Vast.ai listing

vast.ai · Compute · 2 records · JSON

Vast.ai ComputepaidcardcardDretention-clamped
Vast.ai listing structure ComputepaidcardcardMretention-clamped

Vast.ai Compute

paidcardcardDretention-clamped

vast.ai/pricing · GPU cloud · compute/vast-ai · JSON · compare

Free tier	no free tier; prices are set by supply and demand across 40 or more data centres unit none; metered on wall-clock; interruptible instances are stated as 50% or more cheaper and may be reclaimed
Price	none stated
Per-object cap	not applicable
Retention	a stopped instance keeps its disk for a period the host sets, not Vast
Egress	not published on the page read
Identity	card a card and a balance are required, and the minimum deposit is USD 5, quoted from the vendor's quickstart
Limits	retention-clamped interruptible instances may be reclaimed at any time; that is the discount source request-rate no per-account concurrency number is published on the pricing page source prices are set by the host, so the rate for a card changes with supply
Binding	retention-clamped interruptible instances are reclaimable, so preemption is the clock and the discount is its price
Interface	rest, cli REST API, CLI and a web console; SSH into an instance
Tooling	the Vast CLI and API, SSH, Docker images
Composes with	the cheapest GPU hour in the market and the least reliable: the host is a third party
Reliability	no reliability event recorded
Notes	The pricing page renders live card prices client-side, but the marketplace API does not: a probe of https://console.vast.ai/api/v0/bundles/ returned HTTP 200 and 177,005 bytes with 64 offers, including GTX 1660 S at USD 0.021 per GPU-hour, RTX 3090 at USD 0.109, RTX 4090 at USD 0.120 and RTX 5090 at USD 0.213. The earlier note said no rate was recorded; that was true of the pricing page and false of the API, which is the better source. The rates are read from one snapshot on one day and move with host supply.
Evidence rung	documented
Checked	2026-10-22 (live); stale on 2026-11-21

Evidence (15)

date	source	what it proved
2026-10-22	vast.ai/pricing	three tiers: on-demand with per-second billing, interruptible at 50% or more cheaper, and reserved on 1, 3 or 6 month terms
2026-10-22	vast.ai/pricing	prices are set by supply and demand across 40 or more data centres; 68 or more GPU types from RTX 3060 to B200
2026-10-22	vast.ai/launch-your-gpu	quoted: 'Is there a free trial? - Not currently'; 'Do I need a credit card to try it? - Yes - but you can start with just USD 5 in credits'
2026-10-22	docs.vast.ai/guides/get-started/quickstart	quoted: 'The minimum deposit amount on Vast.ai is USD 5'; 'When your balance reaches zero, your running instances will automatically stop'
2026-10-22	docs.vast.ai/guides/concepts	quoted: 'Reliability score - A measure of a machine's historical uptime and health. New machines start at 60% and climb as they demonstrate availability. Higher reliability matters more the longer your rental.'
2026-10-22	docs.vast.ai/guides/concepts	three rental kinds: on-demand fixed and guaranteed until max duration, reserved, and 'Interruptible, bidding-based, lowest cost, may be paused when outbid or when on-demand demand spikes'
2026-10-22	console.vast.ai/api/v0/bundles/?q=%7B%7D	probe.py recorded HTTP 200, 177,005 bytes, sha256 14ae21ac30941206b38860c6107be97bdf05a59132e289ec0b25d6bca7241f78, 64 offers, no challenge; parsing the JSON gives cheapest per-GPU rates of USD 0.021 for a GTX 1660 S, 0.109 for an RTX 3090, 0.120 for an RTX 4090 and 0.213 for an RTX 5090
2026-10-22	docs.vast.ai/guides/instances/choosing/instance-types.md	interruptible instances are 'often 50%+ cheaper than on-demand', may be paused if outbid, and 'Data preserved when paused but instance not functional', quoted
2026-10-22	docs.vast.ai/pricing	billing is by the second for actual usage and the marketplace sets rates per host, so no fixed per-card spot price exists
2026-10-22	console.vast.ai/api/v0/bundles/?q=%7B%22rentable%22%3A%7B%22	The vendor's own unauthenticated search endpoint returned 200 with 64 rows sorted ascending by dph_total. The cheapest row with gpu_frac 1.0 is gpu_name 'GTX TITAN X', dph_total 0.042222222222222223, min_bid 0.03333333333333333, gpu_total_ram 12288, geolocation 'South Korea, KR', reliability2 0.9849486, host_id 154621, verification 'unverified'.
2026-10-22	console.vast.ai/api/v0/bundles/?q=%7B%22rentable%22%3A%7B%22	The same 64 rows carry gpu_frac values of 0.125, 0.142857, 0.166667, 0.2, 0.25, 0.333333, 0.5 and 1.0, counted as 7, 4, 5, 2, 9, 1, 18 and 18 rows. The cheapest row is gpu_name 'Quadro P2000' at gpu_frac 0.125 with dph_total 0.02962962962962963, geolocation 'Nevada, US', host_id 90577, and its gpu_ram and gpu_total_ram both read 5120 MB, which is the whole 5 GB card, not the slice.
2026-10-22	console.vast.ai/api/v0/bundles/?q=%7B%22rentable%22%3A%7B%22	The interruptible query returned 200 with 64 rows on 56 distinct host_ids. The cheapest whole card is gpu_name 'GTX 1060' at dph_total 0.03338889222222222, geolocation 'Vietnam, VN', host_id 172022. The same host 154621 and the same whole Titan X appear at dph_total 0.0356 against 0.0422 on the on-demand query, and min_bid over dph_total across the 64 on-demand rows has a median of 0.825.
2026-10-22	console.vast.ai/api/v0/bundles/?q=%7B%22rentable%22%3A%7B%22	Sorted by card count descending, the largest single offer is gpu_name 'RTX 5090' with num_gpus 16, gpu_total_ram 521712 MB, dph_total 14.296, geolocation 'Taiwan, TW', host_id 51010. The num_gpus distribution in the response is 8x35, 4x11, 6x7, 7x4, 5x2, and one each of 9, 10, 12, 14 and 16.
2026-10-22	cloud.vast.ai/create/	The listing page rendered without login and priced each offer in the same form, for example 'Type #46832290', 'Quebec, CA', '16 GB VRAM', '$0.195/hr plus bandwidth'; 'Type #28418865' shows '2x H200 NVL' at '$8.032/hr plus bandwidth'; 'Type #53243440' shows '1x RTX 6000Ada' at '$0.645/hr plus bandwidth'. Every one of its offer cards ends in the words 'plus bandwidth'.
2026-10-22	vast.ai/pricing	'68+ GPU Types' and 'From RTX 3060 to B200' and 'Prices set by supply and demand across 40+ data centers.' The interruptible tier is described as '50%+ cheaper. Best for batch training.' and reserved as 'Up to 50% Off' with '1, 3, or 6 month terms'. The page contains no price and no rate: the live table it advertises did not render into the extracted text.

