Priced 8 SEO Jobs on Every Claude Model: Here’s the Bill

October 14, 2026
Artificial Intelligence, SEO

Claude for SEO · token economics · October 2026

Claude for SEO: what crawl audits, keyword clustering and content briefs really cost, by model and effort

We measured how many tokens real SEO data eats (crawl exports, GSC, keyword lists, server logs, competitor pages), then priced eight everyday SEO jobs on Claude Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.1 at every effort level. Here's what each job costs, which model and effort to run it on, and when Extra high effort actually earns its keep.

Test job: auditing a 920-URL Screaming Frog crawl (200K tokens)

Token sizes: measured on real export formats

Prices: Anthropic API list rates, Oct 14 2026

Key takeaways for SEOs

Pick the model by job, not by habit. Bulk titles and metas on Haiku 5.5 cost $0.11 per 1,000 pages. The same job on Opus 5.5 at High: $5.18.

Trim exports before you paste. 15 Screaming Frog columns instead of ~70 cuts tokens 52%. Five Semrush columns instead of nine cuts 78%.

Never paste raw HTML or raw logs. Raw HTML runs ~89× the tokens of extracted text. A day of logs for a busy site is ~29× the 1M context window, so have Claude write the parser (about $0.30).

Cache big exports. Ten follow-up questions on a cached crawl cost $1.36 in input instead of $8.00.

Ask for IDs, not URLs. Redirect maps returned as row IDs cost $3.34 instead of $11.63 for 3,000 URLs.

Save Extra high for agent runs. On a one-shot crawl audit it adds 34% to the bill for 2 benchmark points. On 30+ minute Claude Code runs it's where the payoff is.

Claude Haiku 5.5

$0.10 / $0.50

per 1M tokens, input / output

Cached input read: $0.01 / $0.05
Speed at High: 178 tokens/sec
Default effort: Medium
Context window: 1M tokens

Bulk SEO workhorse. Keep each request under 100K tokens: above that the rate jumps 5× to $0.50 / $2.50.

Claude Sonnet 5.5

$2 / $10

per 1M tokens, input / output

Cached input read: $0.10
Speed at High: 92 tokens/sec
Default effort: High
Context window: 1M tokens

Content and schema model. Answers start in about a second at Low and Medium effort.

Claude Opus 5.5

$4 / $20

per 1M tokens, input / output

Cached input read: $0.20
Speed at High: 83 tokens/sec
Default effort: Medium
Context window: 1M tokens

Audit, migration and agent model. #1 on the Artificial Analysis Intelligence Index at Max effort.

Claude Fable 5.1

$10 / $50

per 1M tokens, input / output

Cached input read: $0.25
Speed at High: 55 tokens/sec
Default effort: High
Context window: 1M tokens

Premium tier. On Max plans capped at 50% of your weekly limit; on Pro it bills usage credits.

Fewer tokens: Screaming Frog export trimmed to 15 key columns

52%

453 tokens per URL with every column, 217 trimmed.

Fewer tokens: Semrush keyword export cut to 5 columns

78%

122 tokens per keyword down to 27. 10,000 untrimmed keywords won't even fit in 1M context.

More tokens: raw HTML vs the page's extracted text

89×

Median of 4 real homepages tested (range 63× to 289×). Feed text, not source.

Cheaper input: 10 questions on a cached 200K crawl

83%

$1.36 vs $8.00 on Opus 5.5. Cache reads are 95% off on Opus and Sonnet 5.5.

One crawl audit, every model, every effort level

The test job: a 920-URL crawl export (200K tokens) audited in one request. The pale part is reading the crawl (fixed per model). The solid part is everything Claude writes, including hidden thinking, which is billed at the full output rate.

Input: the crawl exportOutput: thinking + audit (shade follows the model)

Swipe sideways to see the full chart →

Tokens Claude generates per audit

Thinking + answer, by effort level. Extra high roughly doubles High. At Max, Haiku and Sonnet blow past what fits in a single response.

