# Why LLM Output Tokens Cost More Than Input My Weird Prompts — an AI-generated podcast. Daniel Rosehill sends in a prompt; AI hosts Corn and Herman discuss it. Verify important claims independently. - Episode: 6384 - Published: 2026-10-31 - Duration: 23:25 - Audio: https://episodes.myweirdprompts.com/audio/blended-token-pricing-workloads-49591.m4a - Tags: ai-inference, local-ai, gpus, transformer-architecture, reasoning-models - Web page: https://myweirdprompts.com/episode/blended-token-pricing-workloads-49591/ - JSON: https://myweirdprompts.com/api/episodes/blended-token-pricing-workloads-49591.json ## Summary Provider rate cards charge three to five times more for output tokens than input tokens, driven by the physics of memory bandwidth bottlenecks during decode. Comparing models using industry-standard three-to-one blends misrepresents actual costs, especially as agentic harnesses and retrieval-augmented generation skew workloads anywhere from twenty-two-to-one to over a hundred-to-one. Analysis explores roofline limits, hidden reasoning tokens, benchmark weighting, and why a single blended rate hides the primary cost drivers. ## The prompt > AI API pricing is almost always quoted as two numbers: a price per million input tokens and a price per million output tokens, and output is often several times dearer than input. That makes comparing models awkward, because which one is actually cheaper depends on how your own traffic splits between the two. I've wondered whether it makes sense to turn that split into a single "workload multiplier": if my workload is, say, 80% input and 20% output, I weight the two prices accordingly and get one blended number per model that I can compare honestly. Has anyone formalised this, beyond the fixed 3:1 blended prices some benchmark sites publish? What input-to-output ratios do real workloads tend to have, for example chat, RAG, coding agents, summarisation, classification, and long-form generation? And can the formula be extended to cover prompt caching, with cache hit rate, cached-read discounts and cache-write premiums, so that you end up with an effective cost per million tokens for your own usage pattern? Ideally this is something you could compute yourself from your own usage logs, rather than trusting a vendor's headline price to match your bill. ## Show notes Pricing for language model APIs presents a persistent asymmetry: output tokens cost three to five times more than input tokens. This disparity stems from hardware constraints rather than arbitrary vendor markups. Prefill batches entire prompts in parallel compute-bound sprints, whereas decode streams weights sequentially out of memory, yielding compute utilization as low as 1.4 percent on an H100. Standard industry benchmarks often publish blended rates using a fixed three-to-one input-to-output ratio, but actual workloads diverge wildly. RAG setups frequently reach ratios of 160 to one, while agentic benchmark runs hit nearly 100 to one. Because hidden reasoning tokens bill at output rates, reliance on generic blended formulas routinely obscures genuine costs. ## Sources - [Anthropic docs (cache write/read multipliers, TTLs, per-model table)](https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching) - [Multigrid, updated 2026-08-03 (prefill/decode economics, 1.4% utilisation)](https://multigrid.ai/learn/input-vs-output-tokens) - [Multigrid, updated 2026-08-04 (blended formula, caching extension, price-per-task)](https://multigrid.ai/learn/ai-price-history) - [MAPIKA.DEV, published 2026-06-09, reviewed 2026-09-05 (roofline model, ρ = K*/K)](https://mapika.dev/essays/the-price-is-the-physics) - [AI//COST glossary (agent-shaped blend, Artificial Analysis 7:2:1, cache-write treatment)](https://aicost.tools/glossary/blended-rate/) - [Promptster, 2026-08-19 (measured rate-card inversions, verbosity spread)](https://www.promptster.dev/blog/blended-token-cost-how-to-calculate) - [LiveBench cost data, 2026-10-25 snapshot (avg input/output tokens per model)](https://livebench.ai/data/20261025/cost.csv) - [OpenRouter listing, fetched 2026-10-31 (current list prices)](https://openrouter.ai/anthropic/claude-sonnet-5.5) - [Convly AI Price-Performance Index 2026 (114× spread, 3:1 blend)](https://convly.ai/ai-price-performance-index-2026/) - [DataLLM Lab, export 2026-06-12 (75/25 blend, 97 models)](https://www.datallmlab.com/blog/cheapest-llm-api-2026.html) - [CloudQuell, prices as of 2026-10-12 (3:1 blend rationale)](https://www.cloudquell.com/blog/cheapest-llm-api/) - [MyTokenTracker methodology (AI Cost Index, fixed-ratio basket)](https://mytokentracker.io/methodology) - [KickLLM, 2026-08-24 (2× price disagreement on GPT-5.6 Sol)](https://kickllm.com/research/llm-price-source-disagreement.html) - [HN, 2026-05-19 (4:1](https://news.ycombinator.com/item?id=48189375) — 5:1 real-world ratio critique) - [HN, 2026-07-24 (cache = 80](https://news.ycombinator.com/item?id=49040667) — 90% of agentic tokens; Grok 4.5 cache price cut) - [HN, 2026-10-19 (DS vs GLM output-heavy workload comparison)](https://news.ycombinator.com/item?id=50157697) ## Transcript Corn: The second number is always the bigger one. Every rate