Open data: AI coding plans, models and prices

Everything on this site is generated from these files. Use them in your own tools, articles and research under CC BY 4.0: free to copy, change and redistribute, commercially too, as long as you credit codingplans.dev. Last verified 2026-11-01.

Files

JSON is the canonical format. Every file is served with open CORS, so a browser app can fetch it directly. CSV exports flatten lists into ; -separated text.

Dataset	Contents	Download	Fields
plans	73 AI coding subscriptions: price, annual price, included models, usage limit, main limitation, adoption estimate, source and check date.	JSON · CSV	Schema
models	42 models: Artificial Analysis Intelligence Index, speed, latency, context window, cheapest token price, cost per task and effort profiles.	JSON · CSV	Schema
providers	Token prices by provider, including the DeepSeek peak and off-peak rate tables.	JSON	Schema
apps	Coding agents and how to switch models in each (OpenCode, Command Code, Pi, OpenRouter).	JSON	Schema
schedules	DeepSeek peak windows in UTC and the off-peak discount, machine-readable.	JSON	Schema
enterprise	Contract and team plans with seat prices, minimums, terms and security features (SSO, SCIM, audit logs, data residency).	CSV	Schema

How the data is checked

Every record has a source (the vendor page, or Artificial Analysis for benchmarks) and a source_date: the day someone last checked that page, not the launch date.
A validator runs before every build. It rejects a record without a source or date, a date in the future, an annual price above the monthly price, or a schedule whose peak rate does not match its published multiple. A dataset that fails is not published.
Records not rechecked within 45 days are flagged for review.
Prices are copied, not estimated. Where a vendor publishes no number the field is empty and a reason, starting with an em dash, says why.
Benchmark figures are copied from Artificial Analysis without rounding.

Confidence labels

Plan popularity (popularity_confidence): published is a vendor figure, reported comes from credible reporting, proxy is an install or download count, assumed is an explicit estimate. The score is an adoption estimate, not market share; the method is on the About page.
Enterprise seat prices (confidence): published on the vendor page, reported by credible sources, estimated (an explicit estimate; the record's reason field says how it was reached), or unpublished (shown as "Quote").

Plan fields

Field	Type	Meaning
`id`	string	Stable lower-case identifier, vendor-prefixed. Also the anchor of the plan on its vendor page; never changes.
`name`	string	Plan name as the vendor sells it.
`vendor`	string	Company that sells the plan.
`monthly_price`	number	Price in US dollars per month with monthly billing. Per seat for team plans. 0 for free tiers.
`annual_monthly_price`	number or null	Effective monthly price in US dollars with annual billing (the yearly price divided by 12, rounded to cents), or null when the vendor sells no annual option.
`annual_monthly_price_reason`	string	Why annual_monthly_price is what it is, starting with an em dash, e.g. "— (monthly billing only)".
`kind`	string: IDE extension, Editor / IDE, CLI / Terminal, Cloud agent & IDE, Web chat & coding agent, API / PAYG, Coding subscription	Product category: where you use the plan.
`included_models`	array	Models the plan gives access to, as the vendor names them. A note in parentheses marks a restriction, e.g. "(usage credits)".
`usage_limit`	string	The allowance as the vendor describes it, in plain text. Vendors rarely publish exact request counts.
`main_limitation`	string	The most important catch or restriction, in plain text.
`vendor_notes`	string	Further context from the vendor page.
`source`	string	Vendor page the price and limits were taken from.
`source_date`	string	Date (YYYY-MM-DD) the source was last checked. Not the launch date.
`buyer`	string: individual, team	Who buys it self-serve. team = multi-seat with shared billing and admin, bought without a sales contract; it is shown as the entry rung of the vendor's enterprise ladder. Contract plans live in enterprise.json.
`seat_pricing` optional	object	How a team plan's price scales with seats. The price itself stays in monthly_price / annual_monthly_price (per seat for per_seat and base_plus_seat, for the whole team for per_team).
`popularity_score` optional	integer	Adoption index 0-100, log-scaled from the product's published reach and a documented tier share. Not a vendor-published metric.
`popularity_basis` optional	string	The raw adoption signal behind the score plus the tier share used, in plain words.
`popularity_source` optional	string	Where the adoption signal behind popularity_score comes from.
`popularity_source_date` optional	string	Date (YYYY-MM-DD) the popularity source was checked.
`popularity_confidence` optional	string: published, reported, proxy, assumed	How solid the adoption signal is: published (vendor figure), reported (credible reporting), proxy (installs or downloads), assumed (explicit estimate).

