Estimated SCU data

Standard Capability Unit

SCU, or Standard Capability Unit, is a normalized unit for measuring the practical AI capability delivered by model tokens. It compares how much usable reasoning, coding, math, multimodal understanding, agentic execution, reliability, and inference efficiency one million tokens can provide across different models and providers.

SCU is not a hardware compute metric. It does not measure GPU FLOPs, training cost, or raw inference compute. It measures capability-adjusted token output.

TPI measures token price. TPX measures token purchasing power. SCU measures AI capability, not hardware compute. Capability Oracle not yet live.

Live SCU Spot Layer

Real-time delivery quotes moved to Live

SCU base rates stay here as the benchmark layer. The live page now carries the spot quote, market state, change windows, and provider live table.

Open Live

Global SCU Rate

9.97 SCU

1M weighted tokens

Estimated

US Frontier SCU Rate

13.07 SCU

1M weighted tokens

Estimated

China SCU Rate

8.19 SCU

1M weighted tokens

Estimated

SCU Purchasing Power

11.46 SCU

$1 estimated capability

Estimated

What SCU Measures

Dollars buy tokens. Tokens generate outputs. SCU measures how much practical AI capability those outputs actually deliver.

TPI and TPX remain token-economic indices. SCU adds a capability layer on top of token economics: TPI is the price layer, TPX is the purchasing-power layer, and SCU is the capability-adjusted token value layer.

Why SCU Exists

USD is the payment unit, model tokens are the invocation unit, and SCU is the capability delivery unit. Different model tokens cannot be compared by price alone because quality, success rate, reliability, and task coverage differ.

Token-to-SCU Exchange Rates

Prototype rates reuse the current Sources model list and price snapshots, then attach estimated SCU metadata for capability, reliability, autonomy, verification, and speed.

Estimated, not live benchmark data

Anthropic

Claude Opus 4.7

Estimated

Region: United States
USD/1M: 5.0000 / 25.0000
SCU/1M weighted: 14.61
SCU/USD: 2.92
Source reliability: official
Last updated: May 17, 2026, 12:00 AM

OpenAI

GPT-5.5

Estimated

Region: United States
USD/1M: 5.0000 / 30.0000
SCU/1M weighted: 13.41
SCU/USD: 2.68
Source reliability: official
Last updated: May 17, 2026, 12:00 AM

Google

Gemini 3.1 Pro Preview

Estimated

Region: United States
USD/1M: 2.0000 / 12.0000
SCU/1M weighted: 12.53
SCU/USD: 6.27
Source reliability: official
Last updated: May 17, 2026, 12:00 AM

xAI

Grok 4.3

Estimated

Region: United States
USD/1M: 1.2500 / 2.5000
SCU/1M weighted: 11.72
SCU/USD: 9.37
Source reliability: official
Last updated: May 17, 2026, 12:00 AM

DeepSeek

DeepSeek V4 Pro

Estimated

Region: China
USD/1M: 0.4350 / 0.8700
SCU/1M weighted: 9.46
SCU/USD: 21.74
Source reliability: official
Last updated: May 17, 2026, 12:00 AM

Alibaba Qwen

Qwen3 Max

Estimated

Region: China
USD/1M: 0.3590 / 1.4340
SCU/1M weighted: 9.03
SCU/USD: 25.15
Source reliability: official
Last updated: May 17, 2026, 12:00 AM

ByteDance Doubao

Doubao Seed 1.8

Estimated

Region: China
USD/1M: 0.8000 / 2.0000
SCU/1M weighted: 8.39
SCU/USD: 10.48
Source reliability: official
Last updated: May 17, 2026, 12:00 AM

Moonshot Kimi

Kimi K2.6

Estimated

Region: China
USD/1M: 0.6000 / 2.5000
SCU/1M weighted: 8.19
SCU/USD: 13.65
Source reliability: official
Last updated: May 17, 2026, 12:00 AM

Baidu

ERNIE 5.0

Estimated

Region: China
USD/1M: 1.7000 / 5.5000
SCU/1M weighted: 7.60
SCU/USD: 4.47
Source reliability: official
Last updated: May 17, 2026, 12:00 AM

Zhipu

GLM-4.6

Estimated

Region: China
USD/1M: 1.4000 / 1.4000
SCU/1M weighted: 7.49
SCU/USD: 5.35
Source reliability: official
Last updated: May 17, 2026, 12:00 AM

MiniMax

MiniMax-M2.7

Estimated

Region: China
USD/1M: 0.3000 / 1.2000
SCU/1M weighted: 7.21
SCU/USD: 24.03
Source reliability: official
Last updated: May 17, 2026, 12:00 AM

