Take an ordinary workload — 2,500 input tokens, 600 output tokens, 50,000 requests a month — and price it on OpenAI GPT-5.6 Sol at $5.00/$30.00 per million tokens. Input costs 2,500 ÷ 1,000,000 × $5.00 = $0.0125. Output costs 600 ÷ 1,000,000 × $30.00 = $0.018. That is $0.0305 a request, or $1,525 a month. Run the identical workload on Mistral's Ministral 3 at $0.10/$0.10 and you pay 2,500 ÷ 1,000,000 × $0.10 = $0.00025 plus 600 ÷ 1,000,000 × $0.10 = $0.00006, which is $0.00031 a request — $15.50 a month. Ninety-eight times cheaper, same traffic, no negotiation.
That gap is the single most useful thing this calculator surfaces, and the reason the table is sorted by monthly cost rather than by sticker price.
Two prices, and the one everybody misreads
Every LLM price card carries two numbers because the model does two different jobs. Reading your prompt is a parallel operation: the whole input is processed in one pass across the accelerator. Writing the answer is sequential — each token is generated from the full state built so far, one step at a time, and that step cannot be shared with anyone else's request in the same way. Vendors price that asymmetry directly. Claude Opus 5 charges $5.00 in and $25.00 out. GPT-5.6 Sol charges $5.00 in and $30.00 out. Gemini 2.5 Flash charges $0.30 in and $2.50 out. The multiple sits between 5× and 8× almost everywhere.
The practical consequence is that model choice punishes writing far harder than reading. Flip the example above to an output-heavy shape — 500 input, 3,000 output, same 50,000 requests. GPT-5.6 Sol now costs 500 ÷ 1,000,000 × $5.00 + 3,000 ÷ 1,000,000 × $30.00 = $0.0925 a request, or $4,625 a month. It tripled while the total token count barely moved. Meanwhile Llama 3.3 70B via Together, at a flat $1.04/$1.04, costs $0.00364 a request — $182 a month. Grok 4.6 at $2.00/$6.00 costs $950 for the same shape despite an input price only twice Llama's, because its output price is nearly six times higher. Enter your real input:output ratio before you read anything else in the table; the ranking changes underneath it.
Why the tiers are priced so far apart
The calculator groups models as frontier, mid, small and budget, and the "Show" selector lets you price one band at a time. The spread across bands is not a pricing accident — it tracks the cost of serving. A frontier model has more active parameters per token, needs more accelerator memory, and occupies more of a node for longer. Claude Opus 5 at $5.00/$25.00 and Claude Haiku 4.5 at $1.00/$5.00 come from the same vendor with the same billing model, and the ratio between them is a clean 5×. Google's ladder is steeper still: Gemini 3.1 Pro at $2.00/$12.00 down to Gemini 2.5 Flash-Lite at $0.10/$0.40 is 20× on input and 30× on output.
Then there are the models that break the ladder. DeepSeek V4 Flash at $0.14/$0.28 and Ministral 3 at $0.10/$0.10 are priced like infrastructure rather than like a product tier. That is what stretches the spread on this page from the roughly 5× you would expect within one vendor to the 98× you actually see. If you have never priced your workload against the budget band, do it once — the number is usually startling, and choosing a model tier is where you decide whether the quality gap is worth it.
The cached-input share field, and when to be honest about it
Set "Cached input share" to 40 and the calculator computes an effective input rate of 0.6 × standard + 0.4 × cached, then prices the workload at that blended figure. On Claude Sonnet 5 the cached rate is $0.20 against $2.00 standard, a 90% discount on the repeated portion. At a 60% cached share the effective input price becomes 0.4 × $2.00 + 0.6 × $0.20 = $0.92 per million. Our original 2,500/600 workload then costs 2,500 ÷ 1,000,000 × $0.92 + 600 ÷ 1,000,000 × $10.00 = $0.0083 a request — $415 a month against $550 uncached.
You are entitled to a number above zero only when two things are true. First, the repeated part of your prompt is byte-identical between requests and sits at the very front — same system prompt, same tool definitions, same few-shot block, in the same order, with nothing variable spliced in above it. Second, requests arrive densely enough that the next one lands inside the cache lifetime, which is measured in minutes. A support tool handling 40 requests a minute against a fixed 12,000-token preamble will genuinely see 85%+. A nightly job firing 200 requests spread across eight hours will see close to zero. Guess, and every downstream figure on this page inherits the guess. The prompt caching savings calculator derives the hit rate you actually need before caching pays for itself, including the write premium this field ignores.
Note also that "cached input share" is not the same as the batch checkbox. Caching discounts repetition; batching discounts patience. They stack, and both are applied here where a vendor publishes them — every OpenAI, Anthropic and Google model in the index carries a 50% batch discount, while DeepSeek, Mistral, xAI and the Together-hosted models do not.
Reading the "vs cheapest" column
The multiplier in the last column is the ratio of that row's monthly total to the cheapest row's monthly total, computed on your traffic shape. It is not a ratio of price-card numbers, and the difference matters. GPT-5.6 Sol's input price is 50× Ministral 3's and its output price is 300×; the blended multiplier on the 2,500/600 shape came out at 98×. Change the shape and the multiplier moves without a single price changing.
Use it as a triage instrument. A row at 1.4× is inside the noise of a bad month's traffic forecast and should be decided on quality alone. A row at 12× is a budget decision that needs a quality argument to survive. A row at 98× is asking whether you have ever actually tested the cheap option on your own evaluation set — which, for classification, extraction, routing and short summarisation, frequently passes. When it doesn't, the fix is usually to split the traffic rather than pick one model, and the model switch savings calculator prices that split.
One caveat the table cannot show: prices in the index are per-token list rates verified from vendor pricing pages on 15 August 2026. They exclude rate-limit tiers, committed-use discounts, and the reasoning tokens some models bill as output whether or not you see them. If a model's real bill runs 30% above what this page predicts, hidden reasoning output is the first thing to check.
Prices used by this calculator were last verified on 15 August 2026 from the vendors' own pricing pages. See the full price index for every figure and its source, the change log for what has moved recently, and the methodology for how the index is maintained and where its limits are.