Claude Model Comparison: Opus vs Sonnet vs Haiku (and Fable), Which for Which Task

TL;DR: This Claude model comparison, checked against Anthropic's docs on 17 September 2026, comes down to four current models. Fable 5.1 is the most capable and most expensive. Opus 5 is Anthropic's recommended default. Sonnet 5 handles everyday coding and writing at 40% of Opus's price. Haiku 4.5 is the cheapest, fastest pick for high-volume simple work.

Which Claude models are current in September 2026?

Four models make up Anthropic's current lineup: Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5 and Claude Haiku 4.5. That is straight from the comparison table on Anthropic's models overview, which we read on 17 September 2026. If you came here expecting "Opus vs Sonnet vs Haiku", the lineup has grown a tier above Opus, and that changes the answer for a small but expensive set of tasks.

A fifth name, Claude Mythos 5.1, appears on Anthropic's pricing page at the same price as Fable 5.1. Anthropic's model-selection guide says it offers the same capabilities to Project Glasswing participants only, so unless you are in that program, Fable 5.1 is the top model you can call.

Everything else is legacy. Fable 5, Opus 4.8, Opus 4.7, Opus 4.6, Opus 4.5, Sonnet 4.6 and Sonnet 4.5 are listed as "Legacy models (still available)". They still run, but none of them is where a new project should start, and older articles that call Opus 4.8 or Sonnet 4.6 "the latest" are simply out of date.

Here is the spec sheet, row by row, as Anthropic publishes it:

Two rows deserve a second look. The knowledge cutoff gap is large: Haiku 4.5's reliable knowledge stops in February 2025, more than a year before the other three. And Haiku's retirement floor is the nearest date on the whole table, which we come back to below.

Why does Anthropic tell you to start with Opus 5, not Fable 5.1?

Because Fable 5.1 costs twice as much per token and is slower, and for most work the extra capability does not show up in the result. Anthropic's model-selection guide puts it plainly in its selection matrix: "Most workloads start with Claude Opus 5." The same guide describes Fable 5.1 as Anthropic's "most capable widely released model", and then tells you when to reach for it: if your evals on Opus 5 at xhigh or max effort still fall short on demanding reasoning or long-horizon agentic work.

Read that order carefully. It is not "try Opus, then try Fable". It is "try Opus, then turn Opus's effort up, then try Fable". The guide says it directly: "Tuning effort is often a better lever than switching models."

That is the single most useful sentence in Anthropic's selection docs, and it gets lost in most comparisons. Effort is a request parameter on Fable 5.1, Opus 5 and Sonnet 5, and it trades thinking depth for speed and cost inside one model. A task that fails on Opus 5 at the default high might pass at xhigh for far less than doubling your token price.

The matrix Anthropic publishes for Fable 5.1 lists agent sessions that run for hours, multistep deep research, and analysis carried through to a finished document, spreadsheet or deck. If your task is none of those, Fable is probably the wrong default.

Which Claude model should you use for which task?

Use Haiku 4.5 for fast, repetitive, well-defined work; Sonnet 5 for everyday coding and writing; Opus 5 for hard multi-step and agentic work; and Fable 5.1 only when Opus at higher effort still fails. The table below turns that into specific jobs. The model picks follow Anthropic's selection matrix; the task rows are our own mapping of common jobs onto it.

One pattern runs through that table: nobody should start at Fable. It is the escalation path for the hardest work, not a default for people who want "the best one". If you are choosing in the Claude app rather than the API, the same logic applies. Open a new chat on Sonnet for drafting, switch to Opus when the thread gets hard.

How much does each Claude model cost for a real job?

At list price, a job with 10,000 input tokens and 2,000 output tokens costs about $0.20 on Fable 5.1, $0.10 on Opus 5, $0.04 on Sonnet 5 and $0.02 on Haiku 4.5. That arithmetic uses the per-million-token prices on Anthropic's pricing page, checked 17 September 2026:

• Fable 5.1: 10,000 × $10/M = $0.10 input, plus 2,000 × $50/M = $0.10 output. Total $0.20.
• Opus 5: $0.05 input plus $0.05 output. Total $0.10.
• Sonnet 5: $0.02 input plus $0.02 output. Total $0.04.
• Haiku 4.5: $0.01 input plus $0.01 output. Total $0.02.

