Which Claude Model Is Best for Business? Sonnet vs Opus vs Haiku (2026)

Every week a client asks me some version of: "Which Claude model should I use?" The product page lists three, the chat dropdown has more, and the wrong default can quietly 5x your bill or starve a workflow that needed more horsepower. Here's the short, honest answer from someone who installs Claude into small businesses for a living.

What is the best Claude model right now?

It depends on the job, and the lineup changed three times since August. As of October 18, 2026:

• Best default for business: Sonnet 5.5. $2 in / $10 out per million tokens. Fast enough for real-time drafting, and the right model for about 90% of the work in this guide.
• Best for hard reasoning: Opus 5.5. $4 / $20. Strategy memos, contract review, multi-file code changes. Anthropic's own model guide says to start here for most demanding workloads.
• Smartest: Fable 5.1. $10 / $50, and the slowest of the four. Worth it only when Opus 5.5 still falls short. In the app it is not on Free, costs extra usage credits on Pro, and is capped at half your weekly limits on Max and on Team premium seats. Full read: Claude Fable 5.1.
• Cheapest and fastest: Haiku 5.5. $0.10 / $0.50 for prompts up to 100,000 tokens. Tagging, routing, summaries, anything repeated. In my own test it passed 154 of 156 checks to Sonnet's 156, at about a twelfth of the cost: Haiku vs Sonnet.

All four have a 1M-token context window. If you are weighing Claude against ChatGPT's newest model instead, that comparison is in GPT-6 Astra.

Claude model comparison at a glance

Model	Speed	Cost (per 1M tokens)	Context window	Best for
Claude Sonnet 5.5	Fast (real-time)	$2 in / $10 out	1M	Daily business work — drafts, analysis, client emails, light coding, the 90% default
Claude Opus 5.5	Moderate (thinks longer)	$4 in / $20 out	1M	Hard reasoning, deep research, complex code refactors, long agent runs, one-shot heavy lifts
Claude Haiku 5.5	Fastest	$0.10 in / $0.50 out (up to 100K-token prompts; $0.50 / $2.50 above)	1M	High-volume, simple tasks — classification, summaries, batch drafts, real-time chat
Claude Fable 5.1	Slower	$10 in / $50 out	1M	The escalation, when Opus 5.5 at higher effort still falls short
Sonnet 5 / Haiku 4.5 / Sonnet 4.6 / Opus 4.7 / Opus 5 (older, still available)	—	$2 / $10 · $1 / $5 · $3 / $15 · $5 / $25 · $5 / $25	200K–1M	Only if a workflow is pinned to them; the newer models are cheaper or better

Prices re-read from Anthropic's pricing page on October 18, 2026. Cache hits cost 5% of the input price on Opus 5.5 and Sonnet 5.5, 2.5% on Fable 5.1, and 10% on Haiku 5.5.

Correction (September 10, 2026), re-read from Anthropic's pricing page today: the Opus rate above is $5 / $25 per million tokens, not the $15 / $75 an earlier version of this table carried (that was the retired Opus 4.1 rate); Haiku is now Haiku 4.5 at $1 / $5 (Haiku 3.5's $0.80 / $4 is retired); and the Claude 5 generation prices at Sonnet 5 $2 / $10, Opus 5 $5 / $25, Haiku 4.5 $1 / $5, with the new top tier, Fable 5.1, at $10 / $50. Cache hits are one tenth of the input price on every model. The ratios in the sections below were corrected the same day: Opus is about 2.5x Sonnet per token, not 5x, and Haiku is a fifth of Opus, not a twentieth.

Correction (September 27, 2026), re-read from Anthropic's pricing page, model overview and plan table today: the table above now shows the current lineup. Anthropic released Claude Opus 5.5 on September 22, 2026 at $4 / $20 per million tokens, 20% under Opus 5, with cache hits at $0.20, and its model guide now says to start with Opus 5.5 for most workloads. That makes the current Opus 2x Sonnet 5 per token, not 2.5x; the 2.5x figures below are the Opus 5 and Opus 4.7 ratio and still hold for anyone pinned to those models. The recommendation doesn't change: Sonnet by default, Opus for the hard jobs, Haiku on volume. Only the Opus you reach for does. The full launch read is on Claude Opus 5.5.

Three models, three different jobs. Most teams quietly use whichever one is on top of the dropdown that week. That's the bug.