More fields (13)

{
 "shape": {
  "vcpu": null,
  "ram_gib": null,
  "disk_gib": null,
  "gpu_class": null,
  "gpu_count": null
 },
 "billing_model": "allocated",
 "granularity": {
  "min_billed_s": 1,
  "rounding_s": 1
 },
 "session_cap": {
  "wall_clock_min": null,
  "idle_reap_s": null,
  "restart_required": false,
  "note": "interruptible instances can be reclaimed at any moment"
 },
 "keep_rate": {
  "value": null,
  "basis": "not established; disk is on the host machine",
  "source": "https://vast.ai/pricing"
 },
 "persistent_storage": {
  "included_gib": null,
  "price_gib_month": null,
  "durability": "host disk",
  "survives_stop": null,
  "survives_destroy": false,
  "max_retention": "set by the host"
 },
 "network_out": {
  "allow_outbound": true,
  "allow_inbound_listen": true,
  "ports": "SSH and mapped ports",
  "static_ip": null,
  "ipv6": null,
  "per_session_bytes": null
 },
 "free_tier_kind": "none",
 "m10h_usd": null,
 "m24h_usd": null,
 "gap": "the pricing page is client-rendered; the API at console.vast.ai/api/v0/bundles/ returns the live offers and was probed 2026-10-22, so the rate gap is closed and the snapshot date is recorded",
 "wave2": {
  "merged_agents": [
   "a04-decentralized",
   "a05-spot",
   "a09-forums",
   "vast-ai-supply"
  ],
  "merged_evidence_rows": 13,
  "merged_note": "The cheapest whole single GPU that was actually rentable on-demand at 2026-10-22T17:00Z was a GTX TITAN X 12 GB at USD 0.04222 per hour on host 154621, South Korea, reliability 0.985, 20 effective CPU cores, 7.8 GB RAM, 151 GB disk, 606 Mbps down and 404 Mbps up, egress USD 4.00 per TB, host marked unverified. The cheapest single-GPU listing of any kind is not a card: a Quadro P2000 at gpu_frac 0.125, one eighth of a 5 GB card, at USD 0.02963 per hour on host 90577, Nevada, reliability 0.997. Of the 64 cheapest single-GPU on-demand rows, 18 are whole cards and 46 are fractional slices, spread "
 },
 "rate_snapshot": {
  "source": "https://console.vast.ai/api/v0/bundles/?q=%7B%7D",
  "date": "2026-10-22",
  "note": "64 offers; cheapest consumer cards quoted in notes; a marketplace rate moves with host supply"
 }
}