Run that job 50,000 times a month and the gap is $10,000 on Fable against $1,000 on Haiku. That is why model choice is a budget decision long before it is a quality decision.

Three details change the math in practice, and all three come from Anthropic's own pricing notes:

1. Batch requests are 50% off. If the work can wait, the Message Batches API halves every row above.
2. Cache reads are cheap. A prompt cache hit costs 10% of the base input price on most models, and 2.5% on Fable 5.1 and Mythos 5.1. A long, repeated system prompt is where this pays. Anthropic's cost guide calls prompt caching the largest lever it measured.
3. Token counts are not equal across models. Anthropic's pricing page says Claude 4.7 and later models use a newer tokenizer that produces approximately 30% more tokens for the same text, and that Sonnet 4.6 and earlier use the previous one. Haiku 4.5 predates the change, so the same paragraph is fewer billable tokens on Haiku than on Sonnet 5. The price gap is a little wider than the per-token rates suggest.

Output tokens dominate on anything verbose. Every model's output rate is five times its input rate, so a model that writes 3,000 words when you needed 300 costs you more than a pricier model that stops on time. Telling the model the length you want is a cost control, not a style preference.

Which Claude models can you use in the Claude app on each plan?

On the Free plan you get Sonnet and Haiku, with no Opus and no Fable. That is the claude.com pricing table as of 17 September 2026. The paid tiers break down like this:

The table names the model families, not the version numbers, so check the picker inside the app for the exact model you are getting. The context window row on that page reads "Up to 1M" with a note that it varies by model, which matches the API table: Haiku's is smaller.

The practical takeaway for a Pro subscriber is that Opus is your heavy model and Fable is a metered extra. For a Free user, Sonnet is the ceiling, and it is a good one for most writing and analysis.

Do you need to change your prompt when you switch Claude models?

A little, and mostly in the request settings rather than the words. The biggest trap is temperature. Anthropic's Messages API reference marks the parameter deprecated and says models released after Claude Opus 4.6 do not support setting it: a value of 1.0 is accepted for backwards compatibility, and any other value is rejected with a 400 error. Haiku 4.5 predates that cutoff. So a pipeline tuned on Haiku with temperature: 0.2 will fail outright when you point it at Sonnet 5 or Opus 5.

The second difference is thinking. Haiku 4.5 uses extended thinking, the older manual mode with a token budget. Fable 5.1, Opus 5 and Sonnet 5 use adaptive thinking, where the model decides how much to think and you steer it with effort. On Fable 5.1 adaptive thinking is always on. Anthropic's overview notes that the manual extended-thinking mode is not accepted on models after the 4.6 generation, so a Haiku request body carrying a thinking budget needs editing before it moves up.

The third is the wording itself, and here the change is smaller than people expect. The same fundamentals work across all four: clear task, explicit output format, examples where format matters, and structure with system prompts and XML tags. Our guide to XML tags in Claude prompts applies to every model on this page. What does change is how much you spell out. Haiku benefits from tighter, more explicit instructions and a worked example. The adaptive-thinking models do better when you describe the goal and constraints and let them plan, which is the same shift we describe in reasoning models vs chat models.

Here is a request body that stays portable across all four models, because it avoids the settings that differ:

{
  "model": "claude-sonnet-5",
  "max_tokens": 2000,
  "system": "You are a support triage assistant. Classify each ticket into exactly one of: billing, bug, how-to, account, other. Reply with JSON only: {\"category\": \"...\", \"confidence\": \"low|medium|high\", \"reason\": \"one sentence\"}.",
  "messages": [
    { "role": "user", "content": "<ticket>I was charged twice this month and the second charge has no invoice.</ticket>" }
  ]
}

No temperature, no thinking block, no effort setting. Swap the model string to claude-haiku-4-5 or claude-opus-5 and the call still works. Add effort or thinking settings per model once you know which one you are keeping.

