LLM API Cost Calculator
Estimate your monthly API bill across Claude Opus 5.5, Claude Sonnet 5.5, GPT-6.1 Sol and Grok 4.5 at official list prices — including the effect of prompt caching — plus per-image costs for the image models on sale. Text-model rates below are the vendors' published list prices as cited on our model pages (checked 2026-10-16; Grok 4.5 as published by xAI in July 2026). No official price for Grok 4.7 or Gemini 3.8 Flash is documented here yet — see Google's Gemini API pricing page for Gemini.
Via byesu you pay a fraction of this list-price bill — billing uses transparent multipliers (input 1x · output 5x · cache hit 0.1x) and the console shows your live billed price next to the official list price.
See live prices in the console →List prices are the vendors’ published API prices (checked 2026-10-16; Grok 4.5 as published by xAI in July 2026) and may change; image prices are byesu’s per-image prices as of 2026-10-22, the same at 1K / 2K / 4K.
How to read the numbers
- Input / output split. Output tokens cost 3–5× more than input on every model here. Coding agents are output-heavy, so verbosity matters as much as the rate card: according to xAI, Grok 4.5 emits ~14k output tokens on task batches where Claude Opus 4.8 emits ~67k.
- Cache hit rate. At official list prices, cached input tokens bill at 5% of the input rate on Claude Opus 5.5, Claude Sonnet 5.5 and GPT-6.1 Sol, and 25% on Grok 4.5. Long system prompts, big repo contexts and multi-turn agent loops routinely reach 50–80% cache hit rates, which is why the "with caching" number drops so hard.
- Tier choice on GPT-6. GPT-6.1 Sol is listed at $2 / $10 per million tokens; GPT-6 Astra ($10 / $50) is OpenAI's most capable tier and GPT-6 Luna ($0.1 / $0.5) the high-volume tier. See the GPT-6.1 / GPT-6 page.
What you actually pay on byesu
byesu is an AI API gateway: one key covers all the models above through an OpenAI-compatible endpoint plus an Anthropic-native /v1/messages endpoint. Billing is pay-as-you-go with transparent multipliers:
| Billing item | Multiplier |
|---|---|
| Input | 1x |
| Output | 5x |
| Cache hit | 0.1x |
The calculator shows the official list-price bill; the console shows your live billed price next to the official list price after you sign in, so you can verify it per model before committing volume.
Top up with USDT or Alipay. Image requests that are not processed successfully are generally not billed.
Related pages
- Claude Opus 5.5 API
- Claude Sonnet 5.5 API
- GPT-6.1 / GPT-6 API
- Gemini 3.8 Flash API
- Grok 4.5 API — access & pricing
- Full pricing & billing overview
FAQ
How is LLM API usage billed?
Per token: one rate per million input tokens, a higher rate per million output tokens. Official list prices used by this calculator: Claude Opus 5.5 $4/$20, Claude Sonnet 5.5 $2/$10, GPT-6.1 Sol $2/$10, Grok 4.5 $2/$6 per million tokens.
How much does prompt caching save?
At official list prices, cached input tokens bill at 5% of the input rate on Claude Opus 5.5, Claude Sonnet 5.5 and GPT-6.1 Sol, and 25% on Grok 4.5. At a 60% cache hit rate, the input side of your bill drops by more than half — move the slider above to see the effect on your own volumes.
What do image generation APIs cost?
Image models are billed per image, the same price at 1K / 2K / 4K: on byesu, gpt-image-2, gpt-image-2.5 and gemini-3.0-pro-image-ultra are $0.009 per image, gemini-nano-banana-2.1-ultra $0.008 and gemini-3.1-flash-image-ultra $0.007 (prices as of 2026-10-22). No video model is on sale right now.
Does the calculator show what I would pay on byesu?
For text models it shows the official list-price bill; on byesu, billing uses transparent multipliers (input 1x, output 5x, cache hit 0.1x) and the console shows the live billed price next to the official list price after you sign in. Image prices shown are byesu's per-image prices.
Why do output tokens dominate the bill in agent workflows?
Output tokens cost 3–5× more than input tokens on the models in this calculator, and coding agents generate a lot of output. Models also differ widely in verbosity — according to xAI, Grok 4.5 emits roughly 14k output tokens on task batches where Claude Opus 4.8 emits about 67k — so both the per-token rate and the token count matter.