Three Gauges, One Company, and One That Stopped Answering
Three perpetual futures markets sit on three builder-deployed exchanges inside one venue, and all three reference the same private company: Anthropic. One of them is quoted and traded. A second exists but carries no open interest and an empty book. The third has published the same two prices for seven days. Read at 2026-10-06 01:53 UTC, the spread between the one traded mark and the frozen one is 28.9878 percent. That figure was measured on three dates across seven days, and it widened on every one of them: 26.8696 percent on 2026-09-29, 28.9323 percent on 2026-10-05, and 28.9878 percent on 2026-10-06. One of the three instruments moved. Two did not. The disagreement, not the level, is the finding, and the fact that it grows each time it is checked turns a curiosity into a measurement problem.
A Field Called Previous Day Has Not Rolled in Seven Days
The fastest way to learn what a feed measures is to watch a field that is supposed to move. The vntl exchange lists a market called vntl:ANTHROPIC. Its mark price reads 1619.3, and its prevDayPx reads 1619.3 as well. Those two numbers were identical on 2026-09-29, on 2026-10-05 at 06:45 UTC, on 2026-10-05 at 08:57 UTC, and again on 2026-10-06 at 01:53 UTC. Open interest was 0 on every visit. Twenty-four hour notional volume was 0. Funding was 0. The order book held zero levels on both sides. A market with no quotes, no positions and no funding is not a market in any operational sense. It is a row in a table.
Widen the view and the pattern repeats. All 15 vntl markets carry open interest of 0. Twelve of the 15 report a mark equal to their prevDayPx. Only three differ: vntl:OPENAI at 1336.2 against 1344.5, vntl:MAG7 at 63.661 against 64.339, and vntl:INFOTECH at 191.3 against 191.31. Those three are the only places in the family where the two fields have separated at all.
Now the part that carries the weight. The vntl:OPENAI pair read exactly 1336.2 and 1344.5 on 2026-09-29, on 2026-10-05, and again on 2026-10-06. A field named previous day price has not advanced in seven days. If prevDayPx meant the close of the prior session, it would differ on each date by definition. It does not.
Set that beside a market that is genuinely live. The io:ANTH prevDayPx returned four different values across the measurement window: 2084.1 and then 2078.9 on 2026-10-05, 2082.9 later that day, and 2083.1 on 2026-10-06. Same venue, same request shape. One field rolls; the other has not rolled once. A field named previous day that does not roll is not reporting that nobody traded. It is reporting that nothing was reported.
Three Ceilings, 4.5x Apart, and Only Two Have an Oracle Updater
Each of the three markets publishes a streaming open-interest cap. io:ANTH allows $45,000,000. para:ANTH allows $25,000,000. Every vntl market, including vntl:ANTHROPIC, allows $10,000,000. For the same company, the highest ceiling is 4.5 times the lowest. Maximum leverage narrows in the same order: 6 for io:ANTH, 5 for para:ANTH and 3 for vntl:ANTHROPIC.
The oracleUpdater field is populated for io and for para and is empty for vntl. A role that can move the price reference exists on two of the three markets, and on the third the field is blank. A field left open in a specification and settled later is ordinary design, so it is worth reading how a blank specification field gets filled in after the fact before treating the blank as meaningful. The deployer addresses are visible in the feed itself, and the venue publishes its own explorer, where a reader would start to check what an address has done rather than take the role on description.
Nothing about these caps should be read as a constant. The io:OAI streaming cap read $10,000,000 on one visit and $12,000,000 on a later one, and it still reads $12,000,000 now. No changelog for the change was found. A ceiling an operator can raise between two visits, with no public record, is a risk parameter rather than a fixed property of the market.
The Book Flipped Twice in Seven Days
On 2026-09-29 io:ANTH showed 20 levels, with bid size of 34.604 units against ask size of 3.479 units. In notional terms that was about $71,125 against about $7,151, roughly ten to one leaning on the bid side, with a spread of 0.0097 percent. On 2026-10-05 the same endpoint returned a bid of 9.522 units against an ask of 7.827 units, about $19,851 against about $16,347, nearly symmetric, with a spread of 0.0192 percent. On 2026-10-06 at 01:35 UTC the book stood at a bid notional of $18,380.09 against an ask notional of $26,198.58. Eighteen minutes later, at 01:53 UTC, it was bid $11,165.35 against ask $25,806.86, a ratio of 2.31 to 1 on the ask side, a spread of 0.0048%, and impact prices of 2088.318 and 2088.8.
The company did not change across those snapshots. The quoting did. Quoting shape is a property of who is quoting this week, not a property of Anthropic, so any citation of book depth in these markets has to carry a date beside it or it is describing a moment that has already passed.
Funding Moves Between Samples Ten Seconds Apart; the Mark Does Not
Three samples of io:ANTH, taken ten seconds apart at 01:41:38.520, 01:41:48.115 and 01:41:58.248 UTC, returned a byte-identical mark of 2088.8, a byte-identical open interest of 18662.294 and a byte-identical 24h notional volume of 5413054. The funding rate moved on every one of the three samples: 0.0000164701, then 0.0000164647, then 0.0000164592. A second pair of samples, about twenty seconds apart at 01:53:14 and 01:53:21, repeated the result: funding moved from 0.0000165473 to 0.0000165452 while mark, open interest and 24h volume did not move at all.
