Claude Models: Fable, Opus, Sonnet and Haiku Compared

Claude has four current models: Fable 5.1 ($10 in, $50 out per million tokens), Opus 5.5 ($4 and $20), Sonnet 5.5 ($2 and $10) and Haiku 5.5 (from $0.10 and $0.50). All four read text and images, write text, and have a 1M-token context window. They are closed-weight, so you reach them through an API, not a download. Checked 31 October 2026.

The lineup at a glance

Anthropic's models overview describes each model in a single line, and that line is the quickest way to see how the family is meant to be used.

Every row comes from Anthropic's models overview, pricing page and the four model pages. Two Anthropic models sit outside the table. Claude Mythos 5.1 is listed on the pricing page as limited availability, and we do not cover it here. Older models such as Opus 4.8 and Sonnet 4.6 are still served and are listed as legacy.

Who makes it and how it is licensed

Anthropic makes Claude. The weights are not published. There is no open licence, no Hugging Face repository and no way to run these models on your own GPU. If open weights are a requirement, read open source LLMs instead, or the open versus closed models comparison.

What you do get is a choice of where the model runs. Anthropic lists six platforms on each model page: the Claude API, Amazon Bedrock, Claude Models on Google Cloud (Google managed), Microsoft Foundry, Claude Platform on AWS, and Claude Models on Google Cloud (Anthropic managed). Anthropic sets the retirement date for the platforms it operates and says partner-operated platforms set their own.

Modalities, context and output

All four models accept text and images and return text. Anthropic's overview says all current models support multilingual use, vision and tool use. The context window is 1M tokens on every model, which Anthropic converts as roughly 555,000 words on the current tokenizer. Synchronous output tops out at 128K tokens. On the Message Batches API, Opus 5.5, Sonnet 5.5 and Haiku 5.5 can reach 300K output tokens with a beta header named output-300k-2026-03-24.

The reliable knowledge cutoff for all four is June 2026.

Two things to know about thinking. All models use adaptive thinking, where the model decides how much to think and the effort setting steers it. On Fable 5.1 and Opus 5.5 it is always on. Sonnet 5.5 accepts a lower setting called between_tools that turns off up-front thinking, and the docs say a plain disabled request returns a 400 error on Sonnet 5.5 and Opus 5.5. Haiku 5.5 still accepts disabled at high effort or below, though Anthropic says effort is the better lever.

Prices, checked today

All prices come from the Claude pricing page on 31 October 2026, in USD per million tokens.

Cache reads cost 10 percent of the base input price on Haiku, 5 percent on Opus 5.5 and Sonnet 5.5, and 2.5 percent on Fable 5.1. The Batch API takes 50 percent off input and output. Opus 5.5 also has a fast mode priced at $8 and $40.

Three modifiers matter for a bill. US-only inference with inference_geo: "us" adds a 1.1x multiplier on every category. Haiku 5.5 is the only model priced by prompt length: a request over 100,000 input tokens, counting cache reads and writes, pays the higher rate for that whole request. And Sonnet 5 at $2 and $10, which launched with introductory pricing, kept that price: the planned rise to $3 and $15 on 1 September 2026 did not happen.

For plan prices and the full bill-estimation maths, see Claude API pricing and Claude pricing.

Where it is served

Claude runs on Anthropic's own API and on several clouds. Of the hosts in our directory, the relevant one is AWS Bedrock, where Opus 5.5 uses the ID anthropic.claude-opus-5-5. Bedrock and Google Cloud's regional endpoints include a 10 percent premium over global endpoints. For the wider hosting picture, see AI hosting.

Note that Claude Code's aliases do not resolve the same way everywhere. On the Anthropic API, opus, sonnet and haiku point at Opus 5.5, Sonnet 5.5 and Haiku 5.5. On Amazon Bedrock and Google Cloud's agent platform, the docs list sonnet as Sonnet 4.5 and haiku as Haiku 4.5. On Microsoft Foundry, opus is Opus 4.6. If you run Claude Code through a cloud, pin the full model name.

Which coding agents use Claude

Claude Code is Anthropic's own agent and uses Claude only. Its /model command takes the aliases fable, opus, sonnet, haiku, best and a mode called opusplan, which uses Opus in plan mode and Sonnet to execute. See Claude Code for the agent itself.

Other agents offer Claude as one choice among several. GitHub Copilot lists Claude Fable 5 and 5.1, Haiku 4.5 and 5.5, Sonnet 4, 4.6, 5 and 5.5, and Opus 4.8, 5 and 5.5. Cursor has model pages for Sonnet 5.5, Opus 5.5, Fable 5.1 and Haiku 5.5. Cline and OpenCode list Anthropic among their providers. Aider names Claude support in its changelog, though its newest Claude entry that we read was for the 4.5 and 4.6 generation, and its repository's last push was on 22 May 2026, so we found no Claude 5.x aliases in its changelog.

Published benchmarks

Anthropic publishes benchmark tables on each launch page. We reproduce the Opus 5.5 table on the Claude Opus page, with the footnotes, and the Sonnet 5.5 and Haiku 5.5 tables on their own pages. Every number is vendor-reported by Anthropic, and some columns for rival models are, in Anthropic's words, "as reported by OpenAI." We have not run any of these.

Two cautions from Anthropic's own text deserve repeating. The Opus 5.5 launch post says that "benchmark margins have become a less reliable guide to real-world differences." And it says the gap between Opus 5.5 and Fable 5.1 is narrower in Anthropic's own use than the scores suggest. Treat the tables as a starting point for your own evaluation.

