Why is LLM API pricing not a simple multiplication?

Because input and output tokens are priced differently, and your workload profile decides the ratio: summarization is ~8:1 input-to-output, code generation is 1:2, chat is 1:0.4. The same 10K daily requests can cost 100× more on one provider than another depending on the profile. This calculator compares 17 models across 6 providers with caching and batch discounts applied.

What is prompt caching and how much does it save?

Prompt caching reuses the token computation for repeated prompt prefixes at a heavily discounted rate — often a fraction of the input price. For workloads with stable system prompts (chat assistants, RAG), caching is the most underutilized discount in the LLM bill. The calculator models cache hit rates so the savings show up in the comparison.

Why do tokenizers change the effective price?

Tokenizers split text differently per model — 1,000 characters is not 1,000 tokens, and the same English text can tokenize to different counts across providers. Combined with input/output ratio differences, the "same" prompt costs different amounts. The calculator accounts for tokenizer variance so your comparison is apples-to-apples.

Tool Hub

📰📡

Home · JSON · Calculators · Articles · About

LLM API Pricing Calculator

Compare real token costs across 17 models from 6 providers. Workload-aware — chat costs different from summarization. Includes prompt caching (up to 90% off) and batch processing (up to 50% off) discounts. All calculations client-side. No data leaves your browser.

📋 Workload Configuration

Select a workload profile to auto-fill typical token ratios, or configure custom values. Token counts are approximate — different providers use different tokenizers, so the same text produces different token counts. The calculator uses your numbers uniformly; see the methodology note below.

Workload Profile

Daily Requests

Avg Input Tokens / Request

Avg Output Tokens / Request

Prompt Cache Hit Rate: 0%

% of input tokens billed at the cache-read rate instead of the full input rate. Cache reads run 2–20% of input depending on provider.

Batch Processing: 50% off for async on OpenAI, Anthropic, Google and Mistral. DeepSeek and Together AI publish no batch tier.

Why LLM API Pricing Isn't a Simple Multiplication

The pricing page says $2.00 per million input tokens. Your mental math: "10,000 requests × 2,000 tokens × $2.00/M = $40/day." Then the bill arrives and it's $120/day. Four structural reasons the napkin math fails every time.

1. Input ≠ Output Token Counts

Most pricing pages quote input and output prices separately — but your workload determines the ratio. A chat application with system prompts and conversation history averages roughly 1:0.4 input-to-output tokens (2.5× more input than output). A summarization pipeline averages 8:1 (long document in, short summary out). A code generation agent averages 1:2 (short prompt, longer code block out). Since output tokens are priced 5× higher than input tokens across the current flagship tier — 6× on Gemini 3.1 Pro, 3–4× at DeepSeek and Mistral — the workload ratio is a cost multiplier. A summarization workload at 8:1 pays mostly input prices; a code generation workload at 1:2 pays mostly output prices. Same daily request count, same provider — 3× cost difference between profiles.

2. Prompt Caching: The Most Underutilized Discount

A cache read is the cheapest token on any price card. OpenAI and Anthropic both bill one at 10% of the input rate — a 90% discount — and Claude Opus 5.5 goes further, at 5%. Google bills Gemini cache reads at 10%, Mistral at 10% and DeepSeek at 2–3%. If your application sends the same system prompt with every request — and most do — that's 500–2,000 tokens per request that qualify for caching. A 2,000-token system prompt at 1M requests/month saves $7,600/month on Claude Opus 5.5 alone. The discount exists. Most teams never configure it.

3. Batch Processing: Half Price for Patience

OpenAI, Anthropic, Google and Mistral all publish a 50% batch discount for asynchronous processing, traded for a completion window instead of a realtime SLA. For offline workloads — evaluation runs, dataset labeling, embedding generation, nightly summarization jobs — batch pricing cuts the bill in half with no downsides. Stacked with caching, a batched cache hit bills at 5% of the sticker input rate — 50% batch × 90% cache. DeepSeek and Together AI publish no batch tier; there the lever is the cache hit, worth 80–98% off input.

4. Tokenizer Variance: 1,000 Tokens ≠ 1,000 Tokens

OpenAI, Anthropic, Google, DeepSeek and Mistral each ship a different tokenizer, and none of them quotes a token count for your text. The same English paragraph can measure 800 tokens on one provider's tokenizer, 750 on another's and 880 on a third; on Chinese text with interleaved English technical terms the spread is wider still. Across 1 million requests per day, a 10% tokenizer efficiency difference translates to thousands of dollars per month in cost variance — before you even compare the per-token prices. The calculator uses uniform token counts for comparability; in practice, benchmark your actual text across providers before committing.