Haiku 5.5Sonnet 5.5Opus 5.5Fable 5.1

Swipe sideways to see the full chart →

How many crawl audits $100 buys

Budget left after each 920-URL audit at High effort. The dot on the axis is where the money runs out.

Haiku 5.5Sonnet 5.5Opus 5.5Fable 5.1

Swipe sideways to see the full chart →

Burn rate grid: cost per audit, audits per $100, wall-clock time

Darker = drains faster. The multiplier compares each cell to Sonnet 5.5 at High (1.0×). On Pro and Max plans Anthropic doesn't publish per-model multipliers, but usage tracks compute, so this is the best proxy for how fast each setting eats your limit.

Swipe sideways to see the full chart →

The efficiency frontier, mapped to SEO work

Intelligence (Artificial Analysis Intelligence Index) against cost per benchmark task. Points on the dark staircase are the efficient choices: nothing else is both smarter and cheaper. Each model's points run Low to Max effort, left to right. Numbered pins match the SEO job price list below.

Haiku 5.5Sonnet 5.5Opus 5.5Fable 5.1

Swipe sideways to see the full chart →

SEO jobJob cost

Haiku 5.5 · Low

Bulk title tags, meta descriptions, alt text, product copy tweaks

$0.11

Haiku 5.5 · Medium to High

Keyword intent tagging, clustering into a fixed topic list, SERP feature extraction

$0.22

Sonnet 5.5 · Medium

JSON-LD schema, hreflang tables, redirect regex, quick GSC questions. Answers start in about a second

$0.45

Sonnet 5.5 · High

Content briefs, article drafts, on-page rewrites from SERP research

$0.36

Opus 5.5 · Medium

Crawl export audits, GSC trend analysis, internal linking plans

$1.15

Opus 5.5 · High

Migration redirect mapping, cannibalization, canonical and hreflang conflicts

$3.34

Opus 5.5 · Extra high

Agent runs in Claude Code: fix templates, re-crawl staging, validate schema, repeat

$2.73

Opus 5.5 · Max

Go/no-go calls: migration QA, traffic-drop diagnosis where a miss costs rankings

$4.82

Where's Fable 5.1? Off the frontier at every effort level. Opus 5.5 scores higher for less money across the board, so for SEO work it's hard to justify. Keep it for long, autonomous projects where your own side-by-side tests show it wins.

Token cheat sheet: what your SEO data costs to feed Claude

Tokens per row or page on the current Claude tokenizer, how much fits in a 1M-token context window, and what it costs Opus 5.5 to read 1,000 of them. Trim before you paste.

Swipe sideways to see the full table →

SEO data	Tokens each	Fits in 1M context	Opus 5.5 cost per 1,000
Screaming Frog Internal HTML, all ~70 columns✕ Avoid	453 / URL	2,207 URLs	$1.81
Screaming Frog, 15 key columns✓ Use this	217 / URL	4,615 URLs	$0.87
Migration rows: URL, title, H1	86 / URL	11,576 URLs	$0.35
XML sitemap <url> entries	75 / URL	13,290 URLs	$0.30
GSC performance rows (query + page)	76 / row	13,146 rows	$0.30
Semrush keyword export, all 9 columns✕ Avoid	122 / keyword	8,199 keywords	$0.49
Semrush, 5 columns (keyword, intent, volume, KD, CPC)✓ Use this	27 / keyword	36,763 keywords	$0.11
Server access log, combined format	144 / line	6,958 lines	$0.57
Competitor article as extracted text (2,000 words)✓ Use this	3,126 / page	319 pages	$12.50
Web page as raw HTML (median of 4 real pages)✕ Avoid	66,589 / page	15 pages	$266.36

SEO job price list: right model, right effort, real cost

Eight common jobs, sized from the measurements above. Cost includes Claude's hidden thinking at that effort level. The Batch API halves the bill for anything that can wait for results.

