Which Claude Model Is Best? Opus 5.5 vs Fable 5.1 vs Sonnet 5 vs Haiku 4.5

Updated September 25, 2026 · 27 min read · By Lasse Rouhiainen

Are you using the same Claude model for everything? And did you know that since September 22, 2026, the most expensive Claude model is no longer the automatic choice for serious work?

If you only read one paragraph: start with Sonnet 5 for clear, routine work, Opus 5.5 for difficult work where several documents or constraints must agree, Haiku 4.5 for narrow tasks you repeat many times, and try Fable 5.1 only when a valuable problem is still unsolved after a good Opus attempt. Below you will find how to choose in 20 seconds, which models your plan includes, a real test with all four models on the same task, and 21 prompts you can copy.

Which Claude model should you choose? Start here

These are my recommended starting points, not a ranking from a controlled test. The models are not a best-to-worst ladder: Anthropic’s own guidance on choosing a model weighs capability, speed, cost and effort, and recommends testing on your own tasks.

Which Claude model should I try? A 20-second selector

Claude app or API? Two different questions

Most comparisons mix these up. If you pay a monthly subscription, a cheaper model does not lower your fixed subscription charge (usage billed separately, as on Enterprise, is a different story): your questions are about access, usage limits and how much correcting you do. If you pay per token through the API, price per result matters directly.

Which Claude models are included in your plan?

From Anthropic’s pricing page, checked on September 25, 2026:

Prices in US dollars, excluding applicable taxes. Plans and limits change often: check the pricing page before you buy.

How to change the model in the Claude app

Under the box where you type your message, you will see the current model and its effort level, for example “Opus 5.5 · Medium”. Click it to open the model menu and pick another model. Which models appear depends on your plan (interface checked on September 25, 2026).

What affects your usage limits

Anthropic’s pricing page says: “How much you can do depends on the length and complexity of your conversations, the model you choose, and the features you use.” It also notes that your activity on web, desktop, mobile and Claude Code “all draws from the same pool.” So on a subscription, choosing Sonnet for routine work mainly protects your limits for the work that needs Opus.

Good to know: when you use Opus 5 or Opus 5.5, Claude can switch a flagged message to another model for safety reasons, and shows a notice when it does. You can turn this off in Settings > Capabilities. Details in the Claude Help Center: why Claude switched models in your conversation.

I gave the same task to all four models

Recommendations are easy to write. So I tested one. On September 25, 2026, I gave Haiku 4.5, Sonnet 5, Opus 5.5 and Fable 5.1 exactly the same supplier comparison, and checked each answer against a list I wrote before reading any of them.

The method

• Task: compare three supplier offers for 40 laptops, installed and with data migrated, for a fictional dental clinic. The offers were written by me and are fictional.
• The traps in the material: offer A excludes installation in a note at the end; offer B quotes a price with VAT included while the others exclude it; offer B does not state its warranty length; offer C delivers too late and forces a paid support contract.
• Environment: Claude Code on September 25, 2026, using the model aliases haiku, sonnet, opus and fable, which Claude Code’s documentation on that date maps to the latest model of each family (Haiku 4.5, Sonnet 5, Opus 5.5, Fable 5.1). Effort was not set by hand, so each model ran at its own default. No tools, so no calculator. The same prompt for all four, one attempt each.
• Instructions loaded: my own working project instructions were active. They set Spanish as the default language, which is why the Sonnet answer came back in Spanish. I translated its excerpt below; the full answer is shown in the original language.
• Timing: the time each run took from start to finish, as reported by Claude Code. It is one observation, not a speed benchmark.
• What this is: one run of one task. It illustrates the differences; it is not a benchmark.

The results

What each answer said

• Haiku 4.5 got the VAT division wrong (“EUR 47,190 ÷ 1.21 = EUR 39,008.26”) and then wrote that “TechNova is lowest by EUR 792 (ex-VAT) vs. Brightline”, when TechNova is actually EUR 800 more expensive. The brief did not say which country the clinic is in, yet it assumed a jurisdiction and mentioned HIPAA, a US law, alongside GDPR. It also left a blank placeholder in one question. Its question about who is liable for patient data during migration was a good one.
• Sonnet 5 passed every check in the shortest time and recommended not awarding the contract yet (excerpt translated): “Do not award yet. Before committing…” followed by clear questions for each supplier.
• Opus 5.5 went further than the checklist. It spotted that “EUR 45 per device” could mean installation and migration each, and turned the comparison into a decision rule: “B is only the cheapest option if that uplift costs less than 800 excl. VAT.”
• Fable 5.1 also went beyond the checklist, asking whether the forced support contract is useful to the clinic: “Does Northwind need ongoing support anyway? This changes whether C’s EUR 3,600 is a cost or a benefit.” The EUR 3,600 is still a cost either way; the question is whether it buys something the clinic needs. It concluded: “The decision hinges on B’s answers, not on price.”

