Claude Sonnet vs Opus vs Haiku: Costs and When to Use Each

Tagging 10,000 customer reviews by theme costs about $2.67 on Claude Haiku 4.5, $7.20 on Sonnet 5, $14.40 on Opus 5.5 and $36.00 on Fable 5.1 at Anthropic’s API list prices as of September 2026, with 150 words in and 10 out per review at about 1.8 tokens per word (1.33 on Haiku’s older tokenizer). One 2,000-word blog draft costs about 6 cents on Sonnet and 12 cents on Opus, so for a single draft the Claude Sonnet vs Opus price gap is six cents.

Those figures are my arithmetic from Anthropic’s per-token pricing, not measured runs, and they leave out thinking tokens. The cost table further down shows the working for every job.

So the choice turns on two things a price list does not show: how many times a job runs, and how many steps each run takes. Volume multiplies the 2x gap between the tiers. Steps can reverse it: in Anthropic’s internal runs of a coding benchmark (a SWE-bench Pro subset), Opus 5.5 came out cheaper per solved task than Sonnet 5.

My routing rule for marketing work follows from those two facts, and it departs from Anthropic’s Opus-first default for single-pass writing and for jobs repeated at volume:

The rule: Send checkable bulk jobs to Haiku 4.5. Run single-pass writing and jobs repeated at volume on Sonnet 5. Move to Opus 5.5 for multi-step jobs or work whose figures and sources must be right. Try Fable 5.1 only when Opus 5.5 at xhigh or max effort falls short.

If you are still choosing between vendors, Claude vs ChatGPT covers that decision. This page assumes you have picked Claude and now have to pick a model.

Claude Models Comparison: Sonnet vs Opus vs Haiku (September 2026)

Anthropic’s models overview lists four current models, checked on 26 September 2026. MTok means one million tokens.

Model	Input / output price per MTok	Context window	Best for
Claude Fable 5.1	$10 / $50	1M tokens	Hours-long agent runs and deep research, only after Opus falls short
Claude Opus 5.5	$4 / $20	1M tokens	Multi-step agentic work, research synthesis, anything with figures
Claude Sonnet 5	$2 / $10	1M tokens	Single-pass drafting and rewriting, repeated jobs at volume
Claude Haiku 4.5	$1 / $5	200K tokens	Bulk tagging, extraction and sorting you can spot-check

The “best for” column is my routing, built from Anthropic’s model selection matrix, which says “most workloads start with Claude Opus 5.5”. The Sonnet vs Opus section below explains where I follow that default and where I peel off from it.

In the Claude apps, the Free plan covers Sonnet and Haiku while Opus needs a paid plan and Fable access varies by plan, per Anthropic’s plans page; the free tier is also on my list of free AI tools for marketing. Everything below uses API prices, because per-token prices are what let you price a job.

The rest of the spec sheet, from the same overview page:

	Fable 5.1	Opus 5.5	Sonnet 5	Haiku 4.5
Latency (Anthropic’s label)	Slower	Moderate	Fast	Fastest
Reliable knowledge cutoff	Jun 2026	Jun 2026	Jan 2026	Feb 2025
Max output	128K tokens	128K tokens	128K tokens	64K tokens
Thinking	Adaptive, always on	Adaptive, always on	Adaptive	Extended (manual)
Default effort	high	medium	high	Not supported
Retirement, not sooner than	Sep 1, 2027	Sep 22, 2027	Jun 30, 2027	Oct 15, 2026

The knowledge cutoff row tells you how recent a topic can be before you have to paste in sources. The thinking row decides how far a request can run above its visible-text cost, since thinking tokens are billed as output (next section).

What One Marketing Job Costs on Each Claude Model

Stage 3 of the marketer’s AI maturity curve is where a team starts measuring what its AI output costs. This table does that on paper, per job, before you run anything.

