What Every AI Model Costs Per Million Tokens

This is a maintained AI model API price index: what every current model costs per million tokens on input, cached input and output, plus batch, flex, off-peak, priority and fast-mode rates wherever the vendor publishes them, organised by the surface you buy on. Data as of August 30, 2026. Every figure traces to the vendor’s own pricing page. Promotional rates never share a column with list rates.

The reason to build it this way is not tidiness. In the last three weeks two separate frontier rates moved for reasons that were printed in a footnote and dropped by nearly everyone who quoted the number — including, in one case, us. A reader looking at two confident secondary sources had no way to tell which one to believe, because neither carried the footnote that settled it.

So the index leads with those two cases, then gives you the tables: the full rate card, the promotions with their end dates, the surface layer where the same model bills differently depending on where you bought it, the batch slugs that cost more than standard, the long-context cliffs that make any one-number-per-model table incomplete, and the one vendor charging in data rather than money.

A price index earns the right to be cited when it can settle a disagreement that the reader cannot settle themselves. Two cases in August 2026 show exactly what that looks like, and they run in opposite directions from the same underlying mistake.

The first is Anthropic’s Claude Sonnet 5. When Sonnet 5 launched, $2 per million input tokens and $10 per million output tokens were published as introductory pricing through August 31, 2026, with a scheduled increase to $3 / $15 on September 1. Two secondary sources contradicted each other this month: one reported the increase had been cancelled, one reported it was landing. Only the vendor could settle it, and the vendor did. Anthropic’s pricing documentation now carries a callout stating that the $2 / $10 pricing “is now the standard price” and that “the previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur.” The Sonnet 5 launch post carries a dated edit note to the same effect, marked August 10, 2026.

Our own August 5 pricing tracker reported the expiry footnote accurately, quoting Anthropic’s own wording, and built a section on the September 1 step-up. Five days later the vendor withdrew the increase. The post was not careless; it was a correct snapshot of a rate whose footnote the vendor then deleted. That is the strongest argument for a maintained index we can make, and it is about our own work, which is why it belongs at the top of this page rather than buried in a correction note.

The second is OpenAI’s GPT-5.6 Sol, and it is the same error running backwards. Sol’s current $4 / $20 was widely reported — and recorded in our own internal anchors — as a list price. It is not. OpenAI’s model page states plainly that “GPT-5.6 Sol’s promotional pricing is available at least through November 21, 2026.” A promotional decrease was mistaken for a permanent one, where Sonnet 5 was a scheduled increase mistaken for an inevitable one. Both mistakes come from taking a rate without its footnote.

The two vendors also differ sharply in what they disclose about the other side of the promotion. Google publishes both halves for its Gemini 3.7 and 3.6 Flash promotion — the promotional rate, the end date, and the exact figure that applies afterwards. OpenAI publishes a soft floor and no post-promotional rate at all. What it does publish is a characterisation: the cut is “a 20% reduction in input pricing and a 33% reduction in output pricing.” Run that backwards and you land on $5 / $30, which is GPT-5.5’s rate. This index will report that only as vendor arithmetic, and never in a price cell, because OpenAI has not published it as a price.

The habit of reading a rate together with its footnote is the same literacy our model catalogue literacy guide argues for across prices, dates and surfaces. This page is that argument turned into a dataset.

The gaps do as much work here as the prices. A price index that silently drops the rows it could not verify looks more complete than it is, and quietly teaches the reader that everything absent from it does not exist. Every unverified row below is kept in and labelled, and the three labels mean three different things.

Every rate column in this index resolves to one of the pages below. They are listed rather than footnoted so a reader can re-derive any cell, and so a later refresh has an explicit list to re-fetch.

Read the columns carefully. List holds the standard rate. Promotional holds a discounted rate together with its published end date, or the vendor’s own words where no date exists. They are never merged, and a blank promotional cell means the vendor publishes no promotion on that row — not that we did not look.

