← All articles
Guide

Haiku 5.5, Sonnet 5.5, Opus 5.5 or Fable 5.2: how to choose a Claude model

A practical guide to Anthropic's four current Claude models, built from the company's own docs: what each costs, what it is for, and when to raise effort instead of switching models.

Anthropic now sells four current Claude models side by side: Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.2. They share the same context window and output limit, so the choice comes down to price, speed and how hard the work is. This guide pulls together what Anthropic's own pages say about each model, as of mid-November 2026, and turns it into a practical way to pick one. Everything here comes from Anthropic's published documentation and announcements; we have not run our own benchmarks.

The short version

  • Opus 5.5 is Anthropic's recommended starting point for most work, both in the Claude apps and in the API.
  • Fable 5.2 is for the hardest, longest technical projects, and costs twice as much per token as Opus 5.5.
  • Sonnet 5.5 suits well-defined tasks with a clear brief, where speed matters.
  • Haiku 5.5 is for high-volume, simple jobs such as sorting, extraction and routing.

Anthropic's models overview puts it plainly: if you are unsure, start with Opus 5.5, and move up to Fable 5.2 for demanding reasoning and long-running agent work, or when your tests of Opus 5.5 at higher effort still fall short.

What the four models have in common

According to Anthropic's models overview, all four current models accept text and images, reply in text, handle many languages and can use tools. Each offers a context window of 1M tokens and can write up to 128K tokens in one response, and all four list a reliable knowledge cutoff of June 2026. Thinking is adaptive on every model, and on Opus 5.5 and Fable 5.2 the overview describes it as always on.

That means you are not trading away long documents or big codebases when you choose a cheaper model. What changes is how much reasoning you get per request, how quickly answers arrive and what you pay.

Prices at a glance

The list prices below are for usage billed by the token through the Claude Platform, per million tokens of input and output, as shown in Anthropic's models overview and its knowledge-work guide:

  • Claude Haiku 5.5: from $0.10 input and $0.50 output. Anthropic's guide notes that prompts over 100k tokens are billed at $0.50 and $2.50.
  • Claude Sonnet 5.5: $2 input and $10 output.
  • Claude Opus 5.5: $4 input and $20 output.
  • Claude Fable 5.2: $8 input and $40 output.

Fable 5.2 also has an optional fast mode. Anthropic's launch post says it generates tokens up to 2.5 times faster and is priced at $16 and $80 per million tokens for input and output respectively. The same post says Fable 5.2 itself is up to 30% faster than Fable 5.1 and cheaper than its predecessor.

If you use Claude through a subscription rather than the API, prices per token matter less than usage limits. Anthropic's knowledge-work guide says higher-priced models use up plan limits faster, and that, apart from temporary launch offers, Fable 5.2 sits outside the built-in limits of Pro plans and of standard seats on Team and seat-based Enterprise plans, so using it there spends usage credits. At launch, Anthropic offered a promotion that let Fable count toward up to 50% of weekly limits on Pro plans and standard Team seats; check the current terms before relying on it.

Opus 5.5: the default for most work

Anthropic made Opus 5.5 the default model in the Claude apps on Pro, Max, Ultra and Team plans, and its guide from November 12 recommends it as the model to leave selected for research, analysis, writing, spreadsheets and decks.

The company's case rests on a few internal tests it describes:

  • Accuracy with figures: in a research task where an automated grader failed any report containing an invented number or quote, Anthropic says 16 of 18 Opus 5.5 reports passed.
  • Speed on long projects: in a mock merger analysis that involved building an Excel model and a presentation, Anthropic reports Opus 5.5 finished in 63 minutes, against 93 minutes for Opus 5, at half the cost.
  • Vision: Anthropic calls it its best Opus model so far for reading charts, screenshots and diagrams.

The guide also argues that a stronger model can be cheaper per finished task even if it costs more per token, because it needs fewer attempts and less back-and-forth. That is why it suggests teams who keep Sonnet as their everyday model and save Opus for big jobs should reconsider.

In the API, Opus 5.5 is aimed at long-running agentic coding and knowledge work, according to the models overview.

Fable 5.2: when the job is genuinely hard

Anthropic launched Fable 5.2 and its sibling Mythos 5.2 on November 9. The two share an underlying model. Fable carries stronger cybersecurity and biology safeguards so it can be offered to everyone, while Mythos is limited to organisations admitted through Anthropic's verification programmes.