The assumptions, so you can swap in your own:

• Words to tokens. Anthropic’s models overview (markdown version) says 1M tokens is roughly 555,000 words on the current tokenizer, introduced with Claude Opus 4.7, while models before it fit about 750,000 words. That works out to about 1.8 tokens per word and 1.33 tokens per word. The pricing page says Claude 4.7 and later models use the newer tokenizer and Claude Sonnet 4.6 and earlier use the previous one, so Fable 5.1, Opus 5.5 and Sonnet 5 take the 1.8 figure and Haiku 4.5 takes 1.33.
• Visible text only. Anthropic’s thinking documentation says reasoning tokens “are billed as output tokens, even when the thinking text isn’t returned to you”. None are counted here, on any model, so treat every figure as a floor. Sonnet 5 thinks by default unless you pass thinking: {type: "disabled"} (what’s new in Sonnet 5), while Opus 5.5 returns an error if you try (what’s new in Opus 5.5) and Fable 5.1 always thinks.
• List prices. No prompt caching, no Batch API discount, global routing. The pricing page adds a 1.1x multiplier for US-only inference on Claude 4.6 and later models.
• No long-context premium. The pricing page says Claude 4.6 and later models bill the full 1M-token window at standard rates, so the large synthesis job pays the same per token as the small draft.

The jobs:

Job	Words in	Words out
One blog draft	6,000 (brief, source notes, style guide)	2,000
Research synthesis	120,000 (20 interview transcripts of 6,000 words)	2,500
Rewrite 1,000 product descriptions	300 per item (150 of instructions, 150 of original copy)	120 per item
Tag 10,000 customer reviews	150 per item (100 of instructions, 50 of review)	10 per item

What each one costs, calculated from Anthropic’s per-token prices:

Job	Haiku 4.5	Sonnet 5	Opus 5.5	Fable 5.1
One 2,000-word blog draft	$0.02	$0.06	$0.12	$0.29
Research synthesis of 20 transcripts	$0.18	$0.48	$0.95	$2.39
Rewrite 1,000 product descriptions	$1.20	$3.24	$6.48	$16.20
Tag 10,000 customer reviews	$2.67	$7.20	$14.40	$36.00

The blog draft on Sonnet 5, worked through: 6,000 words x 1.8 = 10,800 input tokens, at $2 per million = $0.0216. Then 2,000 words x 1.8 = 3,600 output tokens, at $10 per million = $0.036. The total is $0.0576, which rounds to $0.06, and every other cell follows the same two lines with that model’s prices and tokens-per-word figure.

What the table shows:

• Opus 5.5 costs exactly twice Sonnet 5 on every row. The per-token prices are 2x apart and the two models share a tokenizer, so without caching the ratio never moves. Only the size of the job changes what 2x means in dollars.
• Haiku 4.5 lands at about 37% of Sonnet’s cost, lower than its 50% sticker ratio. Its older tokenizer turns the same words into fewer tokens.
• Fable 5.1 costs 2.5x Opus 5.5 on every row of this table.
• Bulk jobs can run at half price. The Batch API takes 50% off both input and output tokens in exchange for results within 24 hours, so the review-tagging job on Sonnet 5 drops to $3.60.

Cheaper Moves Than Switching Models

Anthropic’s cost and intelligence guide sorts cost levers into “free wins” that cut spend without touching quality and tradeoffs such as model choice. Caching pays only when a long prefix repeats across requests, such as an agent loop or a long style guide sent with every draft. The minimum cacheable prompt is 1,024 tokens on Sonnet 5, 512 on Opus 5.5 and 4,096 on Haiku 4.5 (prompt caching docs), so the 100-word tagging instructions in the table above are too short to cache.

1. Prompt caching, where the prefix qualifies. Cache reads cost 10% of the base input price (5% on Opus 5.5, 2.5% on Fable 5.1). In Anthropic’s measured agent runs, the guide says caching “cut agent-loop cost by a factor of 2.7 to 5.3” and is “worth more than most model-choice decisions”.
2. Batch processing for anything no one is waiting on, at 50% off.
3. Effort. This one is a tradeoff rather than a free win, but the guide’s advice when “costs are too high; quality is fine” is to sweep effort down on your current model before changing models.

Claude Sonnet vs Opus: Where the 2x Price Gap Matters

Per base input and output token, the Claude Sonnet vs Opus comparison is fixed: Opus 5.5 at $4 / $20 against Sonnet 5 at $2 / $10. Cache reads are the exception: Opus 5.5 reads its cache at 5% of its input price, so both models pay $0.20 per MTok for a cache hit.

Sonnet’s rate is also settled. The pricing page says the $2 / $10 rate, announced as introductory pricing through August 31, 2026, “is now the standard price”, and the scheduled rise to $3 / $15 “will not occur”.