August 2026 update: the Claude 5 family is here — same logic, new names

Anthropic's Claude 5 generation began rolling out in 2026, and it adds something the 4.x lineup never had: a tier above Opus. The new flagship is Claude Fable 5 — Anthropic calls the tier "Mythos-class" — the most capable generally available Claude model, with Opus 5 and Sonnet 5 as the generational successors to the models in the table above, and Haiku 4.5 as the current small model. (There is also a Claude Mythos 5 — the same underlying model as Fable 5, offered to approved organizations without the additional dual-use safety measures — so for a business buyer, the top of the menu is Fable 5.)

Here's the part that matters for your bill and your defaults: nothing about the decision logic changes — it just shifts up one generation. The mid-tier model (now Sonnet 5) stays the right default for daily business work. The heavyweight (Opus 5, or Fable 5 where your plan offers it) stays reserved for tasks where reasoning depth is the product — strategy, hard analysis, one-shot heavy lifts. The small model stays on volume work. If your workspace defaults are set the way this guide describes, a generation upgrade is a dropdown change, not a re-architecture — Projects and Skills carry forward untouched.

Two practical notes from the first months of the 5 family. First, check current per-token rates on Anthropic's pricing page before moving any high-volume automation up a generation — the tier logic transfers, but exact numbers move with each release, and the cost trap described below gets more expensive one tier up. Second, resist the new-flagship reflex: the arrival of a tier above Opus doesn't mean your email triage suddenly needs it. The 90% of business work that ran perfectly on the mid-tier yesterday still runs perfectly on the mid-tier today.

September 2026 update: Claude Opus 5.5 is the Opus to use now

On September 22, 2026 Anthropic shipped Claude Opus 5.5, the first model in a Claude 5.5 family. The numbers that matter for this guide: $4 per million input tokens and $20 output, 20% under Opus 5; cache hits at $0.20, 60% under Opus 5; and, in Anthropic's words, it "performs at the level of Claude Fable 5.1 on most work." Its model guide now reads "start with Claude Opus 5.5 for most workloads," with Fable 5.1 kept for demanding reasoning and long-horizon agent work that Opus 5.5 at higher effort still can't handle.

What that does to the decision tree: the Opus branch gets cheaper and stronger, and the Fable branch gets narrower. On Pro, Max, Team and Enterprise, Opus is already in your plan, so there's no new bill. On the API, the gap between Sonnet and Opus drops from 2.5x to 2x per token. That still isn't a reason to make Opus the default. It's a reason to stop reaching for Fable first. Sonnet 5.5 followed on September 28 at Sonnet 5's exact per-token price ($2 / $10), so the Sonnet row did not move on price.

One plan detail that changed with it: Anthropic's plan table now lists Fable 5.1 on Team, but only on premium seats and capped at 50% of weekly limits. Standard Team seats get Opus, Sonnet and Haiku, and that's the right set for 90% of the work in this guide.

October 2026 update: Claude Haiku 5.5 makes the cheap branch much cheaper

Anthropic released Claude Haiku 5.5 on October 7, 2026, completing the 5.5 family. On the API it costs $0.10 per million input tokens and $0.50 output for prompts up to 100,000 tokens, a tenth of Haiku 4.5's $1 / $5, and $0.50 / $2.50 above that line. Anthropic calls it its fastest model at standard speed, and it's selectable in the Claude app on every plan, Free included. The same day, Sonnet 5.5's cache hits were halved to $0.10.

For this guide, the decision tree doesn't change shape. It changes the size of the gap. Sonnet 5.5 now costs 20 times Haiku 5.5 per token on short prompts, so any narrow, repeated job you've been running on Sonnet "to be safe" (tagging, routing, extraction, summaries) deserves a test on Haiku 5.5. Anthropic itself still says Sonnet and Opus are the better choice for complex multi-step agent work and coding. The prices, the 100K line and two worked workloads are in Claude Haiku 5.5.

Sonnet 4.6: the right default for most business work

October 2026: the current version of this default is Sonnet 5.5, at $2 / $10 per million tokens against Sonnet 4.6's $3 / $15. Everything below still applies.

Sonnet 4.6 is the model you should set as your default in claude.ai, in Cowork, and in any Skill you build. Fast enough to feel real-time when you're drafting an email. Less than half the cost of Opus per token (Sonnet 5 is $2 / $10 per million against Opus 5's $5 / $25, as of September 2026). And the reasoning quality is genuinely close to Opus for most business tasks — proposals, newsletters, client deliverables, meeting summaries, lead scoring, basic code review. You will not feel the difference on the vast majority of work.