Swipe sideways to see the full table →

#	SEO job	Run it on	Tokens in / out	Cost	Batch API
1	Bulk title tags + meta descriptions1,000 product pages, 4 chunks of 250Haiku 5.5 · Low	Haiku 5.5 · Low	230K / 174K	$0.11	$0.06
2	Keyword intent + cluster assignment10,000 keywords (5-column Semrush export), 4 chunksHaiku 5.5 · High	Haiku 5.5 · High	280K / 387K	$0.22	$0.11
3	JSON-LD schema for page templates40 templates, one Product/Article block eachSonnet 5.5 · Medium	Sonnet 5.5 · Medium	50K / 35K	$0.45	$0.23
4	Content brief + 2,000-word draftTop-10 competitor pages as extracted textSonnet 5.5 · High	Sonnet 5.5 · High	34K / 30K	$0.36	$0.18
5	Technical audit of a crawl export920-URL Screaming Frog crawl, 15 columnsOpus 5.5 · Medium	Opus 5.5 · Medium	200K / 18K	$1.15	$0.57
6	Migration redirect mapping3,000 old URLs to 3,000 new (URL, title, H1)Opus 5.5 · High	Opus 5.5 · High	520K / 63K	$3.34	$1.67
7	Agentic fix run in Claude Code40 tool turns: edit templates, re-crawl staging, validate schemaOpus 5.5 · Extra high	Opus 5.5 · Extra high	4.51M / 46K	$2.73	Not batchable
8	Migration go/no-go QARedirect map + post-launch crawl + GSC indexing exportOpus 5.5 · Max	Opus 5.5 · Max	607K / 120K	$4.82	$2.41

Overkill check: jobs 1 and 2 on Opus 5.5 at High would cost $5.18 and $7.25, 47× and 33× more for work Haiku handles. Raw HTML check: job 4 fed raw competitor HTML instead of extracted text costs $1.63, not $0.36. Haiku jobs are split into chunks under 100K tokens to stay on its cheaper tier.

High → Extra high: what's going on under the hood

Opus 5.5 measured

Same model, same weights, same price per token. Effort is a behavior the model was trained to follow, not a hard token budget: it sets how readily Claude thinks, how long it thinks, and how much checking it does before it commits to an answer.

Swipe sideways to see the full chart →

Crawl audit, Opus 5.5 at High

$1.29

24K tokens out · about 5 min

Same audit at Extra high

$1.72

46K tokens out · about 9 min

What actually changes

Thinks more often. At High, Claude thinks on most requests that benefit. At Extra high it thinks more readily, so borderline calls (is this canonical intentional?) get reasoning too.

Thinks longer. More hypotheses for why a page isn't indexed, more cross-checks against the crawl, more "let me verify that" passes.

Acts more. As an agent it fetches more URLs, re-runs the crawl, re-validates schema and checks one more template before it says done.

You pay for all of it. Thinking is billed as output tokens at the full rate, even though you only see a summary.

Why it gets better

Long jobs are chains of steps. In a 50-step migration QA run where each step is 98% reliable, the whole run finishes clean only 36% of the time. At 99.5% per step it's 78%.

Extra high buys per-step reliability. That compounds on long agent runs and barely moves a one-shot audit, which is why the payoff shows up past the 30-minute mark.

Benchmark proof: Opus 5.5 at Extra high scores 66.4% on Terminal-Bench 4.0, a long agentic coding test (Anthropic).

Where Extra high pays off in SEO

Claude Code runs that edit theme templates, re-crawl staging and re-test until clean

Building a log-file pipeline: write the parser, verify Googlebot by reverse DNS, chase edge cases

Migration QA where one missed redirect chain costs rankings

Where it's wasted

Meta descriptions, alt text, intent tagging, redirect regex

Structured output like CSV or JSON-LD: it can overthink and you pay for nearly 2× the output tokens for the same answer

Max on top of that: 2.6× more tokens than Extra high for +2 points. Save it for go/no-go calls

Gotchas. Switching effort mid-audit resets the prompt cache: on a 200K crawl export on Opus that's a $1.00 re-write instead of a $0.04 cached read. Set max_tokens to 64K or more at Extra high and Max, or long thinking gets cut off. Follow-up turns that only process tool results (a crawl, a validator report) can skip thinking even at Extra high, so agent runs cost less than the first turn suggests.