Per finished job, the comparison moves. Anthropic’s cost guide tells readers to “compare on cost per completed task, not per token”, and publishes its own measurements:

Benchmark (Anthropic’s internal runs)	Model and effort	Score	Cost
SWE-bench Pro subset (coding)	Sonnet 5, default	77.4%	$0.84 per solved task
SWE-bench Pro subset (coding)	Opus 5.5, default (medium)	92.8%	$0.22 per solved task
SWE-bench Pro subset (coding)	Opus 5.5, low	87.4%	$0.12 per solved task
SWE-bench Pro subset (coding)	Fable 5.1, default	92.3%	$1.19 per solved task
SWE-bench Pro subset (coding)	Fable 5.1, low	88.6%	$0.54 per solved task
DeepResearch Bench II (research)	Sonnet 5	56%	$1.20 per task
DeepResearch Bench II (research)	Fable 5.1, low	66%	$4.66 per task

Anthropic labels these internal results “directional, not guarantees”. On the coding benchmark, Opus 5.5, at twice Sonnet 5’s token price, cost about a quarter as much per solved task ($0.22 against $0.84), while Fable 5.1 at its default cost more than both ($1.19). Anthropic plots all five coding rows on one chart, with a footnote on problem sets: Sonnet 5 and Fable 5.1 at low effort ran on a 482-problem subset, and Opus 5.5 and Fable 5.1 at its default ran on 478 of those problems. The guide’s explanation is that “a more capable model finishes a task with less work: fewer turns, less searching, less re-reading of its own context, and less backtracking.”

Cache pricing also shrinks the gap in loops like these. Anthropic priced each request’s earlier context in those runs as a cache read, which costs $0.20 per MTok on both Opus 5.5 and Sonnet 5, and the guide says cache reads are “routinely the largest single component of task cost” across its measured runs.

A single-pass draft is one request with no turns to cut, so the per-token table above is the starting estimate for it. Sonnet 5 is half the price for the visible text, and the thinking on top depends on the effort each model runs at (Opus 5.5 defaults to medium, Sonnet 5 to high), so set effort explicitly on both before you compare, or turn thinking off on Sonnet 5 for a short rewrite job.

A job that loops (an agent building a script, a research run that reads, searches and re-reads) is where Opus 5.5 can win back its price. On research the cost can climb instead: Fable 5.1 at low effort scored 10 points above Sonnet 5 on DeepResearch Bench II at about four times the cost per task, which the guide puts down to Fable running “a longer research loop over a larger context”.

Claude Opus vs Sonnet for Writing

Anthropic’s selection matrix lists “content creation” among Sonnet 5’s example uses, next to code generation and data analysis, and the spec table above gives Sonnet 5 a “Fast” latency label against Opus 5.5’s “Moderate”. Opus 5.5’s matrix examples are multi-hour coding agents, large refactors and computer use.

What Anthropic does claim for Opus 5.5 on written work sits in its prompting guide: the model “is much less likely to state an incorrect figure or cite the wrong source”, and “the spreadsheets, slides, and documents it produces need less editing before you share them.” The guide describes differences from Opus 5 and never mentions Sonnet, so read both claims as the direction Opus moved rather than a ranking of the two tiers.

One marketer-run test worth reading is Lisa Peyton’s content-marketer showdown, updated June 2026, which ran one newsletter prompt and the same sources through Opus 4.8, 4.7, 4.6 and Sonnet 4.6. She found that Sonnet 4.6 “wrote fluently and quickly, and for my particular voice rules it needed the most editing”, and her verdict was that “there isn’t a single winner right now”.

Every model she tested predates Sonnet 5 and Opus 5.5, and it is one prompt from one writer, so treat it as a named example of the method rather than a result for today’s lineup.

My writing rule from all of that:

• Draft blog posts, emails and landing page copy on Sonnet 5. The price is not the reason for a one-off draft, since the gap is six cents: if Opus 5.5 drafts need less editing against your own voice rules, draft there instead.
• Move a draft to Opus 5.5 when it carries figures, citations or a synthesis of many sources, since wrong figures and wrong sources are the failure Anthropic says Opus 5.5 improved on.
• Keep writing jobs that repeat at volume on Sonnet 5. A run of 1,000 product descriptions costs $3.24 there against $6.48 on Opus 5.5, before the batch discount.
• Run a voice pass yourself on either tier. No model knows your house rules until you write them down and hand them over.