The places Sonnet 4.6 quietly shines: long-form writing with brand voice (especially when paired with a loaded Project), structured outputs like tables and JSON, multi-step reasoning that doesn't require novel research, and any workflow you run more than once a day. If you've been defaulting to Opus "just to be safe," you're paying 2.5x for a difference you can't measure on most tasks.

Opus 4.7: when the harder model is actually worth it

October 2026: the current Opus is Opus 5.5, at $4 / $20 per million tokens against Opus 4.7's $5 / $25. Everything below still applies.

Opus 4.7 is the model you switch to when reasoning quality matters more than speed or cost. It thinks longer, holds more nuance, handles ambiguity better. For the right job, it's worth every cent. For the wrong job, it's a tax on your subscription.

The tasks where Opus actually earns its price tag: deep research where you're asking the model to weigh competing arguments, complex code refactors across multiple files, strategy documents where you want it to push back and find holes in your logic, legal or contract analysis where missing a clause matters, and any "one-shot" lift where you'd rather pay 2.5x once than re-prompt five times.

I use Opus for maybe 5% of my workflow: pricing strategy for a new offer, a board memo, a scope document for a six-figure engagement. Everything else stays on Sonnet. One pattern worth stealing: draft in Sonnet, then ask Opus to critique. You get most of Opus's value at a fraction of the cost because the critique is shorter than the draft.

Haiku: when "good enough" is actually the right answer

October 2026: the current Haiku is Haiku 5.5, at $0.10 / $0.50 per million tokens, a tenth of Haiku 4.5. Everything below still applies.

Haiku is the model nobody talks about, and it's the one that quietly saves businesses the most money on volume workflows. Fastest model Anthropic ships, half the cost of Sonnet 5 per token (Haiku 4.5 is $1 / $5 per million), and handles a surprising range of tasks well.

Use Haiku when you're doing something simple at high volume: classifying inbound leads as hot/warm/cold, summarizing call transcripts, drafting first-pass replies to common support tickets, tagging content, extracting structured data from emails, generating product description variants. It's also the right model for real-time chat experiences — customer-facing bots, internal lookup tools, anything where latency matters more than nuance.

The trap: don't use Haiku for anything that requires real reasoning. It's great at "do this simple thing fast," not "think through this carefully." If your output quality drops, jump back to Sonnet.

How much "good enough" Haiku really is, I measured: on 78 routing, invoice and pricing-math items run twice, Haiku 5.5 got 154 of 156 right to Sonnet 5.5's 156, and the bill came out 12 times lower rather than the 20 the price list suggests. The test, the one mistake Haiku kept making, and the monthly math are in Haiku vs Sonnet.

Cost comparison on a typical business workflow

Real numbers on a workflow most businesses have — triaging 500 inbound emails per week and drafting first-pass replies:

• Haiku 4.5: roughly $2–$5/week. Near-instant. Quality is fine for templated replies and basic triage.
• Sonnet 4.6 or Sonnet 5: roughly $6–$15/week. Noticeably better drafts that need less editing. Worth it if you're sending replies directly to clients.
• Opus 4.7 or Opus 5: roughly $15–$38/week at the $5 / $25 rate (corrected September 2026 from an earlier $40–$75 estimate built on the retired $15 / $75 rate). Marginally better drafts that almost nobody can tell apart from Sonnet's on this task. Not worth it for inbox triage.

That's a 5x cost difference between Haiku and Opus on the same workflow, for an output difference your customers cannot perceive.

How to actually switch models (in claude.ai, Cowork, and Skills)

The mechanics are straightforward once you know where to look.

In claude.ai: there's a model dropdown above the message input. The default is whatever was selected last, which is why most people drift into whichever model was used most recently rather than the one that fits the task.

Inside Claude Cowork: each Skill can specify its own model. When you build a Skill, there's a model selector in the Skill settings. Set Haiku for high-volume classification skills, Sonnet for daily drafting skills, Opus for the rare reasoning-heavy ones. This is where most of the cost optimization lives.

Via the API: the model is set per-request via the `model` parameter (as of September 2026: `claude-sonnet-5`, `claude-opus-5-5`, `claude-haiku-4-5`, and `claude-fable-5-1` for the top tier). Set this once at the workflow level rather than guessing per-call. Inside a Project, the model selection carries through the conversation — you can switch mid-thread to, say, draft with Sonnet and then hit Opus for a critique pass on the same context.