That is the control experiment, and it is the strongest evidence here. Funding is recomputed continuously, so it changes between samples. Mark, open interest and 24h notional volume are published snapshots whose refresh cadence an outside reader cannot observe. A static open interest may mean nobody traded for ten seconds, or that the field is written only every few minutes. The feed does not separate the two cases.
The mark also tracks the quotes rather than an independent feed. Mark 2088.7 sat against a book mid of 2088.75 and an oracle price of 2085.9, a mark-to-oracle premium of 0.1342 percent. When the mark hugs the mid that tightly, it is a summary of the current book rather than a separate source of price.
One scale check keeps the number honest, and it is worth recomputing a published figure from its own inputs instead of citing one. At a mark of 2088.7 and open interest of 18662.294 units, the notional open interest is $38,979,933. That is 86.62% of io:ANTH’s own $45,000,000 streaming cap, and 1.1664 percent of the core BTC perpetual open-interest notional of $3,341,853,052 measured at the same moment. The cap is a field the feed publishes; what the venue does when open interest reaches it is not documented anywhere an outside reader can check, so the honest reading of 86 percent is that 13 percent of headroom is left.
Implied Value Needs an Error Bar
Take the one traded mark and the frozen one on 2026-10-06: io at 2088.7 and vntl at 1619.3. The spread is 28.9878%. The same pair read 26.8696 percent on 2026-09-29 and 28.9323 percent on 2026-10-05, so it widened on each of the three measurement dates, by about 2.06 percentage points across the first six days and a further 0.0555 overnight. para:ANTH sits at 2000.0, which is 4.2467 percent below the io mark, with zero open interest, zero volume and an empty book on both sides.
Three marks, three numbers, one company. Each implies a different value for the same underlying, and no published arbiter stands between them. Anyone who converts one of these marks into an implied valuation is quoting the widest disagreement between three unarbitraged feeds, not a price. The venue publishes these markets as isolated-margin only, so a reader should not assume the three are comparable instruments when nothing in the feed makes them so.
What the Feed Does Not Carry
A call to the allMids endpoint returned 1,237 entries at 2026-10-06 01:53 UTC, including BTC at 85517.5 and ETH at 2703.95. It held no key for io:ANTH, para:ANTH or vntl:ANTHROPIC. The single call most readers would use to get every mid price in the venue carries none of these three markets.
Getting them requires knowing the builder-dex parameter name. The request is a POST to api.hyperliquid.xyz/info with a body of type metaAndAssetCtxs and a dex field naming the exchange; the perpetuals reference describes that call as retrieving perpetuals asset contexts and, in its own words, “includes mark price, current funding, open interest, etc.” The parameter is named dex, not perpDex, and passing the wrong name raises no error; it returns the default set. The same documentation states that dex defaults to the empty string, which is the first perp dex. It does not define prevDayPx, and it publishes no refresh cadence for any field. A reader who guesses the parameter wrong can end up believing three markets have zero activity when they were never queried at all. That failure mode is silent: the endpoint answers, answers completely, and answers about a different set of markets than the one asked about.
| Gauge | What it reported at 2026-10-06 01:53 UTC | What that reading tells you |
|---|---|---|
| io:ANTH | mark 2088.7, midPx 2088.75, oraclePx 2085.9, open interest 18662.294 units, 24h notional volume 5,421,357.91, funding 0.0000165473 per hour, max leverage 6 | A live market: quotes on both sides, open-interest notional at 86.62 percent of its streaming cap, and a funding rate that changes between samples. |
| para:ANTH | mark 2000.0, prevDayPx 2000.0, open interest 0, 24h volume 0, funding 0, empty book on both sides | A newly created market with no activity recorded; its mark equals its prevDayPx for that reason, which is not the same as being frozen. |
| vntl:ANTHROPIC | mark 1619.3, prevDayPx 1619.3, open interest 0, 24h volume 0, funding 0, empty book on both sides | The same two numbers returned on four visits across seven days; a field named previous day that never rolls. |
| the core BTC perpetual on the same venue | mark 85766.0, open interest 38964.77686 units, notional 3,341,853,052 | The scale reference: io:ANTH open-interest notional is 1.1664 percent of it. |
What this reading now accepts as settled is narrow and physical. Three markets reference one private company. Only one quotes and trades, and its mark sits 28.9878 percent away from the third, a gap that widened at each of the three measurement dates. vntl:ANTHROPIC has published the same mark and the same prevDayPx for seven days with no open interest, no volume, no funding and no book. Funding moves between samples while mark, open interest and volume do not. And the generic mid endpoint carries none of the three.
The limit of the reading is just as concrete. No exchange outside this family was checked, so 28.9878 percent cannot be called the widest gap for this company anywhere. allMids carries none of these builder coins, so no independent cross-check of the marks was possible from that endpoint. The funding sign convention was not read from documentation, so only its magnitude is reported. The deployer role can change caps, leverage and the oracle updater, and no changelog was found. Four snapshots over seven days cannot show who decided what inside any of the three exchanges. Nothing here establishes what Anthropic is worth, and no figure in this piece should be read as a valuation.