Local hardware needs

None. Claude models cannot be run locally. The closest local option is a different family, such as Qwen or gpt-oss, which we cover on their own pages.

Retirement and versioning

Every Claude model ID is a pinned snapshot, including the dateless IDs used from the 4.6 generation on. Anthropic commits to retirement dates on the platforms it runs: not sooner than 1 September 2027 for Fable 5.1, 22 September 2027 for Opus 5.5, 28 September 2027 for Sonnet 5.5, and 7 October 2027 for Haiku 5.5. A year of notice is long enough to plan a migration, but the 5.5 models also bring API changes. Opus 5.5 and Sonnet 5.5 reject forced tool use (tool_choice of any or a named tool returns a 400 error). They also tie thinking blocks to the model that wrote them, and Haiku 5.5 rejects temperature, top_p and top_k. Check Anthropic's migration guides before swapping a model ID in production.

How the four relate

Think of the family as three steps and a specialist. Haiku is the cheap, fast model for work that is small and repeated. Sonnet is the daily driver. Opus is the model for long agent runs. Fable sits on top for the hardest reasoning, at two and a half times Opus's price.

Anthropic's launch posts describe the middle two as a pair. The Sonnet 5.5 post calls it "a faster, lower-cost complement to Claude Opus 5.5" and says Opus 5.5 "remains clearly stronger at complex, open-ended work requiring sustained judgment." The Haiku 5.5 post says it "pairs well with Opus 5.5 and Sonnet 5.5 as a subagent on coding work." That framing suggests a design where one larger model plans and smaller models carry out narrow tasks. For a deeper look at the first pair, read Claude Opus vs Sonnet.

What the 5.5 generation changed

Anthropic released the 5.5 models over about two weeks: Opus 5.5 on 22 September, Sonnet 5.5 on 28 September and Haiku 5.5 on 7 October 2026. The launch posts make a few claims worth knowing, all vendor-reported.

For Opus 5.5, Anthropic says it performs at the level of Fable 5.1 on most work and costs 40 percent less to run than Opus 5 on typical workloads. The list price fell 20 percent, from $5 and $25 to $4 and $20, and cache reads fell 60 percent. Anthropic also says Opus 5.5 writes output more than 30 percent faster than Opus 5.

For Sonnet 5.5, the post says it runs 30 percent or more faster than Sonnet 5 and costs up to 30 percent less per task, at the same $2 and $10 list price, because it needs fewer tokens. Cache reads were halved from $0.20 to $0.10.

For Haiku 5.5, Anthropic says it costs around 75 percent less to run on average than Haiku 4.5 (which was $1 and $5), and it is the first Haiku with an adjustable effort setting. The trade is a larger token count for the same text, about 30 percent more than on Haiku 4.5, because of the newer tokenizer.

Read those as the vendor's description of typical workloads. Your prompts will differ, and cost per task depends on effort setting, caching and how many turns an agent takes.

Moving between models

Because the context window, output limit and tool support line up, switching inside the family is mostly a one-line change. Three things catch people.

First, thinking blocks. Each block records the model that wrote it. Opus 5.5 reads blocks from Opus 5 and earlier Opus, Sonnet and Haiku models, and, on the Claude API and Google's managed platform, from Sonnet 5.5 and Haiku 5.5, but not from Fable or Mythos. If you switch mid-conversation to a model that cannot read the blocks, the API drops them before the model sees the request.

Second, token counts. The same text can count as roughly 30 percent more tokens on the newer tokenizer than on older Haiku and Sonnet generations, so recount before you trust an old cost estimate.

Third, sampling. Sonnet 5.5 returns a 400 error if you set temperature, top_p or top_k, and Haiku 5.5 asks you to omit them. If your code sets them for every model, branch on the model name.

Who should pick which

Our take

The most useful thing about the Claude line is how cleanly it is priced. Each step down is a clean halving or better, and the context window and output limit are the same everywhere, so you can move between models without rewriting your chunking logic. The 5.5 generation also made caching much cheaper: Opus and Sonnet cache reads are 5 percent of input, which matters because agent loops resend the same prefix on every turn.

The costs of that tidiness are the closed weights and the API changes between generations. We would pick Sonnet 5.5 as the default for most teams, test Opus 5.5 on the tasks where Sonnet needs two or three retries, and keep Haiku 5.5 for the pipeline stages that do not need judgement. We have not benchmarked these models ourselves, so this advice rests on Anthropic's published positioning and the price list.

For comparison with other vendors, see GPT models, Gemini models and the side-by-side LLM API prices page. To compare an open-weights challenger directly, DeepSeek vs Claude covers price and coding use.

Checklist before you pick

• Count tokens on a real prompt, since the tokenizer counts about 30 percent more than older Haiku and Sonnet generations did.
• Turn on prompt caching for any prefix you resend, and check the 512-token minimum.
• Use the Batch API for anything that can wait, at half price.
• Set effort on purpose. Defaults differ: Opus and Haiku default to medium, Sonnet and Fable to high.
• Pin full model IDs in production, not aliases.
• If you need US-only processing, budget the 1.1x multiplier.
• Recheck the pricing page before you commit, as this table is dated 31 October 2026.

Questions we hear

Is Opus always better than Sonnet? Not on every task. Anthropic says Sonnet 5.5 at max effort is comparable to Opus 5.5 on several evaluations, and that Opus is clearly stronger on complex, open-ended work.

Can I use Claude models for free? The API is pay-per-token. Plan access on claude.ai is a separate matter, covered on Claude AI plans.

Do I need Fable? Anthropic's advice is to try it when your evaluations on Opus 5.5 at higher effort still fall short.