Claude Models Explained: Opus, Sonnet and Haiku

As of September 30, 2026, Anthropic lists four current Claude models: Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 4.5. Think of them as families on a ladder. Haiku is the fastest, Sonnet balances speed and intelligence, Opus is the one Anthropic suggests starting with for most work, and Fable sits above them for the hardest long-running jobs. A fifth model, Claude Mythos 5.1, exists but is invite-only. Everything older, Sonnet 5 included, is now a legacy model.

Model names change every few months, so treat this page as a snapshot we recheck monthly. The source of truth is Anthropic’s models overview↗ (opens in a new tab); keep it open next to anything you read about Claude models, this page included.

The current Claude models at a glance

• Claude Fable 5.1 (claude-fable-5-1), released September 1, 2026: “for demanding reasoning and long-horizon agentic work.” The slowest of the four. 1M-token context window, 128K-token maximum output, reliable knowledge cutoff June 2026.
• Claude Opus 5.5 (claude-opus-5-5), released September 22, 2026: “for long-running agentic coding and knowledge work.” Moderate latency, 1M context, 128K output, cutoff June 2026.
• Claude Sonnet 5.5 (claude-sonnet-5-5), released September 28, 2026: “the best combination of speed and intelligence.” Fast, 1M context, 128K output, cutoff June 2026.
• Claude Haiku 4.5 (claude-haiku-4-5-20251001), released October 15, 2025: “the fastest model with near-frontier intelligence.” 200K context, 64K output, reliable knowledge cutoff February 2025.

All four take text and images in and give text out, and all four support tool use. The legacy models that are still available are Fable 5, Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, Opus 4.5, Sonnet 5, Sonnet 4.6 and Sonnet 4.5.

The Claude model families

Claude Opus

Opus is the family most people mean when they search for “Claude Opus”: the model for work where being right matters more than being quick. Opus 5.5 is Anthropic’s default recommendation and the default model in Claude Code on Pro, Max, Team, Enterprise and the API. Its thinking is always on and its default effort is medium, one level lower than Opus 5.

Claude Sonnet

Sonnet is the everyday model: faster than Opus and lower in cost per token, suited to code generation, data analysis, content and agentic tool use where the task is clear. Sonnet 5.5 replaced Sonnet 5 as the current model two days before this page was published.

Claude Haiku

Haiku is the small, fast family, for real-time replies, high-volume classification and extraction, and subagents that only search or summarize. Haiku 4.5 is a generation older than the others, and it shows in the specifications: a smaller context window, half the output limit, an older knowledge cutoff and no effort setting. Anthropic lists no newer Haiku today.

Claude Fable and Claude Mythos

Fable is the newest family and the most capable one open to all customers. Anthropic points to it when Opus at higher effort still falls short. Mythos 5.1 is the same model as Fable 5.1, with the same specifications, offered only by invitation to Project Glasswing participants. Access goes through your Anthropic, AWS or Google Cloud account team; it is not in the ordinary model lists.

How to pick a Claude model

Anthropic’s guide to choosing a model↗ (opens in a new tab) offers two starting points. Efficiency-first: begin with Haiku 4.5, test, and move up only where it falls short, which suits prototypes, tight latency and high-volume work. Capability-first: begin with Opus 5.5, then lower the effort or move down a model once you know what the task needs. It also says tuning effort is often a better lever than switching models.

• Hard, open-ended or expensive-to-get-wrong work: Opus 5.5.
• Clear, well-specified work where speed matters: Sonnet 5.5.
• Volume, latency and simple subagent jobs: Haiku 4.5.
• Multi-hour autonomous work that Opus at high effort cannot finish: Fable 5.1.

Anthropic’s own pages do not quite agree here. The platform docs say most workloads should start on Opus 5.5, and Claude Code defaults to it, but a Claude Help Center article on models in Claude Code, updated September 22, 2026, still calls Sonnet “the right choice for the large majority of coding work.” The two-model decision has its own page: Claude Sonnet vs Opus.

Context windows, output limits and knowledge cutoffs

The context window is how much the model can consider at once, your prompt and its reply included. Anthropic’s overview puts 1M tokens at roughly 555,000 words on the current tokenizer, and 200K tokens at roughly 150,000 words. Bigger is not automatically better; what fills the window matters more, as AI context windows explained sets out.

