Methodology

1. API-equivalent value

For each plan: the dollar value of the usage it includes if you had bought the same usage through the API at list price. Value × is that figure divided by the monthly price, so 1.0x means no subsidy and 5x means each subscription dollar buys five dollars of API usage.

It is an upper bound on what you can extract: you only realise it if you use the full allowance every cycle.

2. Published vs estimated

Published (8 plans). The vendor states the allotment in dollars or in credits with a dollar price: GitHub Copilot (AI Credits at $0.01), Kiro (credits at the $0.04 add-on rate) and Zed ($5 of tokens at list +10%). Kiro's credit-to-token mapping is not public, so its 2x is measured at Kiro's own retail credit price, not at provider API cost.

Estimated (15 plans). Anthropic, OpenAI, Cognition and Google publish limits as multiples, message ranges or "higher limits", not dollars. Cursor publishes pools but not the per-tier dollar amount on the page I could read. For these I give a low / typical / full-plan range. The default is the full-plan end, which assumes you use your whole allowance every cycle. They are my editorial estimates, built from each vendor's stated ratios (for example Max 5x and 20x against Pro, or Plus message ranges) and typical agentic task sizes. I did not have access to measured usage logs, so treat them as orders of magnitude, not facts.

The estimated plans are marked with an orange dot and a dashed badge in every table and chart. You can overwrite any plan's API value in the leaderboard with your own measurement (for instance from ccusage or the Codex usage dashboard); it is stored in your browser only. Switch to Published only to rank on vendor-stated numbers alone.

3. Output tokens per month

Agentic coding re-reads context on every turn, so a bare output-token price understates the real bill. I convert dollars to tokens with a fixed reference workload. For every 1M output tokens it assumes 40M cache-read input tokens, 2M cache-write tokens and 4M uncached input tokens.

Output tokens per month = API value ÷ all-in cost per 1M output tokens of the best model on that plan (the one shown under the plan name). Kiro uses Claude Opus 5; Zed and Google, whose models have no known API price, are priced as if on Claude Sonnet 5.5 and marked assumed. The ratio is a workload assumption, not a measurement; change it in src/lib/metrics.ts.

4. Intelligence

The Artificial Analysis Intelligence Index of the best-scoring model you can reach on the plan, taking each model at its highest-scoring effort setting. Plans whose top model is not on the leaderboard (for example Kiro, which lists Claude Opus 5 but not Opus 5.5) show n/a, and the composite score re-weights over the metrics that exist. Cost per task is Artificial Analysis's own figure for running their suite.

5. Composite score (0 to 100)

A weighted average of three normalised parts, with weights you control on the Overview: Subsidy (Value ×, log scale from 0.4x to 12x), Intelligence (linear from index 30 to 58) and Capacity (output tokens per month, log scale from 0.1M to 100M). Anchors are fixed so a plan's score does not change when you filter the table. Capacity deliberately rewards bigger plans, to balance Subsidy rewarding small, tightly capped ones.

6. Value over time

Three kinds of evidence, each labelled with where it comes from. Fetched entries were read from vendor pages for this build. Recalled entries (older launches, price changes) come from vendor announcements and press coverage and were not re-verified. Snapshots are written by npm run snapshot and are the only way the chart gains new data points, so the history starts the day you first run it.

7. What this does not capture

• Latency, reliability, editor quality and agent harness differences. A 10x-subsidised plan with a worse agent can still lose.
• Bundled extras: Google AI storage, Claude chat features, Copilot code completions and similar.
• Rate-limit changes. Vendors can tighten limits without notice; check the dated provenance before committing to an annual plan.
• Annual discounts (for example Claude Pro at $17/mo). All prices here are month-to-month list prices.
• Plans I could not read: see "Not scored yet" on the Overview.

Updating the data

Everything lives in src/data/*.json: plans.json (prices and API value ranges), models.json (API prices and intelligence), history.json (events), and aa-leaderboard.json. After editing, run npm run snapshot to log a ledger point.