Check the test yourself

Download the material and the four full answers, exactly as returned (long dashes replaced by colons to match this site’s style):

• The requirement and the three fictional offers
• The checklist, written before the test
• Haiku 4.5, full answer
• Sonnet 5, full answer (in Spanish)
• Opus 5.5, full answer
• Fable 5.1, full answer

What this test shows, and what it does not

In this single test, run without calculation tools, Haiku made arithmetic and ranking errors that looked confident and would have changed the decision. Whatever model you use, check the numbers with a calculator or a spreadsheet before you act on its recommendation. The test also suggests that on a well-prepared, medium-sized task, Sonnet 5 was enough. Opus and Fable added useful judgment on top. On a messier real case, with long documents and more contradictions, that extra judgment is where I would expect the difference to matter most, which is why I still recommend Opus as the starting point for difficult comparisons. One run does not prove a winner: repeat the test with your own documents.

The four Claude models, one by one

Facts in this section come from Anthropic’s models overview (checked September 25, 2026). Claims about performance are Anthropic’s and are marked as such. The “starting point” lines are my recommendations.

Claude Opus 5.5

• Official: “for long-running agentic coding and knowledge work”. Released September 22, 2026. 1M-token context, 128K max output. API price $4 input / $20 output per million tokens.
• Anthropic says: it “performs at the level of Claude Fable 5.1 on most work” and costs about 40% less to run than Opus 5 on typical workloads (announcement). Anthropic’s docs recommend starting with Opus 5.5 for most workloads.
• My starting point for: proposals built from inconsistent inputs, supplier comparisons, spreadsheet reviews, decision memos, small internal tools and coded visuals.
• Not needed for: extracting or reformatting information you already have.

Claude Sonnet 5

• Official: “the best combination of speed and intelligence”. 1M-token context, 128K max output. API price $2 / $10.
• My starting point for: client follow-ups, meeting minutes, proposals from an approved package, content from supplied material, handouts, translation with a glossary, small changes to working code.
• Escalate to Opus when: it misses contradictions between documents, changes assumptions mid-calculation, forgets a key constraint twice, or a fix creates a new problem.
• A pattern that works: let Opus design a recurring report and its checks once; let Sonnet fill it every week. Keep it only if Sonnet’s output passes the same checks.

Claude Haiku 4.5

• Official: “the fastest model with near-frontier intelligence”. 200K-token context, 64K max output. API price $1 / $5.
• My starting point for: support triage, CRM field extraction, review tagging, document routing, form validation and answers drawn strictly from an approved answer bank.
• Watch out: keep outputs short and ask it to return “unknown” when evidence is missing. If a cheap filter silently drops the important document, the stronger model downstream never sees it. In the test above, it made arithmetic and ranking errors on a comparison, so check its numbers independently.

Claude Fable 5.1

• Official: “for demanding reasoning and long-horizon agentic work”. 1M-token context, 128K max output. API price $10 / $50, which is 2.5 times Opus 5.5 at equal token counts.
• Anthropic’s guidance: move to Fable 5.1 when results with Opus 5.5 at higher effort still fall short on demanding reasoning or long-horizon agentic work.
• My starting point for: a valuable, ambiguous decision; an unresolved analytical or technical problem; conflicting requirements; a high-stakes final review.
• How to compare fairly: give Fable the original evidence and acceptance criteria, without showing it the Opus conclusion first. Compare findings, not writing style. Two models agreeing does not prove they are right.

Anthropic has announced that Sonnet 5.5 and Haiku 5.5 will follow “in the coming weeks”. I will update this guide when they are released.

Which Claude model is best for coding?

For difficult coding work, start with Opus 5.5: Anthropic describes it as built for long-running agentic coding, and it is the default model in Claude Code for most plans. Use Sonnet 5 for everyday, well-scoped changes, Haiku 4.5 for simple fast sub-tasks, and Fable 5.1 for the hardest long-running problems when Opus falls short.