Discount tiers collects batch, flex and off-peak. Premium tiers collects fast mode, priority, peak rates and the long-context bands. Both are per-vendor vocabularies rather than a shared one: “batch” means a 50% discount at Anthropic, OpenAI, Google, Alibaba, Mistral and AWS; 60% of standard at Moonshot; a 20% discount at xAI on one model and nothing at all on its flagship; and, on three OpenRouter slugs, a price higher than standard. Section 06 takes that apart.

Two structural observations the table makes that a per-vendor page cannot. First, newer is not reliably cheaper or dearer within a family. Gemini 3.7 and 3.6 Flash sit at a promotional $0.75 / $3.75 while the older 3.5 Flash is $1.50 / $9.00 — double the input and 2.4× the output. Claude Sonnet 5 undercuts both Sonnet 4.6 and 4.5 by a third on input. Grok 4.6 costs more than Grok 4.3 and loses batch support. Whatever rule you were using to guess a price from a version number, it does not hold.

Second, the discount tiers are not a shared vocabulary. Batch is 50% off at six vendors, 60% of standard at Moonshot, 20% off at xAI on exactly one model, and does not exist at all at DeepSeek, Z.ai, MiniMax, Meta or Tencent. Google runs four tiers where flex and batch are both exactly 50% off but are different products with different latency guarantees, and priority is a premium tier at 1.8× on every published row. Reading “batch” as a fixed discount across vendors will misprice a budget by a factor of two.

This is the table that makes the index falsifiable. Every promotional rate in the archive above appears here with whatever the vendor published about when it stops — including, in five rows, the fact that the vendor published no end date at all. A price index that folded these into a single rate column would be unfalsifiable for most of the vendors in it, because there would be no way to tell which cells were about to move.

The most useful thing this index does is refuse to give one number per model. A price is a rate, a unit, an endpoint and a tier — and three of those four change when you move surfaces without the rate moving at all.

The cleanest illustration is first-party and unambiguous. Anthropic bills through AWS Marketplace and Microsoft Foundry in Claude Consumption Units. Token usage is rated in USD at exactly the standard per-model rates in the index above, any negotiated discount is applied, and the result is converted at $0.01 per CCU — 100 CCU represents $1.00 — and reported hourly. Your AWS or Azure bill shows a single CCU line item. The rate is identical. The unit on the invoice is not. Anyone reconciling a cloud bill against a per-million-token model will find nothing that looks like a token.

The second illustration is a capability rather than a unit. Anthropic’s fast mode — $10 / $50 on Opus 5 and Opus 4.8, exactly 2× standard — is available on the Claude API first-party only. It is not available on Claude Platform on AWS or on partner-operated clouds. Same model, same headline rate, and one surface can buy a latency upgrade the others cannot.

The regional dimension deserves its own line, because of all the surface premiums it is the one presented with no price vocabulary attached at all. AWS Bedrock’s regional price list is a premium AWS never labels as one. The only word on those tables is a region name. Measured against the US East and US West rows on AWS’s own table, Gemma 3 27B input runs $0.23 in US regions, $0.27 in Mumbai, Ireland and Milan — a 17.4% premium — and $0.36 in London, which is 56.5% above the US row. Six model families on that list land between +50% and +57% in London. Sydney is the opposite case at around +3%.

OpenRouter is worth stating precisely because the common assumption about it is wrong in a specific way. It takes no per-token markup — that is the vendor’s own claim, made twice on its own pages, and the rate rows in the index above bear it out: z-ai/glm-5.3, moonshotai/kimi-k2.6, qwen/qwen3.8-max and deepseek/deepseek-v4-pro-0813 all match their vendors to the cent. The fee is at the credit-purchase step, at 5.5% on pay-as-you-go, plus 5% on bring-your-own-key above a $25,000 per month list-price allowance. That allowance is measured in list-price inference cost, not in requests — a detail with real consequences for anyone modelling it as a request quota.

