nextpixel

LLM API cost calculator

Work out what an LLM feature actually costs per month across the major models, including prompt caching and batch discounts. Prices are editable, because they change constantly.

Runs entirely in your browser — nothing is sent to us

Your workload

Cached input is charged at 10% of the input rate.

Batch requests receive a 50% discount on both input and output.

Both are applied across all models here. Availability and exact terms vary by provider.

Monthly cost by model

Indicative list prices, August 2026. Editable — verify with the provider.

ModelIn$/1MOut$/1MPer monthPer year
Claude Haiku 4.5Anthropic
Claude Sonnet 5Anthropic
Claude Opus 5Anthropic
GPT-5.6 LunaOpenAI
GPT-5.6 TerraOpenAI
GPT-5.6 SolOpenAI
Gemini 3 FlashGoogle
Gemini 3.1 ProGoogle
Gemini 2.5 FlashGoogle

Cheapest option is highlighted. Cost per request is shown on hover in the monthly column title attribute.

Reading these numbers well

The cheapest model is often the right one. Most production workloads are classification, extraction or routing, where a small model performs indistinguishably from a frontier one at a fraction of the price. Route the hard cases up rather than sending everything to the largest model.

Input tokens are usually the bill. Long system prompts and retrieved context are sent on every single call. Prompt caching typically removes a large share of that, which is why the cache field above tends to move the number more than switching models does.

Conversations grow quadratically. If you resend the full history each turn, a ten-turn conversation costs far more than ten times a single turn. Summarising or truncating history is often the single biggest saving available.

This excludes everything except tokens. No infrastructure, vector storage, egress, or the engineering time to build and operate it. Inference is rarely the largest line in a real AI budget.

Costs looking uncomfortable?

Most production LLM bills can be cut substantially without changing models — caching, batching, routing simple requests to smaller models, and not sending the whole conversation every turn.

Other tools