The benchmark price for AI tokens.
Oil has Brent. Shipping has the Baltic Dry. Tokens have no price their sellers don’t set.
Tokens need a benchmark
Every commodity market that works has a reference price no single seller sets. Brent and WTI price crude. The Baltic Dry Index prices bulk shipping. Contracts, budgets and hedges are written against those numbers.
Tokens have none. Each provider sets its own price and changes it without notice, and more tokens reach buyers subsidized or resold. Nobody can tell a fair price from a subsidy, budget a year ahead, or hedge.
What a benchmark needs
Routes
Standard units everyone prices the same way. Model tiers today, the cost of a defined task later.
A panel
Gateways and large buyers reporting what they actually paid, anonymized and weighted by volume.
A public rulebook
How every value is computed, written down before it is used.
An audit trail
Values anyone can reproduce, administered independently of the sellers.
Not every token is the same token
Crude has grades: light, heavy, sweet, sour. Tokens have grades too: serving precision, context length, uptime. The neocloud spread is the first measurement of that.
One open-weight model is sold by 42 providers at prices 10.0× apart, and the cheapest serves it at unknown.
deepseek/deepseek-v4.1-flash-1017
- Providers
- 42
- Endpoints
- 51
- Cheapest
- $0.20
- Dearest
- $1.95
- Spread
- 10.0×
- Cheapest serves at
- unknown
Where STCI is today
- NowListed prices. Every separately priced model on one aggregator (OpenRouter), normalized to dollars per million tokens and combined into four indices under a public rulebook. Each day carries a verification hash that anyone holding the inputs can reproduce.
- NowGrades and spread. For open-weight models, the price of every provider endpoint selling the same model, and the precision each one serves.
- NextA reported-price panel. Gateways and large buyers report what they actually paid, anonymized and weighted by volume. Listed prices cannot show discounts, credits or subsidies; a panel can.
- LaterContracts that settle against the index. Forwards that let a buyer fix next year’s token cost. This needs the panel and independent administration first.
Today’s index
November 20, 2026. Dollars per million tokens, blended at 3 output tokens per input token.
-
STCI-FRONTIER
$12.03
per million tokens, blended
Flagship models, one per lab. 9 models.
in $3.08, out $15.02
-
STCI-EFFICIENT
$1.07
per million tokens, blended
Cost-optimized models, one per lab. 9 models.
in $0.29, out $1.33
-
STCI-OPEN
$1.90
per million tokens, blended
Open-weight models sold by many providers. 23 models.
in $0.34, out $2.41
-
STCI-ALL
$6.02
per million tokens, blended
Every eligible listed model. 343 models.
in $1.60, out $7.49
The neocloud spread
| Model | Endpoints | Providers | Cheapest | Median | Dearest | Max / min | Cheapest by precision |
|---|---|---|---|---|---|---|---|
deepseek/deepseek-v4.1-flash | 39 | 30 | $0.28 | $0.61 | $9.75 | 34.8× |
|
deepseek/deepseek-v4.1-flash-1017 | 51 | 42 | $0.20 | $0.54 | $1.95 | 10.0× |
|
deepseek/deepseek-v4-flash-0731 | 28 | 22 | $0.12 | $0.25 | $1.10 | 9.17× |
|
z-ai/glm-5.3-flash | 42 | 36 | $0.07 | $0.38 | $0.62 | 9.00× |
|
xiaomi/mimo-v2.6-pro | 5 | 5 | $0.57 | $0.76 | $4.62 | 8.09× |
|
mistralai/mistral-nemo | 6 | 5 | $0.02 | $0.03 | $0.15 | 7.37× |
|
nvidia/nemotron-3.5-super-vl | 11 | 11 | $0.28 | $0.41 | $1.91 | 6.89× |
|
meta-llama/llama-3.3-70b-instruct | 7 | 7 | $0.27 | $0.71 | $1.76 | 6.65× |
|
deepseek/deepseek-v4-pro-0813 | 25 | 19 | $1.65 | $3.30 | $10.58 | 6.41× |
|
meta-llama/llama-3.1-8b-instruct | 4 | 4 | $0.04 | $0.15 | $0.22 | 6.29× |
|
openai/gpt-oss-120b | 24 | 21 | $0.12 | $0.40 | $0.75 | 6.23× |
|
google/gemma-4-31b-it | 18 | 15 | $0.16 | $0.34 | $0.96 | 5.98× |
|
The 12 widest spreads of 57 open-weight models measured on November 20, 2026. Blended $ per million tokens, 3 output tokens per input token. Median spread across all 57: 2.53×.
deepseek/deepseek-v4.1-flash- fp8 from $0.28
deepseek/deepseek-v4.1-flash-1017- precision undisclosed from $0.20
deepseek/deepseek-v4-flash-0731- fp8 from $0.12
z-ai/glm-5.3-flash- fp4 from $0.07
xiaomi/mimo-v2.6-pro- bf16 from $0.57
mistralai/mistral-nemo- fp16 from $0.02
nvidia/nemotron-3.5-super-vl- nvfp4 from $0.28
meta-llama/llama-3.3-70b-instruct- fp8 from $0.27
deepseek/deepseek-v4-pro-0813- precision undisclosed from $1.65
meta-llama/llama-3.1-8b-instruct- fp8 from $0.04
openai/gpt-oss-120b- fp4 from $0.12
google/gemma-4-31b-it- precision undisclosed from $0.16
Precision is what each provider reports. It is a fact about the endpoint, not a verdict on the tokens: OpenRouter’s own test (November 11, 2026) found no clear accuracy gap between fp16, fp8 and fp4 on one model, and could not rule out a point or two. The provider can matter as much as the precision.
Source: listed prices on OpenRouter, from its model catalog and each model’s provider endpoints. Listed prices, not prices paid. As JSON: /v1/index/latest.json and /v1/spread/latest.json.
Where it came from
I started STCI on New Year’s Day 2026 and built the first version in a week: the collector, the indexer, the API and the site. It ran every day for two months. Then I got pulled onto other work and switched it off.
Then Guillermo Rauch argued that tokens are becoming the new unit of energy. Energy has benchmark prices; tokens still did not. So I turned STCI back on, and relaunched it on November 19, 2026, with a rulebook written for today’s models and a measurement of grade the first version did not have.
Jeremy Longshore
Report what you pay
The next rung needs prices paid, not prices listed. If you run an LLM gateway, or your company buys tokens at volume, report what you pay. Reports are anonymized and aggregated before anything is published.
Reproduce it
The rulebook is public. Today’s values carry verification hash
b9d8ab92bd647019: SHA-256 over the sorted eligible listings, the rulebook
and the date. Anyone holding the inputs can recompute it, and the inputs are available on request.