"Free historical options data" is a phrase with three very different meanings. We pulled each source and checked the dates ourselves. Here is what is real.
The free tier, checked
| Source | Starts | Fields | Delisted | Notes |
|---|---|---|---|---|
DoltHub post-no-preference/options | 2020 | bid, ask, IV, greeks | yes | ~1,500 names, breadth, no volume/OI; public API is slow and caps at 1,000 rows/page |
| Public Parquet mirror (R2) | 2008 | bid/ask/sizes, volume, OI, IV, greeks | no | 8 deep names, 2008→2025, 69.8M rows, 1.72 GB; no 15:45; freezes Dec 2025 |
| Kaggle (SPY surface) | 2010 | IV surface, precomputed | one ticker | SPY only, wide format, stops ~2023; great for IV surfaces, not a chain archive |
| OptionsDX | 2010 | bid/ask/last, IV, greeks | limited | paid, per-year orders; strong on a handful of liquid tickers |
The recurring gap is the same everywhere: free does not go back to 2007, free does not have the full universe, and free does not have every field. The Parquet mirror is the only free source with volume and open interest; DoltHub is the only free source that keeps delisted names; no free source has all three.
The paid tier, honestly
OPRA-resourced vendors (Cboe DataShop, the big data providers) go back to 2007 and earlier, carry the full universe and index roots (SPX, VIX, RUT), and include the exchange identifiers that make a rename a single entity. The price is the price: real money, per name or per archive, with a vendor licence if you redistribute. That is the tier you buy when your research depends on 2007, on the full tape, or on IWM/SMH/IGV rather than just SPY and QQQ.
The math on a full chain
Why is a 2008-2025 archive $2,400 and not $50? Count it: the eight-name deep set is 69,856,552 rows and 1.72 GB compressed — that is roughly 4,500 sessions × 15,000 contracts a day, each with a 36-column schema, cleaned, checksummed, and documented. The cost is not the bits; it is the cleaning, the gap-finding, the split mapping, and standing behind the numbers. A $50 "full history" is almost always a back-filled index or a single-name subset wearing a coat.
How we would actually buy it
- Research, one name, one regime: the free Parquet mirror, or one paid name/year.
- IV surfaces / vol research: the Kaggle SPY set, free, is genuinely the fastest start.
- Cross-sectional or screening: DoltHub for breadth, or the breadth tier here for a clean, checksummed delivery with delisted names kept.
- 2007, the full tape, or index roots: pay. There is no free path and pretending otherwise is how research goes wrong.
Start with the sample
SPY + NVDA, a full week of March 2020, 68,250 rows, 4.6 MB, checksummed. If it holds up, the tiers above are the natural next step.
See the tiers Datasheet