bellalingua

For the task in front of you, the cheap model is probably good enough. Here’s the proof.

For nine discrete, bounded jobs — JSON repair, classify, keywords, sentiment, extract, proofread, translate, summarize, and answer — the small model we serve reaches ≈Claude Opus 4.8 quality at ~1–5% of the cost. Seven we benchmark ourselves against Opus on the same inputs (every input, output, and score published to download); the two cited tasks (summarize, answer) we take from the most reliable public board and add our cost column. Don’t take our word for it — check either way.

See the benchmark — per-task tables + downloadable test data →

Nine tasks — seven we benchmark against Opus 4.8, two we cite

Each task is served by a single model — the one that wins on accuracy and cost (for translate, where comprehension scores are statistically tied, the most production-proven model). Seven we run ourselves against Claude Opus 4.8; cited rows (summarize, answer) we take from the most reliable third-party board and add our cost column — summarize’s source (Vectara) did not test Opus 4.8, so it shows that source’s own metric. The full per-model leaderboards, downloads, and source links are on the benchmark page.

Reference = Claude Opus 4.8 (via OpenRouter), pinned at 100%. “Matched” means a measured tie on the published gold set, not a marketing claim — and where a number is ~99% we say ~99%, not “beats.” Rows tagged cited aren’t run by us: the accuracy is the third party’s own metric (answer = % of Opus on FACTS Grounding; summarize = Vectara factual-consistency, where Opus 4.8 wasn’t tested) and we add our cost column — see the benchmark for the source links. You pay is the per-million-token rate you’re actually billed: metered cost + 6% (the served model’s highest OpenRouter provider rate ×1.06, so the real charge lands at or below it — input tokens × the input rate + output tokens × the output rate, billed on tokens delivered). Live OpenRouter pricing, 2026-06-26.

Pick a task

Nine task endpoints — call POST /v1/llm/{task}; the live GET /v1/llm/tasks always shows the chosen model + per-call price. Pricing is cost-plus: you pay the actual provider cost + 6%, itemized on every receipt.

The economics

For bounded tasks you’re overpaying if you reach for the most expensive model by default. The model we serve costs ~1–5% of Opus 4.8’s per-call cost at measured-equal quality — the exact number per task is on the benchmark page, sourced from the published n=100 run. You pay the actual provider cost + a small markup, itemized on every receipt.

Methodology

n=100 inputs per objective task (translate: n=50 per language direction), every model on the same prompt at temperature 0, each example in isolation. Objective scoring against a gold set for each task (JSON-match, exact-match, F1@k, field-match, source-aware GLEU; translate = Belebele foreign→English comprehension via a fixed reader); the reference is Claude Opus 4.8, pinned at 100%. Incomplete runs are removed. Every claim traces to the published per-row data — download it and check the math.