Correct snapshot figures, add plans.jsonl (80 rows)

README.md

@@ -1,5 +1,5 @@
 ---
-license: cc-by-4.0
+license: mit
 language:
 - en
 tags:
@@ -7,8 +7,8 @@ tags:
 - coding-agents
 - pricing
 - llm
-- benchmarks
 - open-data
+- comparison
 pretty_name: AI coding plan prices (tokens and requests per dollar)
 task_categories:
 - question-answering
@@ -18,51 +18,70 @@ configs:
 - config_name: default
   data_files:
   - split: train
-    path: latest.json
+    path: plans.jsonl
 ---
 
 # AI coding plan prices (tokens and requests per dollar)
 
-Daily snapshot of what the flat rate AI coding subscriptions cost and what you get for the money:
-quota numbers, per model spending caps, included models, and the resulting tokens and requests per
-dollar paid. Published by [vibeplan.cc](https://vibeplan.cc/coding-plan-comparison/), the comparison
-that this dataset is built from.
+What the flat rate AI coding subscriptions cost, and what you actually get for the money: published
+quotas, per model spending caps, included models, and the resulting tokens and requests per dollar
+paid. This is the dataset behind [vibeplan.cc](https://vibeplan.cc/coding-plan-comparison/).
 
-Snapshot in this revision: ****, 80 plans,
-0 of them directly comparable.
+Snapshot of this revision: **2026-10-02 19:08 UTC**, 80 plans from 26 providers,
+53 of them directly comparable, built from 30 official sources.
+Disclosure split: 51 disclosed, 21 reference, 4 partial, 2 reported, 1 undisclosed, 1 measured.
 
 ## Files
 
-- `latest.json` - the full dataset, identical to what the website loads
-  ([https://vibeplan.cc/data/latest.json](https://vibeplan.cc/data/latest.json)).
+- `plans.jsonl` - one plan per line, the fields as described below (this is the file to load).
+- `latest.json` - the complete dataset exactly as the website loads it, including the comparison
+  tables, the AI score snapshot, the FX rates and the changelog.
 
-## What is in it
+## Columns in `plans.jsonl`
 
-- `plans[]` - one entry per plan: provider, monthly price with fees included, billing period,
-  the plans own published limits (messages, tokens, requests, hours), per model spend caps,
-  included model list, and whether the plan is comparable at all.
-- `comparable` - `true` only where the provider publishes a hard, recurring limit. Plans whose
-  limits are unknown or only described in vague terms stay `false` and are excluded from the rankings.
-- rates are normalised to tokens per US dollar and requests per US dollar of the price actually paid.
-- each row keeps its source URL and the date it was verified.
+- `id`, `name`, `provider` - plan identity.
+- `price` - effective monthly price in US dollars, **fees included**, plus the billing period.
+- `meter`, `quotas`, `windowCaps` - the plan's own published limits (messages, tokens, requests,
+  hours) and how they reset.
+- `modelRows`, `modelStats` - per model rates and the per model spending caps a plan applies.
+- `disclosure` - how well the plan documents its limits: `disclosed`, `partial`, `reported`,
+  `reference` (kept for context, not comparable) or `undisclosed`.
+- `dataTier` - A/B/M/C/D quality tier of the source situation behind that plan.
+- `sourceIds`, `verifiedAt` - where each figure came from and when it was last checked.
+- `notes`, `priceNote` - the plain language caveats that belong to that row.
+
+Rates are normalised to tokens per US dollar and requests per US dollar of the price actually paid,
+at a target comparison price of 1 USD.
+
+## Loading
+
+```python
+from datasets import load_dataset
+ds = load_dataset("harrytyp/ai-coding-plan-prices", split="train")
+ds[0]
+```
 
 ## How it is built
 
-Rebuilt every day from around 30 official sources (provider pricing pages, docs, plan pages)
-by the build pipeline in the [source repository](https://github.com/harrytyp/coding-plan-comparison).
-Fees are included in the price (`one number, no separate fee marker`). Where a provider does not
-publish a quota, the field stays empty rather than being estimated, and the plan is marked as not
-comparable. Prices are converted with a daily FX feed.
+Rebuilt every day from around 30 official sources (provider pricing pages, docs,
+plan pages) by the pipeline in the
+[source repository](https://github.com/harrytyp/coding-plan-comparison). The build never guesses a
+number: where a provider does not publish a quota, the field stays empty and the plan is marked as
+not comparable instead of being estimated. Prices include card and processing fees, so the figure is
+what you actually pay.
 
 ## Limits, stated plainly
 
-- Only plans whose limits are published can be compared. Several providers keep their caps vague,
-  so the comparable subset is smaller than the total number of plans.
+- Only plans that publish a hard recurring limit are comparable, 53 of 80 at this snapshot.
+  Providers with vague or unpublished caps stay out of the rankings on purpose.
+- `reference` rows are documented plans that cannot be ranked (for example because nothing is
+  published about the included amount).
 - API prices per token are a different question and are not part of this dataset.
-- Promotional prices (first month, annual billing) exist; the dataset uses the effective monthly
-  price and keeps the billing period.
+- Promotional pricing exists (first month discounts, annual billing). The dataset keeps the
+  effective monthly price and the billing period rather than hiding it in a single number.
 
-## License and attribution
+## License and citation
 
-Data is published under CC BY 4.0, the build code under MIT. Attribution:
-[vibeplan.cc](https://vibeplan.cc/coding-plan-comparison/).
+MIT, same as the site. Numbers change daily, so cite with a link to
+[https://vibeplan.cc/](https://vibeplan.cc/) rather than copying figures.
+Corrections and source suggestions: [issues](https://github.com/harrytyp/coding-plan-comparison/issues).