How to Use This Calculator

Pick your workload profile — this sets realistic input/output token ratios.
Set your daily request volume and adjust token counts if needed.
Slide the prompt cache hit rate based on what percentage of your requests share a common prefix (system prompts, few-shot examples, RAG context).
Toggle batch processing for offline/non-realtime workloads.
Compare the "Annual Cost" column (no optimizations) with "Optimized Annual" (cache + batch) — the gap is savings you're leaving on the table.

Pricing Basis, Sources & Assumptions

Every rate on this page is a published vendor list price — no negotiated discounts, private pricing, or credit offsets. Totals are in USD and exclude tax. Rates were last checked against the sources below on 2026-09-23.

Pricing input	Basis used on this page
Region	Global USD list rates. Five of the six vendors publish one worldwide rate card, so the page carries no region selector. The exceptions are stated rather than modelled: Google prices a non-global endpoint about 10% above the global one, Mistral charges 10% for regional inference, and Azure OpenAI is regional by deployment type.
Currency	USD — on-demand list price, tax excluded
Last checked	2026-09-23 — The roster was rebuilt on this date against the vendor price pages linked below. Every model listed is drawn from the vendor's current published line-up, and every rate is the listed per-million-token price for the model named rather than a retired generation's. Where a vendor publishes context-length tiers, peak and off-peak hours, or an introductory price, the assumption list below says which one this page uses. Treat the linked price pages as authoritative.

Modelling assumptions

All rates are USD per 1,000,000 tokens, on-demand list price, tax excluded. Input and output tokens are billed at different rates and are never interchangeable.
Cached input is billed at the discount the vendor publishes for a cache read, and it only applies when the same prefix is reused within the cache window. The saving is real but requires prompt structure designed for it.
Batch processing is modelled at the vendor's batch discount where one is published. Batch trades latency for price and has a completion window rather than a latency SLA.
Context-length tiers are not modelled. Gemini 3.1 Pro doubles both its input and its output rate above 200K input tokens, and the GPT-6 and GPT-5.6 families publish a higher long-context rate. Every model here is billed at the shorter-context rate, so a request that crosses the threshold costs more than shown.
DeepSeek is billed at peak rates. DeepSeek publishes peak and off-peak rates with off-peak at half price. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday; every other hour, including weekends and Chinese public holidays, is off-peak. A workload that runs out of hours pays roughly half of what this page shows.
Gemini 3.8 Flash is billed at its standard rate, not its introductory one. Google's introductory price for the 3.6, 3.7 and 3.8 Flash models is $0.75 / $3.75 per million tokens through 31 December 2026, reverting to the $1.50 / $7.50 used here on 1 January 2027. The page uses the standard rate so the comparison does not expire mid-life — today's effective rate on that row is half what is shown.
Mistral prices Large 3 below Medium 3.5 on its own card ($0.50 / $1.50 against $1.50 / $7.50). That is the vendor's ordering, carried across unchanged rather than corrected.
The open-weight row is a hosting price, not a model price. Kimi K3, GLM-5.3 and MiniMax M3 are billed by Together AI, a third-party host. The same weights are priced differently elsewhere, and serving them yourself costs GPU-hours instead of tokens.
Reasoning and thinking tokens are billed as output tokens on every model that generates them, so a reasoning-heavy request costs more than the headline output rate implies.
Workload shape matters more than the rate: summarisation and chat have very different input-to-output ratios, and the profile selector is what makes the comparison meaningful.

What this model excludes

Enterprise agreements, committed-spend discounts, and volume tiers, which can move the effective rate well below list.
Regional and sovereign deployments priced above the global rate card.
Fine-tuning, hosted fine-tuned model hours, and provisioned throughput — priced on the LLM fine-tuning page instead.
Embeddings, image, audio, and video modalities, and any per-file or per-minute surcharges.
Retries, failed requests, evaluation harnesses, and observability of the calls themselves.

Sources

OpenAI (2026). "OpenAI API Pricing." Per-million-token input, output, and cached-input rates for the OpenAI models listed. developers.openai.com
Anthropic (2026). "Claude API Pricing." Per-million-token rates and prompt-caching discounts for the Claude models listed. platform.claude.com
Google Cloud (2026). "Generative AI on Vertex AI Pricing." Gemini per-million-token rates on both the standard and batch tiers. cloud.google.com
DeepSeek (2026). "DeepSeek API Pricing." Per-million-token rates for the DeepSeek models, including off-peak discounts. api-docs.deepseek.com
Mistral AI (2026). "API Pricing." Per-million-token rates for Mistral Large 3, Medium 3.5 and Small 4, plus the batch, caching and regional-inference terms. mistral.ai
Together AI (2026). "Together AI Pricing." Third-party hosting rates for the open-weight models listed: Kimi K3, GLM-5.3 and MiniMax M3. together.ai

Vendor list prices change without notice — re-check the linked pages before committing spend. jslet takes no vendor sponsorship and carries no affiliate links; see about.