card, every vendor, input on the left, output on the right, and the right one is a multiple of the left. Herman: Three to five times, usually. Which is the whole reason Daniel's asking. Corn: Right, so here's his version. Comparing models on a two-number card is dishonest unless you know your own input and output split. His idea is to turn that split into a single workload multiplier. If your traffic is eighty percent input and twenty percent output, you weight the two prices that way and get one blended number per model. He wants to know whether anyone has formalised that beyond the fixed three-to-one blend that benchmark sites publish. He also wants to know what real workloads look like, from chat to RAG to coding agents. And can the formula absorb prompt caching, so you end up with an effective cost per million tokens computed from your own logs, not a vendor's headline? Herman: There's a lot of material on this, and most of it contradicts itself. Corn: Naturally. Start with why two prices exist, because I'd assumed margin. Herman: Most people do. It's not margin, it's physics. Serving a model has two phases. Prefill reads your entire prompt in one batched pass. All those tokens are processed in parallel, so the chip is compute-bound and busy. Decode then emits one token at a time, and each token is a full forward pass that has to stream every weight out of memory. Corn: So the first phase is a sprint and the second is a very long queue at one door. Herman: Here's the number that made it click for me. A seventy-billion-parameter dense model at BF16 on an H100 does prefill on a two-thousand-token prompt at roughly twenty-eight hundred prompt tokens a second. Decode on the same card tops out around ninety-six tokens a second per sequence. And compute utilisation during decode is about one point four percent. Corn: I've napped with a higher duty cycle. Herman: The chip is nearly idle in compute terms and completely saturated in memory terms. That's what you're paying for when you pay for output. Corn: Though a rate card is a menu, not a quote. Two models can rank in opposite orders depending on your traffic, and the card won't tell you which order is yours. Herman: Right. And that's why collapsing it into one number is tempting, and why several people have tried. Corn: Before that, I want the roofline. You've got the look of a man with a balance point. Herman: I do, and it's a good one. MAPIKA dot dev formalises it. Cost per token goes as two times the parameter count over how well you use the hardware. For a toy seventy-billion model on B200-class specs, that's a hundred and forty gigaflops per token, which is about sixty-two microseconds of compute. But reading a hundred and forty gigabytes of weights at eight terabytes a second takes seventeen and a half milliseconds. Corn: So the arithmetic takes sixty-two microseconds and the fetching takes seventeen thousand. Herman: Roughly two hundred and eighty times longer, and that's the balance point. You'd need a batch of about two hundred and eighty-one sequences decoding together before compute became the limit. Real decode batches sit far below that, so every token pays for a full sweep of the weights, shared among too few riders. Corn: So it's a bus that has to drive the whole route every time, whether it carries three passengers or three hundred. Herman: That's a fair picture. The fare per passenger drops as the bus fills, but the provider can't always fill it, because the people riding are waiting on their own answers and won't hold still for a bigger batch. Corn: Fine, but you said the cost ratio is a hardware ratio. Then why is every card exactly five times? Herman: Good catch, because MAPIKA says the same thing about itself. The line is that the roofline explains a cost pressure, not the whole card. Demand, margins, capacity, product strategy all move the posted price. The physics sets a floor and a direction. It doesn't set the number. Corn: And five times on every card is more marketing than physics. Herman: Look at the current listings. Sonnet five point five is two dollars in, ten out. Opus five point five is four and twenty. Haiku is ten cents and fifty cents. Fable five point one is ten and fifty. All five times, to the cent. MAPIKA sampled Anthropic, OpenAI and DeepSeek and found the ratio sitting between three and five, and Multigrid says usually three to five and occasionally far more. Corn: So the hardware says something like a bandwidth tax, and the price list says round number. Herman: Both can be true. Now, blending. Multigrid gives the cleanest statement. Blended price equals the input price times the input fraction, plus the output price times the output fraction, and the fractions sum to one. Take a model at two-fifty in and ten out. On a four-to-one workload, the input fraction is point eight, and it blends to four dollars a million. On a one-to-one mix, the same model blends to six twenty-five. Corn: Same model, same card, a fifty-six percent difference in what I'd pay. Herman: Their line is, the model did not change price, the workload did. And that's the whole legitimacy argument. If you're comparing two models, you hold the