Maximum output is the longest single reply on the Messages API, thinking included: 128K tokens for the three newest models and 64K for Haiku 4.5. The reliable knowledge cutoff is the date through which a model knows the world well. Anything after it, a library released last month or your team’s new API, the model only knows if you give it the documentation.

Thinking and effort

Thinking is the model reasoning before it answers. On Opus 5.5 and Fable 5.1 it is always on and cannot be switched off. Sonnet 5.5 uses adaptive thinking, deciding for itself how much to think; Haiku 4.5 uses the older extended thinking, which you switch on with a budget.

Effort is the dial for how much the model spends: low, medium, high, xhigh and max. The default is medium on Opus 5.5 and high on Sonnet 5.5 and Fable 5.1, and the scale is calibrated per model, so medium on Opus is not the same amount of work as medium on Sonnet. Haiku 4.5 has no effort setting. Higher effort gives more thorough answers and uses your usage limits faster.

Model IDs vs aliases

Anthropic’s page on model IDs and versioning↗ (opens in a new tab) clears up a common misunderstanding. Since the 4.6 generation, IDs have no date, and a dateless ID such as claude-opus-5-5 is a pinned snapshot, not a pointer to “the latest Opus.” Anthropic ships an update under a new ID. Older models carry a date, and the API accepts a short alias that points to it.

claude-opus-5-5              pinned snapshot (4.6 generation and later)
claude-haiku-4-5-20251001    dated ID (before 4.6)
claude-haiku-4-5             API alias for the dated Haiku ID
anthropic.claude-opus-5-5    Amazon Bedrock
claude-haiku-4-5@20251001    Google Cloud
opus, sonnet, haiku, fable   Claude Code aliases: follow the recommended version

The Claude Code aliases do move, and they resolve differently by provider: on Amazon Bedrock and Google Cloud, sonnet still means Sonnet 4.5. In production code, pin the full ID and change it on purpose. As for “Claude 4”: that generation began with Opus 4 and Sonnet 4, both retired on June 15, 2026, and its later Opus and Sonnet releases, 4.5 to 4.8, are the legacy models listed above. Haiku 4.5 is the one 4.x model still current.

Retirement dates and the model lifecycle

Anthropic’s model deprecations page↗ (opens in a new tab) defines four states: Active, Legacy (no more updates, may be deprecated), Deprecated (still works, has a replacement and a retirement date) and Retired (requests fail). It promises at least 60 days’ notice before a publicly released model retires, by email and in the docs, and those dates apply to Anthropic’s own platforms; Bedrock and Google Cloud set their own.

• Claude Haiku 4.5: not sooner than October 15, 2026.
• Claude Opus 4.5: not sooner than November 24, 2026.
• Claude Opus 5.5: not sooner than September 22, 2027.
• Claude Sonnet 5.5: not sooner than September 28, 2027.
• Claude Fable 5.1 and Mythos 5.1: not sooner than September 1, 2027.

“Not sooner than” is a floor, not a date: Sonnet 4.5’s floor was September 29, 2026, and it is still listed as Active. The two pages also label things differently. The overview calls Sonnet 5 and Opus 5 legacy, while the deprecations table shows them as Active. Read “legacy” as “still works, no longer updated, move when you can.”

Where you choose a Claude model

• Claude apps: the model name next to the send button opens the menu, with More models for the rest and Effort beside it, as Anthropic’s help article on model settings↗ (opens in a new tab) describes. On Enterprise, an administrator can turn models off for your role.
• Claude Code: /model opens the picker, claude --model sonnet sets one session, and "model" in .claude/settings.json sets a project default. The model configuration docs↗ (opens in a new tab) list every alias; Fable is never the default and can bill to usage credits depending on your plan and seat.
• The API: the model field on each request. Setup is covered in the Claude API guide.

A monthly model check for a team

1. Open the models overview and the deprecations page. Note any new model and any retirement floor within 90 days.
2. Search your code and config for every model ID you use, including subagent definitions and CI scripts.
3. For each ID near retirement, file a task to test the replacement on your own prompts before switching.
4. Pin IDs in production; leave aliases for interactive use.
5. Write the team’s default model and effort down once, where every person and assistant reads it.

On a fenbs board, step 5 is a rule on the Decisions and rules page, which every connected AI assistant reads first, and step 3 is an enhancement task per model. fenbs has no due dates, so put the retirement floor in the task title, such as “Move summaries off Haiku 4.5 (floor: October 15, 2026)”, and give it a priority that reflects how close it is.