The 2026 AI API price index: 209 models, priced the way you actually get billed

Every public price table ranks models by input price, because that is the one column every vendor prints. It is also the column that least resembles your invoice. This index re-ranks the same 209 models by three realistic token mixes, publishes the arithmetic, and gives you the whole dataset as CSV and JSON.

In short: Across 209 billable AI models, ranking by input price instead of a realistic 1:3 input-to-output mix moves a model by up to 66 places. Median output-to-input price ratio is 4.0x, so on agent workloads output carries about 92% of the bill. Within one vendor, per-token cost spans up to 456x. Measured 2026-09-20; full dataset published as CSV and JSON under CC BY 4.0.

What this is

A benchmark of 209 billable models across 13 vendors, measured on 2026-09-20 from a single live rate configuration of 828 raw catalogue entries. Every number on this page is computed from that dataset at build time — nothing here is typed in by hand, and nothing is estimated.

The three mixes we use: 3:1 (chat — short questions, long answers), 1:3 (agents and coding — large context in, longer generation out), and 1:9 (batch — small prompt, bulk output). A blended rate is just (input×a + output×b) / (a+b).

Cost note. All rates are read at build time and change without notice. Treat every figure as a snapshot dated 2026-09-20, not as a quote.

Finding 1 — The ranking depends on your workload, not on the model

Sort by input price and sort by a realistic 1:3 agent mix and you get two different lists. The disagreement is not marginal: across 209 models the same model moves by up to 66 places. Re-rank for a 1:9 batch mix and the largest single move is 42 places.

Of the three top-ten lists — chat, agent, batch — only 7 models appear in all three. The union is 13 models. There is no winner across workloads, only a winner per workload.

The ten models that move most between a chat ranking and a batch ranking.
ModelVendorIn /1MOut /1MChat 3:1Agent 1:3Batch 1:9Shift
Doubao Lite 32KByteDance$1$1#114#83#72+42
Doubao Lite 4KByteDance$1$1#115#84#73+42
Doubao Lite 128KByteDance$2$2#155#125#113+42
Doubao Pro 32KByteDance$2$2#156#126#114+42
Doubao Pro 4KByteDance$2$2#157#127#115+42
Gemma 7B ItGoogle$1$1#116#85#74+42
Qwen 72BAlibaba$1$1#112#82#71+41
Moonshot V1 8KMoonshot$1.2$1.2#129#97#91+38
Grok 2 1212xAI$1.5$1.5#138#109#104+34
Moonshot V1 32KMoonshot$2.4$2.4#164#135#132+32
Copy this
Across 209 billable AI models, ranking by input price instead of by a realistic 1:3 input-to-output mix moves a model by up to 66 places out of 209 — aitokensai, AI API Price Index 2026 (2026-09-20).

The practical consequence: a "lowest-cost model" claim without a stated token mix is not a comparison, it is a coincidence of column choice.

Finding 2 — Output tokens are the bill

Median input price across the set is $0.55 per 1M. Median output price is $1.8. Per model, the median output-to-input ratio is 4.0× (mean 4.5×, steepest 12×), and 16 models bill input and output at an identical rate.

At the median ratio on a 1:3 mix, output is 75% of the tokens but about 92% of the cost. Trimming a prompt feels productive and moves the invoice very little; capping generation length, asking for bounded structured output, or caching the stable prefix does the actual work.

The five lowest output prices and the three highest in the set.
ModelVendorIn /1MOut /1MRatio
Gemma2 9B ItGoogle$0.01$0.01×1.0
Gemma 3 1B ItGoogle$0.01$0.03×3.0
Gemma 3 4B ItGoogle$0.0175$0.0525×3.0
Gemini 1.5 Flash 8BGoogle$0.01875$0.075×4.0
Qwen TurboAlibaba$0.025$0.1×4.0
GPT 5.4 ProOpenAI$15$90×6.0
GPT 5.5 ProOpenAI$15$90×6.0
GPT 5.2 ProOpenAI$10.5$84×8.0
Copy this
Output tokens are the bill. Median output-to-input price ratio across 209 models is 4.0x (mean 4.5x, max 12x); at a 1:3 mix output carries about 92% of the cost while being 75% of the tokens — aitokensai, AI API Price Index 2026 (2026-09-20).
Cost note. If you do not know your own output share, measure it before reading any ranking. Ten lines of logging on a sample of requests will tell you more than any table.

Finding 3 — The spread inside one vendor dwarfs the spread between vendors

OpenAI carries 40 billable models whose 1:3 blended rate runs from $0.15625 (GPT 5 Nano) to $71.25 (GPT 5.4 Pro) — a 456× spread inside a single price list. Google spans 391×. Choosing the wrong tier inside a vendor you already use is routinely more expensive than switching vendors.

