LLM tasks

Run a structured language task — classify, extract, keywords, sentiment, proofread, JSON repair, translate, summarize, answer — from a single API call. You pick a task; the router serves a model we’ve benchmarked against Claude Opus 4.8 and meters the spend. Pricing is cost-plus: you pay the actual provider cost plus a 6% markup, itemized on every receipt — no token-math homework, no marked-up-by-3x reseller surprise.

curl -X POST https://api.relaystation.ai/v1/llm/classify \
  -H 'Authorization: Bearer rs_live_<key>' \
  -H 'Idempotency-Key: classify-ticket-20260623' \
  -H 'Content-Type: application/json' \
  -d '{ "input": "My card was charged twice", "params": { "labels": ["billing","technical","other"] } }'

Or on the lodestone path — no account, a signed x402 payment instead of an API key:

curl -X POST https://api.relaystation.ai/v1/llm/classify \
  -H 'X-Payment: <base64 EIP-3009 authorization>' \
  -H 'Idempotency-Key: classify-ticket-20260623' \
  -H 'Content-Type: application/json' \
  -d '{ "input": "My card was charged twice", "params": { "labels": ["billing","technical","other"] } }'

The tasks

The registry is live — tasks are admin-managed rows. Query the live catalog any time: GET /v1/llm/tasks is free, read-only, no auth, and lists each task’s params, output cap, and ceiling. The nine benchmarked tasks (seven we run, two cited):

Task	What it does	Route
Classify text	Assign one label from a set	`POST /v1/llm/classify`
Extract structured data	Pull requested fields out as JSON	`POST /v1/llm/extract`
Repair JSON	Fix malformed JSON	`POST /v1/llm/json_repair`
Extract keywords	Pull key terms as a list	`POST /v1/llm/keywords`
Proofread text	Fix spelling, grammar, punctuation	`POST /v1/llm/proofread`
Sentiment analysis	Positive / negative / neutral	`POST /v1/llm/sentiment`
Translate text	Translate to a target language (benchmarked for gist/comprehension)	`POST /v1/llm/translate`
Summarize text	Summarize to key points (cited: Vectara HHEM)	`POST /v1/llm/summarize`
Answer from context	Grounded question answering (cited: FACTS Grounding)	`POST /v1/llm/answer`

Tiers and ceilings

The tier parameter (budget, value, best) sets the per-increment ceiling you authorize — the most a call can cost. For these nine benchmarked tasks we serve a single validated model that already wins on accuracy and cost (for translate, where comprehension scores tie statistically, the most production-proven model), so every tier resolves to the same model; the tier you pass just caps the spend. You pay the real metered amount, which is almost always less. Ceilings apply per 20 KB of input (input is capped at 100 KB):

Tier	Ceiling (per 20 KB increment)
budget	$0.010
value	$0.050
best	$0.250

Receipt anatomy — what cost-plus means

The charge is ceil(provider usage cost × 1.06). The response’s usage object itemizes the provider’s reported cost and the 6% markup separately, so the math is auditable line by line. On the x402 path the published ceiling is what you sign; the settled amount is the actual cost-plus figure, never more than the ceiling. If the call fails before producing output, the hold is released — you pay for delivered tokens, nothing else. The field-by-field receipt anatomy (and how ceiling-vs-charged works across payment modes) is on the canonical Receipts and cost-plus page.

Which model serves each task

We benchmark each task against Claude Opus 4.8. Seven we run ourselves — six objective tasks on 100 inputs + translate on a gist/comprehension eval (Belebele, n=50 per language direction) — and publish every input, output, and score to download. Two (summarize, answer) we cite from the most reliable public board (Vectara HHEM and DeepMind FACTS Grounding) and add our own cost column. We serve the model that wins on accuracy and cost, typically ≈Opus’s accuracy at ~1–5% of its cost. See the benchmark for the per-task leaderboards, downloads, and cited-source links.

MCP tools

Every task is callable over MCP at https://api.relaystation.ai/mcp as llm_<task> — llm_classify, llm_extract, llm_json_repair, llm_keywords, llm_proofread, llm_sentiment, llm_translate, llm_summarize, llm_answer. Same auth, same cost-plus pricing as the HTTP routes.