AI API Pricing Calculator 2026 — Compare GPT-5.6, Claude, Gemini & 50+ LLM Costs

Q: Why is output more expensive than input?

Generating tokens is sequential and compute-heavy while input is processed in parallel, so output is typically 3 to 8 times the input price. Capping response length is the fastest way to cut cost.

Q: What is prompt caching and how much does it save?

Resending the same large context lets providers cache it and bill those input tokens at about 10% of the normal rate, often cutting 40 to 70% off input cost for agents and RAG with a fixed prefix.

Q: Is DeepSeek really that much cheaper?

DeepSeek V4 Pro is about $0.44 input and $0.87 output per 1M tokens, roughly 10 times cheaper than a US frontier model, with a 1M-token context. Validate quality and data residency before moving production traffic.

Q: Should I self-host an open model or use an API?

Self-hosting only wins at high, steady GPU utilization because idle time destroys the economics. Most teams under about 500M tokens per month are cheaper on an API.

Q: How accurate and current are these prices?

Every frontier price was verified against the official provider pricing page on July 25, 2026. Prices change often, so confirm on the provider's site before large commitments.

Q: Is this calculator free?

Yes, all five tools and 50+ models are 100% free with no signup. An optional Pro plan adds price-drop alerts and CSV export.

Large-language-model APIs bill by the token — roughly 4 characters, or about ¾ of a word. You pay separately for input tokens (your prompt, system message, context and documents) and output tokens (the model's reply). Almost every provider charges 3–8× more for output than input, so long generations dominate your bill far more than long prompts.

The five levers that actually move your bill

Model choice. The spread is enormous: a frontier model like Claude Opus 4.8 ($5/$25 per 1M) can cost 100× more than a budget model like Amazon Nova Micro ($0.035/$0.14). Match the model to the task, not the hype.
Prompt caching. Reusing a large system prompt or context? Cached input tokens are billed at ~10% of the normal rate on OpenAI, Anthropic and Google. For agents with a fixed system prompt, this alone can cut 40–70% off input cost. The calculator's "cache hit rate" slider models it.
Batch API. Non-urgent workloads (evals, back-office processing, embeddings) run at roughly half price if you accept a 24-hour turnaround. Toggle "Batch API" to see it.
Reasoning / thinking tokens. Reasoning models emit hidden thinking tokens billed at the output rate. A "cheap" reasoning model can cost more than a pricier standard model once you count them.
Self-hosting. Renting a GPU only wins at high, steady utilization. The Self-Host vs API tab shows the break-even — most teams under ~500M tokens/month are cheaper on an API.

Input vs output: a worked example

Say you send 1,000 input tokens and get 500 output tokens per call, 100 calls/day, on GPT-5.6 ($5 in / $30 out per 1M). Input costs (1,000 ÷ 1,000,000 × $5) = $0.005; output costs (500 ÷ 1,000,000 × $30) = $0.015 — so 75% of the per-call cost is output, even though you sent twice as many input tokens. Shortening responses is usually the highest-leverage cost cut you can make.

How to cut your AI bill without hurting quality

Route easy requests to a cheap model and reserve the frontier model for hard ones; cap output length; turn on prompt caching for fixed context; batch anything that isn't real-time; and measure per-feature cost so you know where the money actually goes. Teams that do all five routinely cut 50–80%. If you'd rather have it done for you, First Deploy AI implements exactly these changes in production.

How much does the GPT-5.6 API cost? ▾

OpenAI's flagship GPT-5.6 (Sol) is $5 per 1M input tokens and $30 per 1M output tokens, with cached input at roughly $0.50. Lower tiers — GPT-5.6 Terra ($2.50/$15) and Luna ($1/$6) — trade a little quality for big savings. Use the calculator to see your exact monthly cost.

Why is output more expensive than input? ▾

Generating tokens is sequential and compute-heavy — the model runs a full forward pass per output token — whereas input is processed in parallel. That's why output is typically 3–8× the input price, and why capping response length is the fastest way to cut costs.

What is prompt caching and how much does it save? ▾

If you resend the same large context (a system prompt, a document, few-shot examples), providers can cache it and bill those input tokens at ~10% of the normal rate. For agents and RAG with a fixed prefix, that often cuts 40–70% off input cost. The calculator's cache-hit slider models the effect.

Is DeepSeek really that much cheaper? ▾

Yes — DeepSeek V4 Pro lands around $0.44 input / $0.87 output per 1M, roughly 10× cheaper than a US frontier model, with a 1M-token context. For many non-frontier tasks it's extremely competitive; validate quality and data-residency requirements before moving production traffic.

Should I self-host an open model or use an API? ▾

Self-hosting only wins at high, steady utilization. A GPU costs the same whether it's 20% or 90% busy, so idle time destroys the economics. The Self-Host vs API tab computes your break-even; most teams under ~500M tokens/month are cheaper (and far simpler) on an API.

How accurate and current are these prices? ▾

Every frontier price here was verified against the official provider pricing page on July 25, 2026. Prices change often — treat the numbers as a strong estimate and confirm on the provider's site before large commitments. Open-model hosted rates vary by host (Together, Groq, Fireworks) and are shown as typical values.

Is this calculator free? ▾

100% free, no signup, no paywall — all five tools and all 50+ models. An optional Pro plan adds monthly price-drop alerts and CSV export and helps support the tool, but you never need it to get a full answer.

Know your AI bill before you ship

💵 Estimate your monthly API cost

📊 Cost comparison

🔢 Token counter & instant price

📊 Full model pricing matrix

🖥️ Self-host vs API break-even

🔁 Flat subscription vs pay-as-you-go API

💡 Your calculator says you could save a lot/mo

Support chatbot

Code assistant

Document analysis

Content generation

Data processing

AI agent / tool use

RAG pipeline

High-volume classify

OpenAI

Anthropic Claude

Google Gemini

xAI Grok

DeepSeek

Mistral AI

Amazon Nova

Meta Llama (hosted)

Cohere

The five levers that actually move your bill

Input vs output: a worked example

How to cut your AI bill without hurting quality

The calculator is free. Forever.

Spending more on AI than you should?

🔌 One API. Every AI model. Coming soon.