What OpenRouter does not guarantee is which provider serves a slug. With 80-plus providers and 500-plus models, the passthrough promise is per-provider, not per-model. Which is the setup for the next section.

Everywhere else in this index, a batch tier is a discount. It is 50% off at six vendors, 60% of standard at Moonshot, 20% off at xAI on one model. On OpenRouter, three :batch slugs in the current catalogue cost more than the standard slug for the same model. Not a rounding difference — two of them are exactly double.

The mechanism is not a markup, and the passthrough claim in the last section survives intact. Each of the three is passing through a different real vendor rate than the standard slug is passing through. The suffix names an API mode; it does not name a price direction. That distinction is invisible until you check both slugs against the vendor’s own page, which is precisely the work an index is for.

Three of the vendors in this index charge a different rate for the same model depending on how long the prompt is, and in every case the higher rate applies to the whole request, not just the tokens above the threshold. A single input rate per model does not describe what any of them will charge you.

The thresholds differ, the multipliers differ, and one major vendor has no cliff at all on its most-used line. This is the reason the index carries a premium-tiers column rather than folding everything into one input figure — and the mechanics of these thresholds, including how to structure prompts around them, are the subject of our long-context pricing thresholds guide.

One further trap sits underneath this table and is worth naming because it will otherwise silently corrupt a cost model built from it. OpenAI’s headline context window includes output. Its 1,050,000 figure is 922,000 maximum input plus 128,000 maximum output, exactly. Google’s 1,048,576 is an input limit with 65,536 of output on top. The two vendors’ “1M” numbers are not the same quantity and must not be compared directly. The ceilings themselves, by model and by surface, are censused separately in our context window and output limit census.

Every other row in this index is a price in money. Meta publishes one that is not. The Muse Spark 1.2 checkpoint is sold at two prices under two model IDs, and the cheaper one is bought with permission rather than cash.

Meta’s own framing is unusually direct about it: tier “sets the price you pay and whether your data may be used to train future Meta models,” and the Contributor ID carries “heavily discounted pricing in exchange for permission to use your prompts and completions to train future Meta models.” Measured against Meta’s own Standard rate for the same checkpoint, that discount is 92.0% on input, 95.3% on output and 98.7% on cached input.

No other vendor in this index prices on this axis, and it is worth being precise about why that is not the same as saying no other vendor trains on your data. Google runs the closest analogue and runs it the other way round: on the Gemini API free tier, content is used to improve Google’s products; on the paid tier it is not. Same model, same rate card, different data terms decided by the tier rather than by a model ID.

The difference matters for procurement. Google’s split is a consequence of not paying. Meta’s is a price: a published, quantified discount attached to a specific model ID, which means it can be compared against the market rate, put in a budget, and — the part that makes it a governance question rather than a finance one — chosen by a developer with a one-word change to a model string.

The most common way to misuse a price index is to take one cell out of it. Every number above is true of a specific model, on a specific surface, in a specific tier, on August 30, 2026 — and at least one of those four qualifiers is doing real work in most rows.

The wider point is about where the market has moved. Two years ago a price index could be a list of models and two numbers each. The direction of travel since is unmistakable: every large vendor has added tiers, thresholds, endpoints and units, and the differences between them are now the thing that decides a bill rather than a footnote to it. That is unlikely to reverse, because each of these dimensions is a real product with a real cost behind it — data residency is genuinely more expensive to serve, sheddable off-peak capacity genuinely is cheaper. What should be expected instead is more of it: more premium tiers, more units that are not tokens, and more prices that are a function of when and where you called rather than what you called.

The practical response is to stop treating price as an attribute of a model and start treating it as an attribute of a call. Our ledger of AI coding plan limit changes tracks the same volatility on the subscription side, where the ceiling moves rather than the rate. If you want senior help mapping what your workloads actually cost across these surfaces, and what a mid-quarter promotional expiry does to a committed budget, that is the kind of question our AI transformation engagements open with.