Anthropic pitches Fable at large, open-ended technical projects, naming coding, finance and science as examples. Its launch post lists examples from prerelease testing, including large software migrations, designing electronic devices and progress on open physics problems. Its own advice is that most users should still start with Opus 5.5 and switch to Fable when a project's scope calls for it.

A practical way to read that: reach for Fable 5.2 when a task runs for hours, needs many steps or subagents, or has already defeated Opus 5.5 at high effort. For a memo, a spreadsheet or a typical bug fix, the extra cost is unlikely to pay off.

Sonnet 5.5: fast and capable for clear briefs

Sonnet 5.5 costs half as much as Opus 5.5 per token and Anthropic describes it as the best balance of speed and intelligence in the lineup. Its knowledge-work guide recommends it for well-scoped tasks with a clear brief, for example turning finished notes into a polished document or building slides from a template.

There is one catch worth knowing. Anthropic says that when Sonnet 5.5 runs at its higher effort settings, its cost per task gets close to that of Opus 5.5. So if good results only come when Sonnet runs at Extra high or Max effort, the company recommends moving to Opus 5.5.

Haiku 5.5: volume and latency

Haiku 5.5 is the fastest and least expensive model. The models overview lists classification, extraction and routing as its main jobs, and the knowledge-work guide adds quick lookups, summaries and sorting data. Anthropic says this is the first Haiku release to offer effort controls, which lets you request a more careful pass before trying a larger model.

Anthropic's cost guide suggests trying Haiku 5.5 first for high-volume work where you can check the output. Keep an eye on prompt length: above 100,000 tokens, every token of a Haiku request costs more, and the guide recommends compacting long conversations before they cross that line.

Effort: the dial to try before switching models

Every current model has an effort setting, which controls how long it thinks. The models overview lists the defaults as high for Fable 5.2 and Sonnet 5.5, and medium for Opus 5.5 and Haiku 5.5.

Anthropic's advice is consistent across its pages: if a result disappoints, raise the effort first. Only when the model had the right context, made a real attempt and still missed should you switch to a bigger one. For Opus 5.5, the guide says medium effort matched or beat Opus 5 at high effort on Anthropic's coding and knowledge-work evaluations, so it suggests starting at medium and stepping up only when that stalls.

The same guide recommends dropping old prompt habits. Phrases such as asking the model to think hard are no longer needed with Opus 5.5, since it decides how much to think from the request and the effort level, and Anthropic says removing them made replies start sooner without a clear loss in quality.

Cutting costs without changing models

For developers, Anthropic's guide on balancing cost and intelligence lists several savings that do not touch quality:

  • Prompt caching: this is usually the biggest lever, in Anthropic's measurements. Cached input is read at a fraction of the normal price: a twentieth on Opus 5.5 and Sonnet 5.5, and a fortieth on Fable 5.2, compared with a tenth on Haiku 5.5.
  • Batch processing: work that can wait up to 24 hours runs at 50% off.
  • Output caps: if responses are being cut off, the guide suggests raising the limit to 64,000 tokens, noting that in Anthropic's tests this did not raise cost per solved task.
  • Re-running failures: when you can verify results with tests, running at low effort and repeating only the failures at high effort came to roughly 40% cheaper than using high effort throughout, with an unchanged pass rate, according to Anthropic.

Anthropic stresses that these figures come from its own internal tests and are directional, so it recommends measuring on your own workload.

A simple decision path

  1. Start with Opus 5.5 at its default medium effort.
  2. If the task is short, well defined and repetitive, try Sonnet 5.5, or Haiku 5.5 for very high volumes, and check the output.
  3. If Opus 5.5 struggles, raise the effort to high before anything else.
  4. If it still falls short, or the job is a long, multi-step technical project, move to Fable 5.2, and consider fast mode only when waiting costs you more than the higher price. Anthropic says fast mode is offered in Claude Code and on the Claude Platform, among other places.
  5. In the API, turn on prompt caching from day one and compare models on cost per completed task rather than price per token, as Anthropic's cost guide advises.

Availability

All four models are available on the Claude Platform, and the models overview lists IDs for Amazon Bedrock, Google Cloud and Microsoft Foundry. The API names are claude-haiku-5-5, claude-sonnet-5-5, claude-opus-5-5 and claude-fable-5-2. Anthropic says it will not retire Opus 5.5 sooner than September 22, 2027, or Fable 5.2 sooner than November 9, 2027, which gives teams building on them a predictable runway.

Prices, limits and promotions change often, so confirm the current figures on Anthropic's pricing page before you commit a budget.