how many token
中
☰

00Data & method

Data & method

Multiply input, cache-read, output and cache-write tokens by their prices, add the four parts into a per-request cost, then use the model allowance to estimate monthly tokens. CP value is the coding score divided by USD cost per 100M token-equivalent; monthly, five-hour and weekly windows are checked separately.

01

Data and dates

One snapshot carries the plan limits, model rows and external evaluation date together.

Plan dataPlan data updated: 2026-09-06
Evaluation dataArtificial Analysis evaluation data updated: 2026-09-04
CoverageCurrently 145 plan options covering 94 models.

02

How the four token classes enter the estimate

Each request keeps input, cache read, output and cache write separate, then adds them as token-equivalent. Cache read is processed volume; it is not fresh input sent again each time.

Token class Official reference profileGrok 4.6 Illustrative common profile input 1,000 · cache read 10,000 · output 1,000 · cache write 0 Illustrative default: input 1,000, cache read 10,000, output 1,000, cache write 0. It explains the four-token relationship and does not replace the official profile in any table row.
Input3901,000
Cache read32,50010,000
Output1201,000
Cache write00
Total33,01012,000

This is an example of a provider reference profile in the data; different models may use different mixes.

03

Formulas

Every result can be followed back through token classes, prices and allowance in that order.

  1. 01

    Cost per request = sum of each token class multiplied by its price per 1M, divided by 1,000,000.

  2. 02

    Estimated requests per month = floor(model monthly allowance ÷ cost per request).

  3. 03

    Estimated monthly token-equivalent = estimated requests per month × (input + cache read + output + cache write). Cache read counts as processed volume; it is not fresh input.

  4. 04

    Coding score / cost ratio = coding score ÷ USD cost per 100M token-equivalent; it is omitted when inputs are missing.

A short worked example

Multiply each token class by its price, add the four parts into a per-request cost, then divide the model's monthly allowance by that cost to estimate monthly requests for the workload.

04

External evaluation data

Intelligence, coding and agentic scores come from Artificial Analysis and are fetched separately from the four official plan sources. They are comparison fields, not results measured by this site on coding tasks.

Evaluation data version: 4.1 Evaluation data fetched: 2026-09-04

How CP value is calculated

Coding score / cost ratio for the same workload, for comparison.

This ratio divides an Artificial Analysis coding score by the USD cost of 100M token-equivalent, making the score-to-price ratio easier to compare for the same workload. A higher value means more benchmark score per unit of cost.

Using the shared workload above: Muse Spark 1.3 Contributor (OpenCode Go) has a coding score of 76.3 and costs $0.0267 per 1M token-equivalent, which is $2.6667 per 100M token-equivalent. 76.3 ÷ 2.6667 = coding score / cost ratio 28.6.

This ratio uses score and price only. Capacity limits, speed, context length, availability and real-task outcomes need separate judgment.

05

How to read the time windows

The monthly allowance, five-hour limit and weekly limit are separate conditions. The reference profile estimates per-request cost; concentrated requests may meet the five-hour or weekly window first. Command Code GOAT rolls its window from the first request; use the official documentation for the exact conditions.

01How model allowances and shared plan pools work

Command Code GOAT uses one shared credit pool; OpenCode Go lists a usage equivalent for each model. Model allowances cannot be added together as plan capacity.

02Per-request usage changes the available count

The reference request is an estimate basis. Five-hour and weekly windows limit concentrated use; Command Code GOAT rolls its window from the first request.

03Read discounted and list prices separately

The GOAT page shows discounted prices for MiMo V2.5, MiMo V2.5 Pro and MiniMax M3. This site keeps list price, current price and the deal note in separate columns.

04Prices and allowances for speed variants

Command's GLM-5.2 Fast and Kimi K2.7 Code HighSpeed are separate models; compare the price and allowance listed for each one.

05Conditions for free models

Command's free models do not consume a model allowance, but starting a session may still require USD 1 of credits, and capacity limits still apply.

06Peak windows and context tiers change the cost

DeepSeek has UTC peak pricing, and some OpenCode models switch to another price set above a context threshold. The reference profile sits in ordinary short context and does not blend every context branch into one number.

07Included capacity and purchased credits

Command pay-as-you-go credits are not bound by the windows, and OpenCode's Use balance may fall back to Zen balance after the Go limits. Both are listed separately from included capacity.

08Model lists and prices are updated

Official documentation can change its model list or prices. This page shows the current data date; comparable later snapshots will be listed in the update history.

06

Sources

Why OpenCode Zen capacity fields stay empty OpenCode Zen bills per token. Its official page publishes no reference request profile and no plan allowance, so its cost per 1M token-equivalent uses this site’s common workload, and capacity fields such as calculated_requests_month and calculated_tokens_month stay empty rather than being filled with 0.

07

Data scope

Capacity, external evaluations and plan limits answer different questions.

  • 01

    Capacity figures are either published by the provider or estimated from official prices and a selected workload.

  • 02

    External evaluation scores are shown separately from results on real coding tasks.

  • 03

    The current data date and sources are listed in the sources section; later comparable snapshots will appear in the update history.

08Technical data

Raw fields, source hashes, full timestamps and evaluation match states remain in the data files and exports for readers who need to trace a figure.

opencode.ai
2026-09-05T22:18:04+00:00 · c308e789823081b9a79cdbb81f7cc3f2ee58b5256d4a9b815fad34069fd7dfee
commandcode.ai
2026-09-05T22:18:04+00:00 · 551cd4a9eab82d8d6a7877cda92c47ba395eac320f57de94642bf56813d1e5b0
commandcode.ai
2026-09-05T22:18:04+00:00 · cea5e1f0013e9daf558442c289b1cb82620384dfbb02f79d0eb7f8a511e59fe6
commandcode.ai
2026-09-05T22:18:04+00:00 · 76dd3408a5a84ac405a68447bafde262a097f2874b5c30714197369539956fc1
opencode.ai
2026-09-05T22:18:04+00:00 · 33756f95d8ba03afe7996760231ecb8eaaff2928b38d6642e3ec73849b451b1b
generated_at
2026-09-05T22:24:27+00:00

09

Update history

Keep comparable snapshots in one place so the current data is not mistaken for a historical trend.

When a comparable snapshot exists, changes to prices, allowances, model lists and published figures will be collected in the updates page.

View updates ↗