Customer support
10,000 requests / month · 1 call per attempt
1,960 input + 420 output tokens / call, estimated from 1,400 and 300 words.
Try customer support exampleKINGY AI / MODEL ECONOMICS
AI API Cost Calculator
Compare text API costs across major model providers. Start with your workload or try an example below.
Estimate a multi-model workflow
Add up to five stages using selected model routes. Enter monthly calls for each stage, including retries and failures. All stages use the cache, surcharge, other-charge and scheduled-pricing settings above. These temporary stages are not included in saved setups, share links or comparison exports; use the separate workflow download below.
Add a stage to begin.
Your estimated monthly API bill
Compare up to five models. Prices do not establish quality on your task.
Browse or replace comparison models
Find a model
The shortlist is an editorial starting point across providers and workloads, not a quality ranking. Categories suggest models to investigate; suitability remains unmeasured unless a matching test is shown. Search covers all routes unless other filters limit it. Capability filters include only routes with recent source evidence; absence does not establish that a capability is unsupported.
OpenRouter prices are its published top-provider rates. Direct API quotes are labeled separately. Lowest token cost does not establish best quality. Invalid or unsupported scenarios have no estimated price.
Loading source coverage…
While this page is visible, current public source evidence is rechecked every five minutes. Review dates advance only when the maintained source evidence does.
Plan a billing period, service tier and currency
What could your workload cost?
Three editable examples use the same maintained prices as the calculator. These are estimates, not measured workloads or invoices.
Document extraction
1,000 documents / month · 1 call per attempt
2,100 input + 280 output tokens / call, estimated from 1,500 and 200 words.
Try document extraction exampleCoding
500 tasks / month · 3 calls per attempt
8,400 input + 2,800 output tokens / call, estimated from 6,000 and 2,000 words.
Try coding example| Direct API model | Support | Extraction | Coding | Source |
|---|---|---|---|---|
| gpt-6.1-solOpenAI direct · Standard globalVerified 2026-11-12T04:20:04Z Review due 2026-11-12T10:20:04Z | $81.20 | $7.00 | $67.20 | Official pricing |
| Claude Sonnet 5.5Claude direct · Standard globalVerified 2026-11-12T04:20:12Z Review due 2026-11-12T10:20:12Z | $81.20 | $7.00 | $67.20 | Official pricing |
| Gemini 3.8 FlashGemini direct · Paid StandardVerified 2026-11-12T04:20:19Z Review due 2026-11-12T10:20:19Z | $30.45 | $2.63 | $25.20 | Official pricing |
Uncached text, standard pricing, all attempted calls included. No added reasoning tokens, tools, taxes or labor. Word counts use 1.4 estimated tokens per word. Expired or held prices are unavailable. These examples do not measure model quality.
How to estimate AI API costs
For uncached text, multiply input tokens by the input price, add output tokens multiplied by the output price, then multiply by monthly calls. Prices quoted per million tokens must first be divided by 1,000,000.
Include the whole request. System instructions, conversation history and repeated documents can all count as input. Word counts here use an editable estimate of 1.4 tokens per word; language, formatting and model tokenizers change the actual count. Use provider-reported tokens when available.
Count retries and agent steps. Each billed call contributes to the monthly total, including failed attempts. The coding example includes three calls per task. Caching, extra reasoning, paid tools, regional charges and taxes can change the bill; add applicable charges in the advanced settings.
Check quality separately. A lower token price does not show whether a model can do your job. The Kingy test costs tab contains dated coding and math results; those results do not measure the support or extraction examples.
Compare GPT-6.1 Sol with Claude Opus 5.5 · Compare Claude Sonnet with Opus
RETAINED MEASUREMENTS
Cost per strict pass
Import your evaluation locally
Import a JSON record containing exact model and route, a task ID and SHA-256 of its definition, test time, declared deterministic scoring, and every attempt including failures. Use user_estimate or user_reported_billed for costBasis; use null when an attempt cost is unknown. Do not include prompts, responses, API keys or personal content. Records stay on this device and are not merged with Kingy measurements or saved automatically. Token usage is recorded; user-supplied costs are not repriced.
View input format
{"schema":"kingy-task-evaluation/1.0.0","modelVersion":"model-version","billingRoute":"provider standard","taskId":"task-version","taskDefinitionSha256":"0000000000000000000000000000000000000000000000000000000000000000","testedAt":"2026-10-31T00:00:00Z","scoringType":"deterministic_checks","scoringMethod":"schema and exact-value checks","costBasis":"user_estimate","attempts":[{"id":"attempt-1","status":"pass","inputTokens":100,"outputTokens":10,"reasoningTokens":0,"costUsd":0.01}]}No evaluation imported.
Total estimated provider charges for every attempt, including failures, divided by the number that passed the deterministic checks.
These costs use recorded token usage and the rates retained with each run. Settled invoice attribution is unavailable. A strict pass is a test outcome, not human acceptance of a production answer.
12 tasks × 3 repetitions per model and category. No LLM judge. Provider-specific medium reasoning settings are not equal compute. These estimates are not eligible for a billed-cost winner. Different cohorts are not paired comparisons; v1 and v2 also use different prompts and scoring. Hardware, human repair, taxes and tool fees are outside this cost boundary. The historical ledger rounds each attempt upward to one microdollar. No new inference was run for this calculator.
DeepSeek pricing is included above. This retained coding/math cohort has no DeepSeek acceptance-and-cost record, so its cost per accepted answer is unmeasured here. Inspect other retained task-cost evidence ↗
PRICING PROVENANCE
Inspect the snapshot
Each route retains its source link, pricing fields, model ID, context limits and verification date. These are text-token estimates: multimodal input billing, tools, regional uplifts, purchase fees and special service tiers can change your bill.
OpenRouter's base rates come from its top provider and may change when the route changes. Conditional long-context and UTC schedules are retained. Conservative mode uses a per-field ceiling across the published rates applicable to your input length; it may overestimate a particular time window.
Official sources
- OpenRouter live model and price API · units, routing and override rules · purchase and service fees
- OpenAI direct API pricing
- Claude direct API pricing and cache TTLs
- Gemini Developer API pricing and announced changes
- DeepSeek direct API pricing and peak hours
Previous calculator snapshot
The earlier 12-model snapshot is preserved with its original dates and expiry. It is historical evidence, not a current fallback price.
Download the previous 12-model ratesDownload the raw verified OpenRouter response
Saving is opt-in and can be removed. Share links include workload settings and selected model IDs. Anonymous aggregate completion counts are recorded after you calculate a valid result. Workload inputs, selected models and results stay private; no model requests are made. Hosting security cookies may still apply.
Other API costs
Estimate selected image-generation token usage, transcription, speech, embeddings or OCR using dated provider list prices. Calculations stay on this device. This is an estimate, not an invoice.
These separate modes cover listed units only; each route states its exclusions. Source evidence expires and is withheld when review is overdue. These values are not included in text-comparison exports or saved setups.