LLM API Cost Calculator

How Is LLM API Cost Calculated?

Large language model APIs like GPT, Claude, Gemini and Grok are not billed per request, they are billed per token. Every call consumes two kinds of tokens: the input (prompt) tokens you send to the model, and the output (completion) tokens the model generates. Providers price these separately, and output tokens are almost always 4 to 8 times more expensive than input tokens, because generating text costs far more compute than reading it.

This tool computes cost per request as: (input tokens / 1,000,000 x input price) + (output tokens / 1,000,000 x output price). Daily cost multiplies that by your daily request count, and monthly cost multiplies the daily figure by 30. By entering your real production traffic (average prompt length, average response length, daily request volume) you can estimate your monthly API bill before shipping a feature. To measure token counts precisely, try our Token Counter tool.

Model Price Table (per 1 Million Tokens, USD)

The prices below were checked against the providers' own pricing pages on 5 September 2026. Providers change rates without notice and promotional rates expire, so confirm with the provider before making a decision.

Related reading

DeepSeek V4.1 Flash takes two rows here because DeepSeek bills by the clock: 01:00-04:00 and 06:00-10:00 UTC on weekdays are peak hours, everything else (weekends included) is half price. A batch job scheduled outside that window runs the same model on the same work at $0.15/$0.60. The architecture, the benchmark numbers and the retirement of V4 Pro are covered in our DeepSeek V4.1 Flash write-up.

GPT-6 Astra and Claude Fable 5.1 sit on the same row here: $10 in, $50 out. But Fable's cache reads are four times cheaper, and Astra raises its rate on requests above 272K tokens. We work out how one price tag turns into two very different bills across three scenarios in our GPT-6 Astra vs Claude Fable 5.1 comparison.

Muse Spark 1.3 appears twice, because Meta sells it through two endpoints: standard at $1.25/$4.25 and Contributor at $0.10/$0.20. The price of that 12.5x cut is permission for Meta to train on your prompts and completions. We cover both tiers, the gated max mode and the benchmark scores in our Muse Spark 1.3 writeup.

The most expensive OpenAI model on the list is GPT-6 Astra: $10 in and $50 out per million tokens, 2.5x GPT-5.6 Sol. OpenAI's answer is that "the price per task is what matters". We cover its ARC-AGI-3, DeepSWE and OSWorld scores, the Astra Pro split and its Critical cyber classification in our GPT-6 Astra writeup.

Qwen3.8-Max has no tiered pricing, and repeated input drops to $0.25. We cover its pricing, context window and benchmark scores in our Qwen3.8-Max review.

The cheap end of the same family is Qwen3.8-Flash: $0.16 in and $0.47 out per million tokens, roughly a twelfth of Max. We cover its architecture, 1M token context and benchmark scores in our Qwen3.8-Flash-Next writeup.

The cheapest natively multimodal option on the list is GLM-5.3-Flash: list price is $0.15 in and $0.50 out per million tokens, but Z.ai currently runs a 50% discount, so you actually pay $0.075 and $0.25. The calculator uses the discounted rate. We cover its 320B total / 18B active architecture and its benchmark scores in our GLM-5.3-Flash writeup.

Gemini Omni 1.1 Flash is the only video-generating model on the list, which is why it has two rows: input is $1.50 per million tokens either way, while output is $9 for text and $17.50 for video. Video is billed as tokens per second: one second of 720p is 5,792 tokens, roughly $0.10. The 360p draft mode costs a third of that ($0.03/s), 1080p is $0.15/s and 4K is $0.30/s. We cover its 40-second scene extension and first/last frame control in our Gemini Omni 1.1 Flash writeup.