Model fields

Field	Type	Meaning
`id`	string	Stable lower-case identifier. Also the anchor of the model on its vendor page; never changes.
`name`	string	Model name as the vendor publishes it.
`vendor`	string	Company that trains the model.
`intelligence_index`	number or null	Artificial Analysis Intelligence Index score at the highest measured reasoning effort, copied without rounding; null when unmeasured.
`intelligence_index_reason`	string or null	Why intelligence_index is null, starting with an em dash.
`intelligence_index_version` optional	string or null	AA Intelligence Index version the score was measured on, e.g. 5.0. Null when no score is published for the model.
`median_speed_tps`	number or null	Median output speed in tokens per second (Artificial Analysis); null when unmeasured.
`is_median_speed_est`	boolean	True when median_speed_tps is an estimate rather than a measurement.
`ttft_seconds`	number or null	Seconds to the first answer token, including thinking time (Artificial Analysis).
`end_to_end_seconds`	number or null	Seconds to a complete 500-token answer, including thinking time (Artificial Analysis).
`context_window`	string	Maximum context window, e.g. "1M" or "256K" tokens.
`open_weights`	boolean	True when the weights are publicly downloadable.
`specialized` optional	boolean	True for a specialised decision model kept below the index floor; see specialized_reason.
`specialized_reason` optional	string or null	Why a specialised model is kept despite a low index.
`cheapest_token_price`	string	Lowest list price as "$input / $output per M" (US dollars per million tokens).
`cheapest_provider`	string	Where cheapest_token_price is offered.
`cost_per_task`	string	Artificial Analysis cost to run its Intelligence Index on this model at API list prices, in US dollars; a guide to relative cost, not a bill.
`source`	string	Main source page for the record.
`benchmark_source`	string	Artificial Analysis model page the benchmark figures come from.
`pricing_source`	string	First-party pricing page (or OpenRouter) for the token price.
`source_date`	string	Date (YYYY-MM-DD) the sources were last checked.
`effort_profile` optional	object	Score and cost at each reasoning-effort level on an agentic coding benchmark, used to recommend an effort level.

Units: prices in US dollars; token prices per million tokens; speed in tokens per second; times in seconds; dates as YYYY-MM-DD. The full rules for every file are in the JSON Schemas.

Research: OpenCode Go allowance benchmark

The raw data behind OpenCode Go usage limits, measured: 26 models, 4 agent tasks each, every model call logged, run 2026-10-31 on OpenCode v2.1.4. With these files you can rerun the benchmark on your own OpenCode Go account and check the projection.

File	Contents
`results.csv`	One row per measured model: requests, cost, mean tokens per call, published and projected requests per five hours for Go and Go Plus, and runway. Generated from the run below.
`raw.jsonl`	Every model call in the run: model, task, step, input, cache read and write, output and reasoning tokens, cost in US dollars and latency in milliseconds.
`summary.json`	Per-model totals, per-request means and medians, and the result of each hidden check.
`analysis.md`	The projection to 50K, 100K and 150K-token sessions, as written by analyze.mjs.
`published-2026-10-31.json`	OpenCode's published prices, monthly caps and requests per window, parsed from its usage-limits page.
`tasks.json`	The four agent tasks each model ran, in one session.
`verify.mjs`	The hidden checks: the bug is fixed, the billing change and rename are made, and the test suite passes.
`bench.mjs`	Runs every paid opencode-go model through the tasks in its own copy of the fixture and logs each call.
`analyze.mjs`	Turns a run into the projection used by the brief.
`fixture/`	The small Node project each model works on, with one real bug: `AGENTS.md`, `package.json`, `src/cost.mjs`, `src/format.mjs`, `src/limits.mjs`, `src/prices.mjs`, `src/report.mjs`, `test/cost.test.mjs`, `test/format.test.mjs`, `test/limits.test.mjs`, `test/report.test.mjs`.

To rerun: copy the files into one folder, keep fixture/ beside the scripts, then run node bench.mjs and node analyze.mjs runs/<timestamp>. It needs the OpenCode CLI signed in to OpenCode Go and spends real allowance.

How to cite

Credit codingplans.dev and link the page or file you used. For example:

codingplans.dev (2026). AI coding plans dataset. Retrieved 2026-11-01 from https://codingplans.dev/data/plans.json. CC BY 4.0.

Spotted a wrong number? Email hello@codingplans.dev with the record id and a source.