Provider	Model	Region	USD/1M Input	USD/1M Output	Output/Input Weight	Capability Score	Reliability	Source Reliability	SCU per 1M Weighted Tokens	USD per SCU	SCU per USD	Data Status	Last Updated
Anthropic	Claude Opus 4.7	United States	5.0000	25.0000	5.00	0.92	0.95	official	14.61	0.3422	2.92	Estimated	May 17, 2026, 12:00 AM
OpenAI	GPT-5.5	United States	5.0000	30.0000	6.00	0.94	0.96	official	13.41	0.3729	2.68	Estimated	May 17, 2026, 12:00 AM
Google	Gemini 3.1 Pro Preview	United States	2.0000	12.0000	6.00	0.89	0.91	official	12.53	0.1596	6.27	Estimated	May 17, 2026, 12:00 AM
xAI	Grok 4.3	United States	1.2500	2.5000	2.00	0.86	0.88	official	11.72	0.1067	9.37	Estimated	May 17, 2026, 12:00 AM
DeepSeek	DeepSeek V4 Pro	China	0.4350	0.8700	2.00	0.84	0.82	official	9.46	0.0460	21.74	Estimated	May 17, 2026, 12:00 AM
Alibaba Qwen	Qwen3 Max	China	0.3590	1.4340	3.99	0.83	0.82	official	9.03	0.0398	25.15	Estimated	May 17, 2026, 12:00 AM
ByteDance Doubao	Doubao Seed 1.8	China	0.8000	2.0000	2.50	0.81	0.81	official	8.39	0.0954	10.48	Estimated	May 17, 2026, 12:00 AM
Moonshot Kimi	Kimi K2.6	China	0.6000	2.5000	4.17	0.82	0.81	official	8.19	0.0732	13.65	Estimated	May 17, 2026, 12:00 AM
Baidu	ERNIE 5.0	China	1.7000	5.5000	3.24	0.80	0.80	official	7.60	0.2236	4.47	Estimated	May 17, 2026, 12:00 AM
Zhipu	GLM-4.6	China	1.4000	1.4000	1.00	0.79	0.79	official	7.49	0.1869	5.35	Estimated	May 17, 2026, 12:00 AM
MiniMax	MiniMax-M2.7	China	0.3000	1.2000	4.00	0.78	0.78	official	7.21	0.0416	24.03	Estimated	May 17, 2026, 12:00 AM

SCU Formula

Output/Input Weight

lambda_m,t = output_price_m,t / input_price_m,t

Output tokens are weighted by their relative price so input and output work share a comparable unit.

Weighted Tokens

WeightedTokens_m,t = InputTokens_m,t + lambda_m,t x OutputTokens_m,t

Weighted tokens normalize mixed input and output usage before capability is compared.

Benchmark Weight

omega_b,t = base_weight_b x freshness_b x anti_leakage_b x challenge_b,t
challenge_b,t = 4 x p_b,t x (1 - p_b,t)

Benchmarks receive more weight when they are current, leakage-resistant, and still challenging.

Capability Score

Capability_m,t = Sum_k W_k,t x C_m,k,t

Capability combines reasoning, math, code, long context, agent, multimodal, safety, and efficiency scores.

Effective SCU Score

SCU_m,t = Capability_m,t x Reliability_m,t x Autonomy_m,t x Verification_m,t x Speed_m,t

The first prototype discounts raw capability by operational quality factors.

Token-to-SCU Rate

SCU_per_1M_weighted_tokens_m,t = 1,000,000 x SCU_m,t / WeightedTokens_m,t

This is the Base SCU Rate. It is benchmark-derived and updates slowly.

Live SCU Spot

Live_SCU_per_1M_tokens_m(t) = Base_SCU_per_1M_tokens_m x DeliveryFactor_m(t)
DeliveryFactor_m(t) = ReliabilityFactor x LatencyFactor x AvailabilityFactor x QueueFactor x RoutingFactor

The live layer adjusts the base rate for delivery conditions that can move every 1 to 5 seconds.

USD per SCU

USD_per_SCU_m,t = Cost_m,t / SCU_m,t

This converts the benchmark run cost into a capability affordability estimate.

Live USD per SCU

Live_USD_per_SCU_m(t) = Base_USD_per_SCU_m x LatencyPenalty x RetryPenalty x QueuePenalty x CapacityPenalty x RoutingPenalty

Live spot pricing applies real-time delivery penalties to the base capability cost.

SCU Spot Price

SCU_Spot_Price(t) = weighted average of top 3 lowest Live_USD_per_SCU_m(t)

The page quote uses the most affordable live capability delivery among eligible models.

SCU Purchasing Power

SCU_per_USD(t) = 1 / SCU_Spot_Price(t)

This answers how much estimated capability one dollar can buy.

Capability Oracle

Click each capability block to inspect how it contributes to the Standard Capability Unit.

Limitations

The first SCU release is a research prototype based on estimated and normalized data. Benchmark results may be affected by data contamination, leaderboard overfitting, task mismatch, prompt sensitivity, and provider-specific execution differences.

Before any SCU rate is treated as live, the system should incorporate live benchmark feeds, real-task validation, historical time-series calibration, and transparent methodology notes.