One prompt habit to revisit on Opus 5.5 is the chain-of-thought pattern in my ChatGPT prompts for marketing templates. The prompting guide’s section on safeguard refusals says requests that push the model to reproduce its internal reasoning in the response text “can be declined”, and tells you to remove prompt instructions that ask it to write out its reasoning. Keep the steps that produce something you want to read, such as a topic list, and cut the ones that only ask it to show its work.

Claude Haiku vs Sonnet: When the Cheapest Model Is Enough

The Haiku vs Sonnet decision has more spec differences than the Sonnet vs Opus one, because Haiku 4.5 predates the other three models and their newer tokenizer:

	Haiku 4.5	Sonnet 5
Price per MTok (input / output)	$1 / $5	$2 / $10
Cost of the same text	About 37% of Sonnet, older tokenizer	Baseline
Context window	200K tokens	1M tokens
Max output	64K tokens	128K tokens
Reliable knowledge cutoff	Feb 2025	Jan 2026
Thinking	Extended thinking, set manually	Adaptive, on by default
Effort parameter	Not supported	Supported, default high
Anthropic’s example uses	Real-time apps, high-volume processing, sub-agent tasks	Code generation, data analysis, content creation

Anthropic’s cost guide draws the line on where Haiku stops. On the GPQA Diamond question set, Haiku 4.5 answered at about a fifth of Opus 5.5’s cost per question, with 63% accuracy against 92%, and “fell much further behind on long coding tasks”. The guide’s summary: “It fits high-volume work with checkable outputs, not long agentic loops.”

Marketing jobs where a person can spot-check a sample against a rule suit Haiku 4.5:

• Tagging reviews, support tickets or survey answers by theme and sentiment
• Pulling company name, role and use case out of form fills
• Normalising messy job titles in a lead list
• First-pass sorting of inbound requests before a person reads them

Keep Haiku away from anything that depends on events after February 2025 without sources in the prompt, from long agent loops, and from final copy. If a hand-checked sample of 50 Haiku outputs shows more errors than you will accept, rerun the same sample on Sonnet 5 before changing the whole job.

Haiku 4.5 is listed as active on Anthropic’s model deprecations page, with retirement “not sooner than October 15, 2026”, and Anthropic commits to at least 60 days’ notice before retiring a publicly released model. Pin the model ID in any recurring job and check that page before you build on it.

Claude Fable 5.1: The Tier Above Opus

Fable 5.1 is, in the words of Anthropic’s model selection page, its “most capable model open to all customers”. It costs $10 / $50 per MTok and carries Anthropic’s “Slower” latency label.

Anthropic’s own advice is to reach it through Opus: start on Opus 5.5, and if “evals at xhigh or max effort still fall short on demanding reasoning or long-horizon agentic work, move to Claude Fable 5.1.” Two of the selection matrix’s Fable examples describe marketing work: “multistep deep research, analysis carried through to a finished document, spreadsheet, or deck”. A full competitive analysis that has to end as a deck fits that description.

Against Opus 5.5, the measured numbers argue for restraint. On the SWE-bench Pro subset, Opus 5.5 at its default matched Fable 5.1 at its default “for about a fifth of the cost per solved task ($0.22 against $1.19)”.

Against Sonnet 5 the result flips: Fable 5.1 at low effort solved 88.6% for $0.54 per solved task, which the guide sums up as “11 more points for 35% less per solved task”. On research, the guide concludes that “Fable 5.1 earns its price only at low” effort.

Claude Mythos 5.1 shares Fable 5.1’s capabilities and prices but is open only to Project Glasswing participants, so for everyone else Fable 5.1 is the top of the ladder.

Which Claude Model Should You Use for Marketing Work?

Hand checkable bulk jobs to Haiku 4.5, single-pass writing and repeated jobs to Sonnet 5, and multi-step or figure-heavy jobs to Opus 5.5. Anthropic’s escalation order for when output is not good enough: “If you lowered effort, restore it; otherwise try the next tier up at low effort”.