In Claude Code, type /model to open the picker, or add an alias to switch directly. In the picker, Enter switches and saves the model as your default, and s switches for the current session only. To start a session with a model, use the flag shown below. According to the Claude Code documentation (September 25, 2026):

/model opus
claude --model sonnet

Tip: try changing the effort before changing the model. Anthropic’s docs say “tuning effort is often a better lever than switching models”. In Claude Code use /effort high or start with claude --effort high. On Opus 5.5 the API default is medium.

Which Claude model is best for writing?

My starting point: Sonnet 5 for everyday business writing when the facts and a tone example are in front of it. Opus 5.5 when the text must stay consistent across many sources: a proposal built from contradictory emails, a board memo, a tender response. Haiku 4.5 only for short, rule-bound replies from an approved answer bank.

In my experience, the biggest jump in writing quality comes from giving Claude a real example of your tone and the list of facts it may use. That matters more than the model.

21 copy-ready prompts, organized by what you want to do

Open a card to see the prompt and copy it. Each card tells you what to attach before you send it. Where missing information would change the result, the prompt asks you for it first; where it would not, it just gets on with the job.

Write and communicate

Using the meeting notes I attach, draft a follow-up email to the client.
Include only agreed decisions, next actions, named owners and stated dates.
Do not turn suggestions into commitments. Mark any missing owner or date.
Match the tone of the example email I attach. Maximum 180 words,
then an action table.

Prepare a client proposal using the briefing and the approved service sheet I attach.
The service sheet controls scope, prices and exclusions.
Include objective, deliverables, delivery plan, client responsibilities,
price and exclusions. Flag any contradiction before you finalize.
Do not invent discounts, case studies, deadlines or guarantees.
If something essential for the proposal is missing, ask me before you start.

Turn the attached transcript into decisions, actions and open questions.
For every action, include the owner, the deadline if stated and the timestamp.
Separate real decisions from proposals. Write "not stated" for missing data.
Finish with the three unresolved matters most likely to delay progress.

Create three LinkedIn post drafts from the approved material I attach.
Audience: business owners who are not technical. Each post needs one useful
lesson, one concrete example and one open question. Maximum 180 words each.
Keep a source note for every factual claim. Do not invent results or testimonials.

Update the approved report I attach using the new figures in the second file.
Keep its structure and wording except where the new evidence requires a change.
Return only the changed sections, plus a table with old value, new value and source.
Flag any conclusion that no longer follows from the data.

Analyze and decide

Compare the supplier proposals below for our requirement. First normalize scope,
quantities, tax treatment, installation, support, lead times and exclusions.
Cite the source for each important term. Show which prices cannot yet be compared
fairly. Then recommend the next decision and the questions to ask before committing.
If information that would change the recommendation is missing, say so instead
of assuming.

Review the attached workbook against the attached definitions of our KPIs.
Check formula consistency, duplicate entries, missing periods, units and totals.
Do not change the original: work on a copy if you can.
Return an issue log with sheet and cell references, how much each issue matters
and the proposed fix. Never change assumptions silently or overwrite source data.

Prepare a decision memo from the evidence I attach.
Separate facts, assumptions, estimates and unresolved questions.
Compare the realistic options against the criteria I give you. Include a downside
scenario, the smallest reversible next step and the evidence that would change your view.
Keep the memo to two pages; put calculations and sources in an appendix.
If the decision or the criteria are unclear, ask me before you start.

Review the proposed decision and the supporting evidence I attach.
Identify the three assumptions whose failure would change the outcome most.
For each one, give the supporting and conflicting evidence and a cheap way to test it.
Do not manufacture objections. Separate decisive problems from minor concerns.
Finish with the new evidence that would change your assessment.

Independently analyze the unresolved problem I describe, using the original evidence.
Do not assume a solution exists. Identify missing evidence and contradictions.
Produce a testable proposed solution, or explain the exact condition that blocks one.
Keep every conclusion traceable to the evidence.
If the constraints or the acceptance tests are unclear, ask me before you start.

Process many records

Classify each message using only the labels in the list I attach.
Return record_id, label, urgency, supporting phrase and needs_review.
Set needs_review=true when the label is ambiguous or context is missing.
Do not draft replies or infer facts that are not in the message. Keep every record ID.

From each sales note, extract company, contact, stated need, next action, owner and date.
Use null for missing facts. Keep the original record_id and a supporting excerpt.
Do not infer buying intent from polite language.
Flag any record with contradictory dates or next actions.