weights fixed and rank them. It's a weighted average. It's not clever, but it's honest. Corn: It's also something I'd have written on the back of a receipt. What I want to know is whether anyone's done it properly or whether everyone just picks three-to-one. Herman: Mostly the latter. Convly's price-performance index uses three times input plus output, divided by four, and reports a hundred and fourteen times spread across frontier APIs, eighteen cents to twenty dollars a million. DataLLM Lab ranks ninety-seven models by seventy-five percent input plus twenty-five percent output and calls it the ratio they see in typical production traffic. CloudQuell uses the same formula and calls it typical of RAG and chat, and then concedes that your real ratio changes the order. Corn: So the benchmark site publishes the ranking and then footnotes that the ranking is wrong for you. Herman: In small print, yes. MyTokenTracker is a bit different. It's an index over a fixed basket with a fixed ratio, on purpose, so the index measures price movement rather than model mix drift. That's a legitimate use of a fixed ratio. You want the thermometer to stay the same thermometer. Corn: And the parameterised ones? Herman: Multigrid's formula page, which is manual. You plug in your own fractions. And AI slash COST publishes an agent-shaped blend, mostly input and most of that cached, and notes that Artificial Analysis uses seven parts cache hits to two parts uncached input to one part output. Their caveat is the one to carry. A blend is good for ranking models on one kind of traffic and not good for predicting your bill if your mix differs. Corn: Multigrid has a warning too, didn't you say? Herman: That a single blended number hides the number you're actually optimising. The blend is a ranking device. The second you use it as a forecast, you've thrown away the one variable you control, which is the split. Corn: Hm. So my instinct is, don't optimise the blend, optimise the thing inside it. Cut output tokens first. Herman: Multigrid says that almost verbatim. A thirty percent cut in output length is worth several times a thirty percent cut in input length. Corn, you're the one who defends the slow option. Would you take a model that thinks longer and costs more per token if it got the answer right? Corn: Anything worth computing is worth waiting for. But I'd want it priced per answer, not per token, because a model that thinks for ten minutes is just billing me for its nap. Herman: You'd know. Corn: I bill for mine at a flat rate. Herman: And that per-answer point is where part of this goes next. Multigrid has an example where an extraction job at ten dollars in and thirty out, four thousand tokens in and four hundred out, cost five point two cents. A reasoning model at one dollar in and four out, emitting three thousand hidden reasoning tokens, cost one point seven six cents. Per-token prices fell ten times and seven and a half times, and the bill fell three times. Corn: That's a strange result. Prices fell by close to an order of magnitude and the saving is a third of that. Herman: Because the hidden tokens ate most of the discount. Twelve thousand dollars' worth of per-token cleverness, if you like, spent on thinking nobody sees. Corn: Hidden reasoning tokens bill as output, so the output share isn't whatever the fixed mix says it is. Herman: And that's something no fixed blend sees. Corn: And here's a voicemail from Opal in Tel Aviv. Opal: Hi Corn, hi Herman, this is Opal from Tel Aviv, I am actually calling from the queue at the pharmacy on Ibn Gabirol, they are very slow here so I have time to think about your show. So today is about the tokens, the input ones and the output ones, and I wanted to ask, when you do the blending thing with the percentages, does it matter which way round you write the ratio, because I always get confused if it is input to output or output to input, and my cousin who works in insurance says ratios are always written the wrong way round on purpose. Also I once spent a whole weekend in two thousand and one trying to work out the average price of a falafel across the whole city, weighting by how much I liked each place, and I never finished it, so I am very glad someone is doing this properly. Herman: Write input to output and label it, Opal, so four to one means four input for every one output, and your falafel average was a weighted blend of exactly this shape, just with tastier weights. Corn: Now, the part Daniel wanted most. What does a real ratio look like? Because I'd bet chat is somewhere near three to one. Herman: That's probably where the folklore came from, and I'll say upfront that nobody has published a measured per-workload table. What we have are fragments, and they pull in different directions. Three to one is asserted as typical of RAG and chat by one site and as typical production traffic by another. Someone on a forum cites four or five to one for general real-world mixes. Agentic harnesses run at twenty-two to ninety-eight to one. Corn: Ninety-eight. Herman: LiveBench publishes average input and output tokens per model in its cost snapshot. Claude Fable five point one is about