How do you test which Claude model is good enough for your task?

Build a small eval set from real inputs, run it on the cheapest plausible model first, and only move up when that model misses. Anthropic's selection guide describes two starting points: efficiency-first, which begins on Haiku 4.5 and upgrades only for specific capability gaps, and capability-first, which begins on Opus 5 and optimizes down over time. Both end in the same place: your own test set decides, not a leaderboard.

A 20-minute version that works for most teams:

CLAUDE MODEL EVAL, 20 MINUTES

1. Collect 15 real inputs for the task. Include 3 you know are hard.
2. Write down what a correct answer looks like for each, before running anything.
3. Run all 15 on the cheapest model you think could work (Haiku 4.5 or Sonnet 5).
4. Score each: correct / fixable with a prompt edit / wrong.
5. For every "fixable", edit the prompt once and re-run. Do not change models yet.
6. If 2 or more are still wrong, run the same 15 on the next model up.
7. On Opus 5, try effort xhigh before trying Fable 5.1.
8. Pick the cheapest model that got all 3 hard cases right.
9. Record the model ID, effort level and prompt version together.

Step 5 matters most. A surprising share of "Sonnet can't do this" turns out to be "the prompt never said what done looks like". Fixing the prompt is free; moving up a tier doubles or more your cost on every future call.

Can you combine Claude models instead of picking one?

Yes, and for high-volume systems it is often the cheapest way to get top-tier quality. Anthropic's cost guide describes two shapes that paid off in its measurements: an executor that escalates hard decisions to an advisor, and an orchestrator that delegates bulk work to lower-cost workers. In both, most tokens are billed at the cheaper model's rate.

In practice that looks like Haiku 4.5 or Sonnet 5 doing the bulk of the reading, extracting and drafting, with Opus 5 called only for the planning step or the final judgment. A research agent might use Opus 5 to break a question into sub-questions, hand each one to a Sonnet 5 worker, and bring the results back to Opus 5 to write the answer. The expensive model sees a few thousand tokens; the cheap ones see the rest.

The same guide is candid that these multi-model setups are narrower wins than prompt caching. Start with caching and batch pricing, then add a second model only if the numbers still hurt.

Where does Prompt Architects fit in a Claude workflow?

Prompt Architects is not a model and does not compete with any of the four above. It is a browser extension and web app that helps you write and store the prompt you send to Claude, ChatGPT or Gemini, so it sits in front of whichever model you pick.

That is useful in two places this comparison touches. First, when you step down a tier to save money, the prompt usually needs to become more explicit, and our generator and enhancer are built for exactly that rewrite. Second, when you keep one tested prompt per model, a saved library stops you from pasting a Haiku-tuned prompt into an Opus chat by accident. Saving prompts starts on Pro (50 saved prompts), with unlimited storage on Advanced; the Free plan gives you 5 prompt enhancements per day, per our FAQ, but cannot save prompts.

If you work in Claude Desktop or Claude Code, our MCP server exposes the same improve and refine tools inside the chat, so you can tighten a prompt without leaving the model you chose. For a broader look at how Claude differs from ChatGPT on prompting style, see ChatGPT vs Claude prompts. For how prompting changes between hosted and self-run models, see open-weight vs closed models.

The short version

Start on Opus 5 if you are unsure, because that is what Anthropic recommends. Drop to Sonnet 5 for everyday writing, analysis and code, and to Haiku 4.5 for high-volume simple work, keeping an eye on its October 2026 retirement floor. Turn up effort before you reach for Fable 5.1, and treat Fable as the escalation path for multi-hour agents and deep research, not a default. Every figure here was checked on 17 September 2026 and will move with the next release, so check Anthropic's models overview before you commit a budget to it.