When NOT to use Opus (the cost trap that catches almost everyone)

The single most common mistake I see when I audit a client's Claude setup: Opus is the default model on every workflow, and nobody remembers turning it on. This is the easiest way to 2.5x your Claude bill without realizing.

Don't use Opus 4.7 for: short questions you could answer in your head, formatting and rewriting tasks, summaries under a page, anything you're running on a schedule (cron jobs, automated reports, batch processing), real-time chat with customers, or "warm-up" prompts where you're still figuring out what you want.

The pattern that costs the most money: someone sets up a daily automation, leaves it on Opus by default, and it runs 30 times a month for six months before anyone audits the bill. The fix is always the same — downgrade to Sonnet, save 60%, see no quality drop.

The practical decision guide

Five branches that cover most business decisions:

• If you're drafting client-facing copy (emails, proposals, newsletters): Sonnet 4.6.
• If you're doing daily analysis, summaries, or "rewrite this in a friendlier tone": Sonnet 4.6.
• If you're running a high-volume automation (classification, batch drafts, support triage): Haiku.
• If you're making a real decision (strategy, pricing, board memo, contract review): Opus 4.7 for the heavy reasoning pass, Sonnet for the rest.
• If you have no idea and just want to get unstuck: Sonnet 4.6. It is the right default 90% of the time.

This is the exact decision tree I install with clients when I set up Claude for their business — default Sonnet at the workspace level, Haiku on the high-volume Skills, Opus on the 2–3 Skills where reasoning depth actually moves the needle.

Set the right default and stop thinking about it

The honest takeaway: 90% of small businesses should set Sonnet 4.6 as the default everywhere — in claude.ai, in Cowork, in every Skill — and only switch up to Opus or down to Haiku when a specific task asks for it. Most people overpay for Opus they don't need, or underuse Haiku on volume work that would save them real money.

This is the same decision tree I configure when I install Claude for Small Business for clients. Default Sonnet, Haiku on the high-volume Skills, Opus on the 2–3 workflows where reasoning depth actually moves the needle. The setup takes an afternoon. The savings show up in the next billing cycle.

If you want the broader cost picture, I wrote a full Claude AI pricing breakdown — and the verdict on whether the $20 subscription itself is justified lives in Is Claude Pro worth it. Still deciding between Claude and another tool? The Claude vs ChatGPT for small business comparison is the head-to-head. Or skip the trial and error — my done-with-you Claude install ships with the model defaults already configured per Skill, so your business gets the right model on the right task from day one.

Claude Opus vs Sonnet: the math the pricing page does not do for you

Anthropic's pricing page gives you two numbers per model and stops. As of September 10, 2026 those numbers are Sonnet 5 at $2 per million input tokens and $10 per million output, Opus 5 at $5 / $25 (Opus 4.7 is priced the same), Haiku 4.5 at $1 / $5, and the new top tier, Fable 5.1, at $10 / $50. What the page does not do is the arithmetic that decides Opus vs Sonnet for a business. Three pieces of it. First, the ratio is 2.5x, not the 5x of the Opus 4.1 era, so the "just use Opus" tax got smaller this year; on a workflow that costs $10 a week on Sonnet, Opus is $25, not $50. Second, on a subscription you do not pay per token at all, but the same choice still costs you: the pricing page says every plan meters usage on a rolling five-hour window, that "how much you can do depends on the length and complexity of your conversations, the model you choose, and the features you use," and that there is no fixed message count. Opus on Pro drains the window faster than Sonnet does, so the model dropdown is a limits decision even when the bill is flat. Third, cache hits are one tenth of the input price on every model, so a Project with a long, stable system prompt makes the Opus premium smaller in practice than the list rate suggests. Fable is not on the Free plan and is metered separately on Pro and Max per the plan table, so it is not the default anywhere. September 27 update: the newest Opus, Opus 5.5, is $4 / $20, so the current ratio is 2x, not 2.5x, and a workflow that costs $10 a week on Sonnet 5 is $20 on Opus 5.5; the plan table also now lists Fable on Team premium seats at 50% of weekly limits. The decision rule survives the new numbers: Sonnet as the default, Opus on the two or three Skills where reasoning depth pays for itself, Haiku on volume.

Model choice is half the decision; the plan decides which models each seat gets. If you're buying for two or more people, my Claude Team plan breakdown covers seat prices, premium seats, and what admins control.