five million input against fifty-two thousand output, ninety-eight to one. Sonnet five point five is eleven million against a hundred and fifty thousand, seventy-five to one. GPT-6 Luna is fifty-two to one, Gemma four thirty-one-B is thirty-two, and Qwen three point eight Max is twenty-two. The least input-heavy model in the table is still about a hundred and twenty-one to one. Corn: Hold on, that last one is least input-heavy at a hundred and twenty-one and Fable is ninety-eight? Those can't both be right. Herman: No, you're right, I garbled that. The model with the highest ratio in the table is the one at a hundred and twenty-one, so that's the most input-heavy, not the least. Scratch my phrasing. The point stands, the whole table sits between twenty-two and the low hundreds. Corn: Fine. And these are benchmark harnesses, so agents hammering a task with tool output and files. Herman: Which is the point. They aren't chat, but they show how far real traffic can sit from three to one. Promptster's line is the one I like. The standard three-to-one blend describes an average workload, which is to say nobody's. Corn: So let's reason it through, with the caveat that we're guessing from fragments. RAG? Herman: Very input-heavy. Promptster's example is a pipeline stuffing eight thousand tokens of context to get a fifty-token answer. That's a hundred and sixty to one. The output rate barely matters to that bill. You could halve it and not notice. Corn: Chat is the closest to balanced, though even there the history gets resent every turn, so input piles up. Herman: Yes, my guess is chat is higher than three to one once conversations run long, but I wouldn't swear to it. Think of a twenty-turn conversation. Turn twenty resends all nineteen earlier exchanges, so the input on that single turn dwarfs the short reply coming back. Long-form generation is the reverse of RAG. A short brief and a long essay, so output dominates and the input rate is a rounding error. Summarisation is input-heavy, a long document in and a short summary out. Classification, I'd say, is near balanced or even output-light, a label out for a paragraph in, although it depends how much you stuff in the prompt as instructions. Corn: And coding agents are the extreme end. Herman: Cache reads make up eighty to ninety percent of tokens in a real agentic task, per one practitioner. An agent re-reads the same files and the same conversation every step, so almost everything is repeated input. Corn: That has a consequence for three to one, doesn't it? If agents sit at twenty-two to ninety-eight, three to one is over-weighting output. Herman: Systematically, and in a specific direction. It flatters models with cheap output and penalises models with cheap input. That's a bias baked into published rankings, not a neutral average. Corn: So the rankings aren't neutral, they have an accent. Herman: And caching makes it worse, because caching breaks the two-number model harder than output pricing does. Multigrid extends the formula. Effective input equals the input price times the uncached fraction, plus the input price times the cache discount times the cached fraction. At two-fifty a million, eighty percent cached, with cache reads at ten percent of full price, you get two-fifty times point two plus two-fifty times point one times point eight. Seventy cents a million. Corn: Seventy-two percent off, from a feature nobody puts on a price chart. Herman: That's their phrasing, nearly. Now the write side, which most blends skip. A cache miss is priced at the cache-write rate when the vendor bills writes, because a miss is the request that writes the prefix again. So low-hit-rate traffic can cost more than no caching at all. The write premium is a real cost, not a free option. Corn: So the discount has a deposit. Herman: Look at Anthropic's card for Sonnet five point five. Two dollars in, ten out. Five-minute cache writes are two-fifty, one point two five times input. One-hour writes are four dollars, twice input. Hits and refreshes are ten cents, point oh five times input. Opus five point five is four and twenty, with writes at five and eight and hits at twenty cents. Fable is ten and fifty, with hits at twenty-five cents, which is two and a half percent of input. Corn: Four meters on a two-number card. Herman: Default time-to-live is five minutes, refreshed free on each use. And the clock counts streaming time. If a response takes four minutes to stream, the follow-up that reuses the prefix has to start within about a minute. Corn: So slow answers cost you your cache. I find that unfair, personally. Herman: Corn, would you pay double for the hour-long version, if your usage were bursty? I honestly can't tell which way you'd go. Corn: Bursty usage with a premium for patience? I'd take the five-minute one and gamble that I'm back in time. Which is how everything I own got cold. Herman: Then you'd take the loss. At low hit rates, the two-times write is the one that stings. Corn: So what's the recipe? Daniel wants to compute it from his own logs. Herman: It's short. Pull the four usage fields per request: input tokens, cache read tokens, cache creation