookstore's customer-facing return policy. It is a good test subject: it needs both faithfulness to company rules and clear writing. We run the same brief with two settings and compare them: Sonnet 5.5 at default effort and Opus 5.5 at High. The brief:

Draft the customer-facing return policy for Bookstore, an online bookstore.

Rules (do not add or change any):
- Customers can return a book within 14 days of delivery.
- Damaged or wrong books: free return shipping, full refund.
- Other returns: the customer pays return shipping.
- Institutional orders (20 or more copies) can be returned in part;
  the institution's approver must confirm the return.
- E-books cannot be returned once downloaded.

Output:
- Plain language, at most 300 words, short headings.
- A "Questions we could not answer" list at the end for anything
  the rules above do not cover. Do not invent policy to fill gaps.
- Write it in Turkish.

Two parts of the brief are deliberate: we say explicitly not to go beyond the rules, and instead of filling gaps we ask for a list of "questions we could not answer". A good answer is then one that notices the gaps in the rules (for example who returns a book bought as a gift, or by which method the money comes back) and writes them in the list; a bad answer fills those gaps with rules it made up. The draft is not a legal text; before publishing it has to be checked by the company's legal team and against the statutory withdrawal rules.

The Claude Academy guide recommends this method for a comparison: define what a good result is before you try, run both models with the same prompt and the same files, and count how many follow-up turns and corrections each needs. Putting this into a small table helps:

You can tell which setting is enough for this job with this table, not with a guess; you have to measure the result in your own account, with your own work. Measuring the same job automatically through the Claude API with different models and effort levels is the subject of the upcoming "Building Applications with the Claude API" series.

In Production

The model list changes. During 2026 many models came out, from Opus 4.6 to Fable 5.2, and old models left the menu; for example, Opus 4 and 4.1 were removed from the model selector in January 2026. On the API side, the retirement date of Claude Sonnet 4.5 is 30 November 2026 and the recommended replacement is Sonnet 5.5 (https://platform.claude.com/docs/en/about-claude/model-deprecations). If your team documents say "use this model for this job" by model name, you will need to update them as models come out; a more durable way is to write the job type and the effort, and to refer to the model by family, such as "the current Sonnet".

The admin sets the team default. On Enterprise the starting model and effort of new conversations come from the organization settings; users can change them within a conversation. Setting the default to the strongest model means every small job of every person starts expensive.

Retry your common jobs when a new model comes out. The Academy guide recommends this too: a new model may give the same result at lower effort, and that directly eases the team's limit.

Common Pitfalls

Giving every job to the biggest model. The strongest model is not the best choice for every job; on small jobs it is slower and uses up the limit faster.

Treating effort as a quality knob. Effort does not guarantee a correct answer; it changes the budget for thinking. Max effort does not rescue a badly described task; we talk about task briefs in the third post.

Memorizing model names. The list in the picker changes by plan, organization and date. A model you saw in a blog post or an old team note may no longer be there; check the current list in the picker and on the models overview page.

Thinking Fable raises your quota. The "50 percent" on Max and Ultra is a cap, not an extra share of the limit. Every hour spent with Fable comes out of the pool the other models can use.

Comparing with a single try. A model that gave a better answer once is not always better. Try at least one short and one long job, with the same brief and criteria written in advance.

A Small Experiment

★☆☆ Open the model menu and note which models and which effort levels appear in your account. Is the level marked "Default" the same for every model?

★★☆ Run the return policy brief from this post with Sonnet 5.5 at default effort and Sonnet 5.5 at High. Fill in the evaluation table for both tries. Which criterion did raising effort change?

★★★ List the five jobs your team does most often and propose a starting model and effort level for each. Work with this mapping for a week, compare the weekly bar in Settings > Usage with the previous week, and correct the mapping.

The right model and effort make a difference on a well-described job. In the third post we look at what to write into Instructions for Claude, project instructions and skills, the parts of a good task brief, and habits for verifying Claude's answers, using a Bookstore campaign e-mail.