FAQ: Claude for SEO work

How much does it cost to analyze a Screaming Frog crawl with Claude?

A trimmed Internal HTML export runs about 217 tokens per URL (453 with every column). A 920-URL crawl is about 200K tokens and costs $1.15 to audit on Opus 5.5 at Medium, or $0.64 on Sonnet 5.5 at High. Past about 4,600 URLs a trimmed export won't fit in the 1M context window, so filter to issue rows or have Claude write the analysis script.

Which Claude model is best for SEO?

Use Haiku 5.5 for bulk work (titles, metas, alt text, intent tagging), Sonnet 5.5 for content briefs, drafts and schema, and Opus 5.5 for audits, migrations and agent runs. Fable 5.1 doesn't beat Opus 5.5 on measured intelligence or cost at any effort level.

Is Extra high effort worth it for SEO audits?

For a one-shot audit of an export, usually not: on the crawl audit it adds 34% to the cost ($1.29 to $1.72) for a small quality gain. It pays off when Claude works as an agent for 30+ minutes, like fixing templates in Claude Code, re-crawling and re-testing until clean.

How many URLs fit in Claude's 1M-token context window?

About 4,615 URLs of a 15-column Screaming Frog export, 2,207 with every column, 13,146 GSC rows, 36,763 trimmed Semrush keywords, 319 competitor articles as text, or only about 15 pages of raw HTML.

How do I cut Claude API costs on SEO jobs?

Trim export columns, feed extracted text instead of HTML, ask for row IDs instead of full URLs, cache any export you'll ask about more than once, send bulk jobs through the Batch API (50% off), and keep Haiku 5.5 requests under 100K tokens.

Can Claude analyze server log files for crawl budget?

Not by pasting them. One day of logs for a site with 200,000 hits is about 29M tokens, ~29× the context window. Have Sonnet 5.5 write a parser that verifies Googlebot and aggregates hits by template and status code (about $0.30), then send Claude the summary.

Method. Token sizes were counted with Anthropic's tokenizer on sample exports built in the real Screaming Frog (Internal HTML), Google Search Console, Semrush keyword, combined access log and XML sitemap formats, plus four real homepages for raw HTML vs extracted text, then scaled +30% for the Claude 4.7+ tokenizer per Anthropic's pricing docs. Job output = the deliverable's size + Artificial Analysis's measured average output per task (thinking + answer) for that model and effort, Oct 2026 (about 2,176 tasks per run); for report-style jobs the measured average covers the report. Input billed at list price, global endpoint. Haiku 5.5 requests over 100K tokens pay its higher tier. The agentic run assumes 40 turns over a cached context growing from 30K to 190K tokens. Time = output tokens ÷ measured output speed. Burn multipliers, plan-usage reading and the reliability chain are estimates or illustrations.
Sources. Anthropic: Pricing, Effort, Steering thinking and cost, Claude Opus 5.5 overview, Introducing Claude Opus 5.5, Claude Sonnet 5.5 overview. Claude Help Center: Change the model, effort and thinking settings, Claude Fable models on your plan. Artificial Analysis model pages for Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.1 at each effort level.

SEO MEDITATIONS

If AI can't find you, understand you and trust you, it won't recommend you.

Keywords tell you what they typed. Intent tells you what they meant. Fan-out tells you what they'll ask next.

Be a thing, not a string. AI remembers entities, not keywords.

The click isn't dead. It just has to be earned twice.

Schema is the business card you hand to every machine that reads your site.

If Googlebot can't reach it and AI can't parse it, it doesn't exist.

Good SEO grows quietly. Then all at once.