Claude API Pricing per Token

What a Claude request actually costs: per-token math for all four Keyo lanes, worked examples, and how to check the bill before you send.

Per token, not per request

Claude-class APIs bill on tokens — roughly three-quarters of a word each. Rates are quoted per million tokens, so the per-token price is just that number divided by a million. On KeyoAPI the four Claude lanes cost:

Model ID	Input / output per 1M	Per 1K tokens in / out	Best for
`claude-sonnet-5`	~$0.88 / $4.41	$0.00088 / $0.00441	Everyday production chat
`claude-opus-5`	~$1.10 / $5.52	$0.0011 / $0.00552	Hard agent and coding escalations
`claude-fable-5-1`	~$7.35 / $36.77	$0.00735 / $0.03677	Planner-tier flagship work
`claude-fable-5`	~$8.82 / $44.12	$0.00882 / $0.04412	Flagship chat / agent twin

Rates are indicative sell prices from the live catalog — confirm the interactive rows on /pricing/claude-sonnet-5, /pricing/claude-opus-5, /pricing/claude-fable-5 and /pricing/claude-fable-5-1 before contracting volume. The full comparison table lives on /claude-api-pricing.

A worked example

Say a support bot sends a 2,000-token prompt and gets a 400-token answer on claude-sonnet-5:

input:  2,000 / 1,000,000 × $0.88  = $0.00176
output:   400 / 1,000,000 × $4.41  = $0.00176
                                     ---------
total per conversation             ≈ $0.0035

Ten thousand such conversations ≈ $35. The same conversation on claude-fable-5 would run ≈ $0.036 — ten times the cost for flagship quality you may not need on every turn. That ratio, not the headline price, is what per-token budgeting is about: route the routine 90% of traffic to the cheap lane and escalate only the hard cases.

Check the math on real traffic

Every response carries a usage object — read prompt_tokens and completion_tokens and multiply by the table above. Point any OpenAI SDK at the relay and pass a Claude model string:

curl https://www.keyoapi.xyz/v1/chat/completions \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"claude-sonnet-5","messages":[{"role":"user","content":"hi"}]}'

Cutting the bill

Escalate, don't default. Run drafts and classification on claude-sonnet-5; promote only failing slices to claude-opus-5 or the Fable lanes.
Shift bulk work off Claude entirely. High-volume drafts and classifiers often do fine on gpt-6-luna or deepseek-v4.1-flash on the same key — see /openai-api-pricing and /deepseek-api-pricing.
Prototype for free. Fixed $0 catalog IDs ending in :free bill against your signup gift credit at fair-use limits — draft and smoke-test there, then switch model= to a Claude lane: /free-models.
Retry around overload, don't re-send. Claude occasionally returns HTTP 529 overloaded; the retry and fallback playbook at /brand/blog/article/ncx234u2/ keeps those from double-billing you.

One prepaid balance covers every lane above plus GPT-class, DeepSeek, OCR and TTS — no separate Anthropic invoice. Base URL: https://www.keyoapi.xyz/v1.