Marketing job	Start on	Move up when	Why
Blog posts, emails, landing page drafts	Sonnet 5	The draft carries figures or cites sources, or Opus drafts need less of your editing: Opus 5.5	Content creation sits in Sonnet 5’s row of Anthropic’s matrix
Rewriting product descriptions or ad variants at volume	Sonnet 5, through the Batch API	Spot-checks keep failing: Opus 5.5 at low effort	One pass per item, and the batch discount cuts a 1,000-item Sonnet 5 run to $1.62
Tagging reviews, tickets and survey answers	Haiku 4.5	A hand-checked sample misses your error bar: Sonnet 5	”Checkable outputs” is Anthropic’s own description of where Haiku fits
Keyword clustering and content briefs, the first steps of an AI SEO strategy	Sonnet 5	Clusters span thousands of keywords and many sources: Opus 5.5	Mostly single-pass structuring
Research synthesis across transcripts, reports and competitor pages	Opus 5.5	Hours-long multi-source research still misses: Fable 5.1 at low effort	Figures and sources must be right
Competitive analysis carried through to a finished deck	Opus 5.5	Output needs heavy rework: Fable 5.1	Anthropic’s Fable examples name the finished deck
Building scripts, tools and pages in Claude Code	Opus 5.5 (Claude Code’s default)	Long autonomous tasks stall: Fable 5.1	$0.22 per solved coding task at default in Anthropic’s runs, against $0.84 for Sonnet 5
Codebase search and exploration subagents	Haiku 4.5	Explorations miss files you know exist: inherit the main model	Claude Code’s docs suggest Haiku to control subagent costs

Claude Code Model Settings: Default, /model and Subagents

I built this site, its nine free tools and its research scripts (a Search Console CLI, a Semrush keyword CLI and an IndexNow pusher) with Claude Code. Model choice there works differently from the API: the tool picks a default for you, and subagents can run on a different tier from the main session. Everything below comes from Anthropic’s Claude Code model configuration docs as of September 2026.

The default. The default setting resolves to Opus 5.5 on Pro, Max, Team, Enterprise and Anthropic API accounts, and on Amazon Bedrock, Google Cloud’s Agent Platform and Claude Platform on AWS. Microsoft Foundry defaults to Sonnet 4.5. Opus 5.5 needs Claude Code v2.1.280 or later, and before that version Pro and Team Standard accounts defaulted to Sonnet 5, so run claude update if your session header shows an older model.

The aliases:

Alias	What it selects
sonnet	Latest Sonnet (Sonnet 5 on the Anthropic API)
opus	Latest Opus (Opus 5.5 on the Anthropic API)
haiku	Haiku, for simple tasks
fable	Fable 5.1, never the default on any plan
best	Fable where available to you, otherwise Opus
opusplan	Opus in plan mode, then Sonnet for execution
default	Clears any override and returns to your account’s default

How to switch:

1. In a session, run /model sonnet (or any alias). This also saves it as your default for new sessions; open the /model picker and press s on a row to switch for this session only.
2. At launch, run claude --model haiku to set the model for that session.
3. Permanently, set "model": "sonnet" in your settings file.
4. Check what you are on with /status.

Effort. /effort sets how much the model reasons, starting from each model’s default in the spec table above. The docs spell out the bill: “You are charged for all thinking tokens generated, even when collapsed or redacted”. They also say the effort scale “is calibrated per model, so the same level name does not represent the same underlying value across models”, which is one more reason to price a sample rather than compare settings by name.

Subagents. A custom subagent’s definition can set model: to sonnet, opus, haiku, fable, a full model ID or inherit. Anthropic’s subagent docs list “routing tasks to faster, cheaper models like Haiku” as one way subagents control costs.

A /model switch reaches subagents that inherit the main model. Since v2.1.198 the built-in Explore subagent inherits your main model too (capped at Opus on the Claude API), so to keep codebase exploration on Haiku you define your own subagent named Explore with model: haiku.

The full list of session commands, including /model, /effort and /status, is in my Claude Code commands reference.

Claude Sonnet vs Opus: Price the Job Before You Pick

The Claude Sonnet vs Opus decision gets easier once each job has a number on it. Take one recurring marketing job, write down its words in and words out, and price it on Sonnet 5 and Opus 5.5 with the 1.8 tokens-per-word figure above.

If it runs in a single pass, start on Sonnet 5. If it loops through many steps or ships figures, run a sample on both at a fixed effort level and keep the model with the lower cost per finished job.