Answer the question only from the approved answer bank I attach.
Include the matching answer ID. If no entry answers it, return NEEDS_REVIEW.
Do not combine entries into a new policy or promise an exception.
Keep the answer under 90 words.

Check each record against the list of required fields I attach.
Return record_id, missing_fields, contradictory_fields and review_reason.
Use only the supplied record. Do not fill in missing values or interpret
unclear attachments. Return the table only, with no commentary.

Build tools, agents and visuals

I want to turn a task I repeat every week into a pilot for an AI agent.
Define its trigger, inputs, approved sources, steps, output, owner and checks.
Say which actions the agent can draft and which need human approval.
Show one normal case, one case with missing data and one exception.
Do not assume any connector or scheduling feature is already enabled.
Ask me about the task before you start if my description is not enough.

Build a simple calculator for the decision I describe, using the formulas I give you.
List any missing input before you build it. Keep assumptions visible and editable.
Use the smallest implementation that works in this environment.
Test typical values, boundary values and invalid input.
Deliver the working file, the source and a short verification note.

Create a 15-second vertical 9:16 animation, built in code, that explains our service.
Seconds 0 to 4: show the problem. 4 to 11: show the real solution in three actions.
11 to 15: show the outcome and the call to action I give you.
Use a cream background, dark text and one accent color. At most seven words on screen at once.
Keep important content away from the bottom and right edges.
Use only claims I approve. Deliver editable source, a preview and an MP4 if rendering is available.
If you do not have my service description or call to action, ask me for them first.

Work better with any model

Complete the deliverable I describe, using only the sources and tools I name.
Give short progress updates at meaningful milestones and keep working.
Stop when the acceptance criteria pass, the budget I set is reached,
or a missing fact or permission blocks you.
After two failures at the same step, diagnose the cause before trying again.
Report only work actually completed; never describe planned actions as finished.

Prepare a handoff of no more than 500 words for the next work session.
Include the objective, approved decisions, authoritative files, current outputs,
remaining tasks, unresolved questions and acceptance checks.
Keep exact IDs and source pointers. Leave out speculative reasoning.
Flag any critical detail that cannot safely be shortened.

Evaluate the attached output against the acceptance checklist I attach.
For each failure, cite the exact location and explain its practical effect.
Separate confirmed errors from preferences and unverified concerns.
Recommend only changes that materially improve the result.
Return pass or fail, the issue table and the minimum necessary corrections.

I will describe a task I do at work. Recommend which Claude model to start with
(Haiku 4.5, Sonnet 5, Opus 5.5 or Fable 5.1), and why, based on how hard
the task is to complete and to check. Define what a good result looks like,
the check I should run before trusting it, and the signal that means I should
escalate to a stronger model. If my description is not enough to decide, ask me first.

For API and automation users: prices, cost and routing

This section matters if you pay per token. API prices from Anthropic’s models overview, in US dollars per million tokens, checked September 25, 2026:

Opus 5.5 also has a fast mode at $8 input and $40 output per million tokens, according to the announcement.

A worked example

A job with 20,000 input tokens and 3,000 output tokens, no caching, costs: Haiku 4.5 $0.035 · Sonnet 5 $0.070 · Opus 5.5 $0.140 · Fable 5.1 $0.350. Real jobs use different token counts on different models, and these figures exclude tools, retries and review time.

Caching repeated inputs

With Opus 5.5, writing a prompt prefix to the five-minute cache costs $5 per million tokens and reading it costs $0.20. Reusing the same 20,000-token prefix ten times: $0.80 without caching, $0.136 with one write and nine reads. That is 83% less for that prefix, not for the whole job. Caching is set up in the integration; writing “cache this” in a prompt does not activate it.

The cost metric that matters

Cost per accepted result =
(model cost + external tools + rendering + human review + rework)
/ number of results that pass the agreed checks

In the test above, Haiku was the cheapest model per token but the only answer that needed correcting before use. That is the whole point of this formula.

A routing policy you can adopt

1. If the task is deterministic, use a formula, a filter or a script.
2. Narrow output format with some language judgment: test Haiku.
3. Routine drafting or synthesis with clear inputs: test Sonnet.
4. Difficult constraints, several files or production steps: test Opus 5.5.
5. Valuable and still unresolved: compare Fable on the same evidence.
6. Keep the option with the lowest total cost for an acceptable result, including review and rework.