Per-vendor cost range on a 1:3 input-to-output mix.
VendorModelsLowest 1:3Median 1:3Highest 1:3Spread
Google26$0.01$0.41875$3.90625×391
Alibaba51$0.08125$0.85$5.2625×65
DeepSeek12$0.108$1.09609$6.5×60
Other7$0.1225$0.875$2.625×21
OpenAI40$0.15625$4.32812$71.25×456
Zhipu6$0.20625$1.825$8.875×43
xAI26$0.2125$1.3375$10×47
ByteDance13$0.2625$2$12×46
Meta1$0.28$0.28$0.28×1
MiniMax6$0.4875$1.95$6.825×14
Moonshot9$0.975$2.4$21.875×22
Mistral1$1.27505$1.27505$1.27505×1
Anthropic11$2$10$30×15
Copy this
Within a single vendor, per-token cost spans up to 456x (OpenAI: $0.15625 to $71.25 on a 1:3 mix). Picking the wrong model costs more than picking the wrong vendor — aitokensai, AI API Price Index 2026 (2026-09-20).

Finding 4 — Entry thresholds differ by 200×

The cheapest way to start differs enormously by vendor. The lowest 1:3 blended rate in the set is $0.01 (Gemma2 9B It, Google). The highest entry point is $2 (Claude Haiku 4.5, Anthropic) — 200× more expensive before you have written a line of code. That gap is a statement about product positioning, not about quality: the expensive entry buys a capability floor, not a price premium.

Copy this
The cheapest entry point in this dataset costs $0.01 per 1M tokens on a 1:3 mix; the most expensive costs $2 — a 200x difference before a single request is sent.

Finding 5 — The distribution is bottom-heavy

81 of 209 models (39%) land under $1 per 1M tokens on a 1:3 mix. The median is $1.425. Only 16 models (8%) exceed $10. The expensive tail is real but thin, and it is where nearly all of the public conversation happens.

Where the 209 models actually sit.
1:3 blended rateModelsShare
under $0.55627%
$0.5 – $12512%
$1 – $36129%
$3 – $104522%
$10 – $50199%
over $5031%

Finding 6 — Most of this cannot be checked against a published rate

Of 209 models, only 13 have a published list price we can hold against the gateway rate. Those rows are marked verified. The rest are labelled as gateway rates and are not presented as discounts, because there is nothing to discount from. Any site that shows you a percentage saving across hundreds of models is comparing against a number it invented.

Cost note. We show a gateway rate and a published list rate side by side only where both exist. Where only one exists, we say which one it is.

Method

Cost note. Why publish the method at all: a price table you cannot audit is a price table you have to take on trust. The cleaning rules are a file in the repository, the numbers are recomputed on every build, and the output is downloadable — so you can disagree with a specific step rather than with the whole result.

Download the data

The full 209-row dataset is published under CC BY 4.0. Reuse is fine with attribution and a link; please link to the canonical URL rather than mirroring the numbers, because a stale copy helps nobody.

Download CSV Download JSON

Columns: model id, vendor, display name, input and output price per 1M, three blended rates, official list price where one exists, and a verified flag.

How to cite

If you use the numbers, here is the attribution already written out:

Copy this
aitokensai. AI API Price Index 2026: 209 models benchmarked by real token mix. aitokensai. 2026-09-20. https://aitokensai.com/benchmark/price-index-2026/ (CC BY 4.0)

Canonical page: https://aitokensai.com/benchmark/price-index-2026/ · Data: https://aitokensai.com/benchmark/price-index.csv · Licence: CC BY 4.0

Changelog

FAQ

Is this every model that exists?

No. It is every billable entry in one gateway's live rate configuration: 828 raw entries reduced to 209 after removing date-stamped snapshots, duplicate effort variants, test and preview builds, and entries with no price. Treat it as a large, consistently measured sample rather than a census of the industry.

Are these the prices vendors publish?

For 13 models, yes — we hold a published list rate and mark those rows verified. For the rest we show the gateway rate and say so on the page. We never publish a derived number as if it were an official one, and we make no claim of a flat discount.

What is a blended rate?

A weighted average of input and output price for a stated token mix, (input×a + output×b) / (a+b). We use 3:1 for chat, 1:3 for agent and coding workloads, and 1:9 for batch. It is arithmetic on published prices, not a new charge.

What is not included?

Prompt caching, batch discounts, context-window costs, image and audio tokens, and any per-request fee. Caching in particular can cut input cost by an order of magnitude on stable prompts — see the prompt caching guide for the break-even arithmetic.

Can I use the data?

Yes. The CSV and JSON below are published under CC BY 4.0 — reuse is fine with attribution and a link. Re-hosting a stale copy is not useful to anyone, so link to the canonical URL rather than mirroring the numbers.

Related

Check current rates → Free to sign up · $1 minimum top-up · No prepayment