AI Cost Calculator for LLM API Calls

Multiply input tokens by the model input rate, output tokens by the output rate, and divide by 1,000,000. On Gemini 3.6 Flash ($0.75 in / $3.75 out) a 4,700-input, 390-output call is (4,700 × 0.75 + 390 × 3.75) ÷ 1,000,000 = $0.00499. Every provider bills in exactly that shape, so the only hard part is knowing your real token size.

That is a measured production request for a planning or tool-calling task: a large system prompt plus tool schemas going in, a short structured decision coming back. It is deliberately not a chat workload, which inverts the ratio and changes which model is cheapest. Move the sliders to match your own traffic — and expect real calls to run 1.5 to 2 times larger than a thin test sample.

Generating tokens is sequential, while input can be processed in parallel, and providers price accordingly. The ratio is typically 5× on Anthropic, Google Flash and Kimi K3, 6–8× across the OpenAI GPT-5 line, and 3× on DeepSeek V4. Because reasoning tokens are billed as output, a model that always thinks can cost more than its visible answer suggests.

Because that provider does not publish a flat per-token cache-hit rate. OpenAI, Anthropic, Moonshot, DeepSeek and Z.ai all do, so the slider applies their real discount — typically 0.1× the base input rate, and as low as 0.03× on DeepSeek. Only Google is disabled here: its context-caching price carries a separate per-hour storage charge, so a single per-token rate would understate the real cost.

At the default call size, gpt-5-nano ($0.05 / $0.40 per million) and GLM-4.7-FlashX ($0.07 / $0.40) are the cheapest paid options, followed by GLM-4.5-Air and gpt-5.6-luna at around $0.0014 a call. GLM-5.3-Flash ($0.15 / $0.50) is new in that group. The DeepSeek Flash model left it in August 2026 at $0.44 / $1.32 and is back on its edge since September as deepseek-flash at $0.30 / $1.20. Cheapest only matters if quality holds for your task, so A/B a candidate on real requests before switching, and note that naming is not a price guide: Kimi K3 costs half as much again per call as Claude Sonnet 5.

Charge from the token counts the provider returns, not a flat price per request, because a flat rate always undercharges your largest calls. Weight output tokens by the model output-to-input price ratio, divide by a fixed tokens-per-credit constant, and settle after the call succeeds. One credit then backs a fixed amount of provider cost for any input/output mix.

Every rate was read from the provider's own documentation on 13 September 2026 — never from aggregator sites, which drift. Published pricing changes often, and introductory rates expire — Gemini 3.6, 3.7 and 3.8 Flash all double on 1 January 2027 — though not always: Claude Sonnet 5's planned move from $2/$10 to $3/$15 on 1 September 2026 was cancelled, and $2/$10 is now its standard rate. The change log on this page records every rate movement we have caught since 29 July 2026, including a 50% cut on Gemini 3.6 Flash and a roughly 3× rise on DeepSeek V4. Verify against the provider before you budget against any figure here.