Twelve habits that cut cost and rework

1. Define “finished”: artifact, format, audience and acceptance criteria.
2. Say which source wins when files conflict.
3. Ask only for missing inputs that would change the result.
4. Request edits to the affected section, not a full rewrite.
5. Separate research from production.
6. Send a relevant evidence pack. A large context window is a limit, not a target.
7. Keep source pointers in every summary.
8. Escalate only the difficult items, not the whole batch.
9. After two failures at the same step, diagnose before retrying.
10. Approve ten outputs, including hard cases, before processing a thousand.
11. Turn every failure into a new check in the next brief.
12. Avoid switching models mid-task without a written handoff (prompt 19).

For developers moving to Opus 5.5

• Start at the default medium effort and measure before raising it. Anthropic’s docs say tuning effort is often a better lever than switching models.
• Thinking is always on and uses the output budget, so very small output limits can cut results short.
• Forced tool use (tool_choice of type any or tool) is not supported on Opus 5.5. Review older integrations that relied on it.
• Read the breaking changes before moving production traffic from Opus 5.

What I see when I train teams, and my verdict

I train teams at large companies in Spain and Europe. Three things I see again and again:

1. Context beats the model. In my trainings, every participant builds their own Claude Project with the context of their real work. Wherever we have done that, the difference in the answers has been striking, with the same model.
2. Real documents show what a model can really do. Fictional examples are useful to teach a method, as in the test above. But to judge whether Claude is useful for your business, use representative real documents: with a participant’s own material, the same model suddenly looks far more capable.
3. Check what was actually delivered, not what was planned. A mistake in one of my own email campaigns taught me this: since then, anything we publish is re-read exactly as it was saved, not as it was intended. The same habit catches delivery and editing mistakes; numbers and facts still need their own independent check.

My verdict: Sonnet 5 for routine work, Opus 5.5 for difficult deliverables, Haiku 4.5 for narrow repeatable processing, and Fable 5.1 only where it proves an advantage on your own tasks. Choose the model with the lowest total cost for an acceptable result, counting review and rework, not just the token price.

If you want to go further, read my guides on using Claude and ChatGPT together for business tasks, how memory works in ChatGPT and Claude and the AI strategy questions every CEO should ask.

Frequently asked questions

What are the different Claude models?

As of September 25, 2026, the current models are Fable 5.1 (for demanding reasoning and long-horizon agentic work), Opus 5.5 (for long-running agentic coding and knowledge work), Sonnet 5 (the best combination of speed and intelligence) and Haiku 4.5 (the fastest). Anthropic has announced Sonnet 5.5 and Haiku 5.5 for the coming weeks.

Which Claude model is best?

There is no single best model for every task. Anthropic recommends starting with Opus 5.5 for most workloads. For routine work with clear inputs, Sonnet 5 is often enough, as the test in this guide shows, and for narrow high-volume tasks Haiku 4.5 is the most economical.

Which Claude model is best for coding?

Start with Opus 5.5 for difficult coding; it is the default in Claude Code for most plans. Use Sonnet 5 for everyday, well-scoped changes, Haiku 4.5 for simple sub-tasks and Fable 5.1 for the hardest long-running problems.

Which Claude model is best for writing?

Sonnet 5 for everyday business writing. Opus 5.5 when the text must stay consistent across many sources, such as proposals, board memos or tender responses.

Is Opus 5.5 better than Fable 5.1?

Not on everything. Anthropic says Opus 5.5 performs at the level of Fable 5.1 on most work, while Fable remains its most capable widely released model. In the test in this guide, both passed every check, and each added a different useful insight. Fable costs 2.5 times as much per token.

Which Claude models are on the free plan?

Sonnet and Haiku, according to Anthropic’s pricing page on September 25, 2026. Opus is included from Pro. Fable comes with usage credits on Pro, at 50% of weekly limits on Max and Team, and on Enterprise.

How do I change the model in Claude Code?

Type /model to open the picker, or /model opus (or sonnet, haiku, fable) to switch directly. To start a session with a model: claude --model sonnet.

Sources, checked on September 25, 2026: Anthropic, Introducing Claude Opus 5.5; Models overview; Choosing the right model; Claude Code model configuration; Claude pricing; Claude Help Center. Test run by the author on September 25, 2026, as described above. Recommendations are starting points to test on your own tasks.

Lasse Rouhiainen is an international AI keynote speaker and bestselling author. He trains teams at large companies in Spain and Europe on AI agents and applied artificial intelligence. More about Lasse.