---
title: "Methodology: how real AI prices per token are computed | Tokenews"
description: "How Tokenews computes a real price per million tokens (monthly fee ÷ usable monthly tokens), the standard workload mix, how leaderboard scores keep their context, and how evidence is graded."
url: https://tokenews.xyz/methodology/
updated: 2026-10-08
source: Tokenews (https://tokenews.xyz), data from Real API Pricing (https://github.com/Fei2411/real-api-pricing, MIT)
---

Methodology · how to read the data

# Every number has a story.

Our contribution is the real price axis. Capability scores come from independent leaderboards, with their original context preserved.

**Real unit price** = Monthly fee ÷ usable monthly tokens (USD / 1,000,000 tokens)

## 01 A floor, not a promise

Allowances assume saturated use. Lower usage increases your effective price. All token types count equally: input, output, cache reads, and cache writes in measured totals. We do not adjust for tokenizer differences or utilization.

A month defaults to 4 weeks. Kimi has an independent monthly pool equal to five weekly pools; Droid's month is 2× its weekly allowance. Model allowances within the same plan are alternatives and must not be summed.

## 02 One comparison workload

Dollar/credit pools priced at public rates and metered APIs are converted using the same three-part workload; for Anthropic models the input share is priced at the 5-minute cache-write rate. Measured samples that include a token breakdown are converted to list-worth at public prices and then to the channel's workload (low-cache for Google and StepFun, Anthropic for Claude); samples without a breakdown and official token tables keep raw totals and are flagged as not workload-normalized. Vendor dashboard dollars are calibrated from measured usage, not treated as public-price dollars.

- **97%** Cache reads
- **2.5%** Fresh input
- **0.5%** Output

This is a comparison convention, not a measured workload. Cache-write charges are modeled only in the Anthropic workload (input share at the 5-minute cache-write rate); other channels with write fees may still see overstated allowances.

## 03 Keep the benchmark context

Each leaderboard uses one selected snapshot; different benchmark versions are never mixed. Code Arena refers to WebDev Overall, not general coding ability. The default view is each model's highest-score configuration; every archived configuration and the reasoning-effort filter remain available. Harness, reasoning effort, service mode, and known score intervals are retained.

Mappings do not verify a subscription’s harness or quota-measurement effort. Missing scores remain missing. Qualitative confidence is not a numerical error bar, and score intervals do not change frontier membership.

## 04 Follow the evidence

High: dashboard back-calculations, controlled saturation measurements, or official absolute-token/credit tables. Medium: usage logs plus quota reports, or official ratios applied to a strong baseline. Low: single reports, third-party summaries, and cross-plan assumptions.

Conflicting historical evidence is archived. The adopted decision and its rationale are exposed for every point. Original evidence text is preserved in its source language.

## Independent leaderboards

[Artificial Analysis Intelligence Index v4.3Intelligence Index · Snapshot 2026-10-01](https://artificialanalysis.ai/leaderboards/models)[Terminal-Bench 4.0Resolution Rate % · Snapshot 2026-09-21](https://www.tbench.ai/leaderboard)[Terminal-Bench 4.0 (AA)Resolution Rate % · Snapshot 2026-10-01](https://artificialanalysis.ai/leaderboards/models)[Code Arena | WebDev OverallArena Score · Snapshot 2026-09-30](https://arena.ai/leaderboard/code)[Agent Arena · Overall (Ranking / Pareto)Net Improvement % · Snapshot 2026-09-30](https://arena.ai/leaderboard/agent)[Artificial Analysis Coding Agent Index v1.5Coding Agent Index · Snapshot 2026-10-01](https://artificialanalysis.ai/agents/coding-agents)[OpenDesign Arena · Quality rankingAverage task score · Snapshot 2026-10-01](https://open-design.ai/zh/llm-arena-for-design/)[DeepSWE v1.1Pass@1 % · Snapshot 2026-09-03](https://deepswe.datacurve.ai/)

Currency conversion: 1 USD = 6.7787 CNY · 2026-09-04 · [CFETS central parity, via Xinhua](https://www.xinhuanet.com/20260904/4c1ce1cbf39d416997cb9c237d6f5024/c.html)

[Full attribution & licensing](https://github.com/Fei2411/real-api-pricing/blob/main/SOURCES.md) · [Download adopted data](https://tokenews.xyz/data/adopted.csv) · [All configurations](https://tokenews.xyz/data/benchmark-configurations.json)

[Explore the data](https://tokenews.xyz/)