tokens, output tokens. Multiply each by its own rate, sum, and divide by total tokens. That's your effective cost per million, and it's measured, not quoted. Then redo it monthly, because the mix drifts every time you add an agent. Corn: And this is why the headline price won't match the bill. Herman: Promptster ran eleven model configurations on identical tasks. The rate card predicted the wrong winner five times out of five. Kimi K3 charges fifteen dollars a million output, half of GPT-5.6 Sol's thirty, and cost fifty-four percent more because it emitted three point one times the tokens. Gemini three point six Flash charges more on both rates than GPT-5.6 Luna, and cost thirty-two percent less because Luna used one point nine times the tokens. Corn: So the cheap model talked more. Herman: Output counts on the same three questions ranged from two hundred and eighty-six to seventeen hundred and fifty-one. That's a sixfold spread, set against a three-to-five-times price ratio. Verbosity is a bigger lever than price. And repeat runs of Claude Opus five on the same battery varied twenty-six percent, which sets a floor. Any comparison that hinges on a gap under about twenty-five percent is unresolved. Corn: So if I rerun the same prompt tomorrow, I might get a bill a quarter different, and that's the same model on the same day's price. Herman: Which is why one run proves nothing. You want a battery, repeated, before you believe a ranking. Corn: That reminds me of the old fight over transcription scores. A headline error rate told you very little about whether the transcript was any use. Herman: Same gap, between a headline metric and the thing you care about. So I'd say the blend is a ranking tool and the effective cost is a log-derived measurement. They aren't the same object. AI slash COST says it plainly. If your cache hit rate is low, your answers long, or your prompts cross a long-context threshold, your effective rate will differ from any published blend, sometimes by a multiple. Hilbert: Typical. I'd argue with the word. Nobody I know has a typical anything until they've counted. Herman: Fair, Hilbert, though that's the three-to-one claim exactly. It's repeated until it sounds like a measurement. Hilbert: Mr. Barak downstairs was quoted a copier lease this week. Per-page rate for black and white, a much higher one for colour. The salesman assumed three to one. I took Mr. Barak's receipts home, thirteen months of them, and counted. Eleven to one. He's been overpaying since the last lease. The counting is in a ring binder, section four, and I turn the pages with one claw because the rings weren't made for the rest of me. Corn: Eleven to one, and the salesman's three was the number on his own sheet, so of course he picked it. Herman: It's the same shape of error as a bad cache assumption. A wrong ratio on a two-rate card inflates the bill for years and nobody checks. Hilbert: The copier went back to the lessor in March. I don't use the lease, I use the binder. I won't count my own model usage, mind. The logs are the vendor's logs, and I don't trust a man to count his own till. Herman: Hilbert, you can export the usage fields per request from the API side, though. You'd be multiplying them yourself. Hilbert: Herman said earlier the mix drifts monthly. I'm going to recount Mr. Barak's in December. Dalia loads his paper tray every morning at ten past eight, and she says the colour toner goes twice as fast as the black. Eleven to one is a hundred and forty-four pages of colour a month. Corn: Which means the salesman's assumption is probably wrong in the other direction on the toner. Herman: Well, the pages-per-cartridge differ, so possibly. The useful thing from the binder is that the number came from receipts, not from a sheet. Corn: There's a version of Daniel's question where that's the entire answer. Count first, then argue. Herman: And that matters for the closing, because there's another place the number gets lost. KickLLM found two independent public price sources disagreeing on the input price for the same first-party model, with the widest gap at two times on GPT-5.6 Sol. So even the card isn't always the card. Corn: The one number I thought I had, the input price, has an argument going on inside it. Herman: In at least one case, yes. Corn: Here's what I'm left with. If three to one over-weights output, how much of the published ranking is just a ratio nobody measured? The hundred and fourteen times spread is real, but the order inside it may not be. Herman: Caching makes effective cost a function of your architecture, not your model choice. The unit may be shifting from price per million tokens to price per task, and nobody has a standard for that yet. The model did not change price, the workload did, and the rate card was never what you were buying. Corn: Thanks to our producer, Hilbert Flumingtop, for keeping the desk running. Herman: If you want more of this, try episode nine, Benchmarking Custom ASR Tools - Beyond The WER; and episode five, Fine-Tuning ASR For Maximal Usability. This has been My Weird Prompts. If you've enjoyed it, a review helps more people find the show. Corn: See you tomorrow.