Skip to content
AnglinAI Docs

Cost Model — Accessibility Skill

What it costs to run one accessibility audit + ACR through the chat skill, and how to keep the numbers current as models and prices shift.

There are two distinct cost buckets that hit different bills:

  1. Platform side (your Anthropic API bill) — server-side LLM calls made by api.theaccessible.org.
  2. User side (the user’s Claude bill) — orchestration tokens consumed by Claude Desktop / Claude.ai / Claude Code while running the skill.

These are paid by different accounts. Don’t conflate them when modeling unit economics.

Per-page cost — platform side

The URL scan pipeline is almost entirely deterministic. Server-side LLM cost per page is just the AI pre-grader.

ComponentLLM?ModelNotes
URL fetch executor (workers/batch/src/url-fetch-executor.ts)no—Puppeteer + axe-core + cms-detector
Auto-fix loop (workers/api/src/services/axe-fixer.ts)no—Deterministic regex-based fixes; despite the name, no LLM call
AI pre-grader (workers/api/src/services/verification-prograder.ts)yesHaiku 4.5One call per not-verified criterion
ACR composer (workers/api/src/services/acr-composer.ts)no—HTML template + weasyprint sidecar

So per scan, the only LLM bill is the pre-grader.

Pre-grader cost arithmetic

VariableValueSource
Modelclaude-haiku-4-5-20251001verification-prograder.ts:19
Input rate$1 / MTokAnthropic published rate (Haiku 4.5 release pricing)
Output rate$5 / MTokAnthropic published rate
Calls per page11–18 (typ. 15)One per not-verified criterion. Range observed in real scans.
Input tokens / call~1,500Artifact JSON (truncated to 8KB max, typical 1–3KB) + system prompt + criterion description
Output tokens / call~150Structured JSON verdict + brief reasoning

Cost per call: 1,500 × $1/M + 150 × $5/M = ~$0.0023

Cost per page (15 calls): ~$0.034

At scale

Pages / monthPlatform LLM cost
100~$3.40
1,000~$34
10,000~$340
100,000~$3,400

These are LLM costs only — they don’t include Puppeteer compute, R2 storage, weasyprint sidecar, or other infra. Those are tracked separately (see aws-deploy-next-steps.md and the cost-tracking follow-up in accessibility-skill.md).

Per-walkthrough cost — user side

A full skill walkthrough (scan → verify ~15 items conversationally → propose fixes → generate ACR) consumes meaningful tokens in the user’s Claude account because tool result payloads (HTML, queue items, conformance reports) inflate the input side fast.

Standard rates (current published list)

ModelInput ($/MTok)Output ($/MTok)
Haiku 4.5$1$5
Sonnet 4.6$3$15
Opus 4.7$15$75

Walkthrough scenarios

PatternInput tokensOutput tokensSonnet 4.6Opus 4.7
Single page, full walkthrough (verify all + ACR)~200K~50K~$1.35~$6.75
Scan + summary only (no verification)~30K~5K~$0.17~$0.83
Scan + auto-approve high-confidence + generate~80K~15K~$0.47~$2.30

Most Claude Pro/Max users won’t see this directly — it eats into their monthly cap. API users see it on their bill.

Phase 6 (screenshots) — incremental cost

Per-element JPEG screenshots inlined into acr.queue responses for visual verification.

  • Server side: zero new LLM calls (Puppeteer + R2 only).
  • User side: 15 items × ~10KB JPEG ≈ 150KB total per response, ~6K extra input tokens. Adds:
    • ~$0.018 / walkthrough on Sonnet 4.6
    • ~$0.090 / walkthrough on Opus 4.7

Negligible compared to the visual-context value for criteria like 1.1.1 (alt text), 1.4.1 (color use), 1.4.3 (contrast).

Pricing implications

If you’re modeling a per-audit charge, the platform cost floor is ~$0.034 LLM + small infra ≈ a few cents. Reasonable per-audit pricing tiers:

TierPer-audit priceLLM cost as %Notes
Self-serve$0.50–$22–7%Dashboard or skill
Pro / agency$5–$15<1%Bulk + sign-off authority
Enterprise$50–$200<0.1%SLA + dedicated review

The platform LLM cost is too small to gate aggressive pricing. Compute (Puppeteer rendering on .4 / EC2) and the cost of human review time on your side are the real cost drivers, not the AI bill.

Optimization opportunities

OpportunityMechanismEstimated savings
Prompt cachingAnthropic’s prompt cache — system prompt + criterion descriptions are identical across all pre-grader calls~80% of input cost on cached prefix → drops pre-grader to ~$0.01/page
Batch APIAnthropic Batch API at 50% discount for non-urgent pre-gradingHalves pre-grader cost; trades latency (24h vs seconds)
Conditional pre-gradingSkip the pre-grader for criteria the user has decided in a prior scan of the same URLSaves ~30–50% of pre-grader calls on repeat scans
Confidence threshold tuningCurrently ≥0.85 && verdict !== 'partial' auto-decides. Lowering to 0.80 would auto-decide more items without re-promptingMarginal pre-grader cost change; bigger UX win

None of these are wired up yet. Worth revisiting once you have ≥1,000 pages/month of real traffic to optimize against.

Keeping this document current

  • When the pre-grader model changes: update verification-prograder.ts:19 AND the rates in this doc’s pre-grader table.
  • When Anthropic adjusts published rates: update both rate tables. Re-derive the per-page number.
  • When real traffic data is available: replace the typed-15-criteria estimate with measured medians from human_verification row counts per job_id. SQL: 
    SELECT percentile_cont(0.5) WITHIN GROUP (ORDER BY n) AS median_calls
    FROM (
      SELECT job_id, COUNT(*) AS n
      FROM human_verification
      WHERE source IN ('ai-auto', 'ai-suggested-human-confirmed', 'ai-suggested-human-overrode')
      GROUP BY job_id
    ) t;
  • When prompt caching is enabled: revisit “Optimization opportunities” — savings should land in the main per-page table.
  • docs/admin/accessibility-skill.md — full architecture and operations reference
  • apps/web/content/guides/claude-desktop-skill.md — end-user install + usage guide