Unknown or withheld values (4)

$.egress.metered egress-unknown
$.per_object_cap.notes process
$.per_object_cap.stated not-applicable
$.reliability_note none-observed

Vast.ai listing structure Compute

paidcardcardMretention-clamped

console.vast.ai/api/v0/bundles/ · GPU cloud · compute/vast-ai-listing-structure · JSON · compare

Free tier	The public offer endpoint answers without authentication and returns a 64-offer page of 176,798 bytes as application/json on 2026-10-22, so the marketplace prices can be read without an account. What that page contains is not what it appears to be: only 3 of the 64 offers carry rentable true, none is an interruptible bid, and several prices are for a fraction of a card. The cheapest listing is an RTX 3090 at USD 0.0881 an hour at a GPU fraction of 0.5, which is USD 0.175 for a whole card, and the cheapest RTX 4090 is listed at USD 0.1215 at a fraction of 0.125, which is USD 0.972 for a whole card. Whole-card equivalents from this page: RTX 5070 Ti USD 0.101, RTX 4080S USD 0.109, RTX 3090 USD 0.175, RTX 5090 USD 0.213, A100 PCIe USD 0.536, RTX 4090 USD 0.802. unit GiB; metered on wall-clock; none
Price	none stated
Per-object cap	not applicable
Retention	The existing record's fact holds and is what binds: interruptible instances may be reclaimed at any time, and that is the discount rather than a defect.
Egress	not stated
Identity	card Unchanged from the existing record: a card and a USD 5 minimum deposit.
Limits	retention-clamped The existing record's fact holds and is what binds: interruptible instances may be reclaimed at any time, and that is the discount rather than a defect. source The existing record's fact holds and is what binds: interruptible instances may be reclaimed at any time, and that is the discount rather than a defect.
Binding	retention-clamped The existing record's fact holds and is what binds: interruptible instances may be reclaimed at any time, and that is the discount rather than a defect.
Interface	rest inferred from the tooling the provider documents; not probed
Tooling	not stated
Composes with	Because the endpoint needs no key, a nightly script can record the whole-card floor for a card class over time and detect when a rate moves, which is the only price history that does not depend on a third-party tracker.
Reliability	no reliability event recorded
Notes	Supplies the reason the existing vast-ai.json record carries no rate, and refutes the naive reading of this endpoint: a session that took dph_total at face value would record a 3090 hour at USD 0.088 and a 4090 hour at USD 0.122, both of which are 1/2 and 1/8 of a card. Three candidate explanations for the low rentable count were considered: that the default page is a catalogue sample rather than the live bookable set, that the account-less view hides rentable rows, and that most offers are genuinely idle hosts. The evidence cannot separate them, so no floor is published. What would settle it is one authenticated search with the same filters, or a documented query syntax from the vendor's API reference.
Evidence rung	measured
Checked	2026-10-22 (live); stale on 2026-12-06

Evidence (7)

date	source	what it proved
2026-10-22	console.vast.ai/api/v0/bundles/	probe.py record: HTTP 200, 176798 bytes, sha256 5b76c6ed15fd5854ccbeb98b0637827464cbe90016c1f14fed517e278672ca89, content-type application/json, full body, no challenge, no false 200
2026-10-22	console.vast.ai/api/v0/bundles/	64 offers across 16 GPU models, of which 3 carry rentable true and 0 carry is_bid true; the response sets truncated false
2026-10-22	console.vast.ai/api/v0/bundles/	cheapest row is gpu_name RTX 3090 at dph_total 0.0881 with gpu_frac 0.5, and the cheapest RTX 4090 row is dph_total 0.1215 with gpu_frac 0.125, so both listed prices buy a slice rather than a card
2026-10-22	console.vast.ai/api/v0/bundles/?q=%7B%22rentable%22%3Atrue%2	a query string asking for rentable whole-GPU offers returned 0 offers, which establishes either that the query syntax is wrong or that no such offer is on the default page; the two cannot be separated from this evidence
2026-10-22	cloud.vast.ai/api/v0/bundles/?q=%7B%22limit%22%3A500%2C%22or	GET returned 64 offers; the three cheapest 8 GB-VRAM rows were RTX 3060 Ti at dph_total 0.0287 (reliability 0.811), 0.0289 (reliability 0.999) and Tesla P100 16 GB at 0.0296
2026-10-22	cloud.vast.ai/api/v0/bundles/?q=%7B%22limit%22%3A500%2C%22or	cheapest rows by VRAM: 24 GB at 0.0689 (GTX TITAN X) and 0.0708 (RTX A2000); 32 GB at 0.0563 (Tesla P100); 64 GB at 0.1096 (Tesla P100 pair)
2026-10-22	vast.ai/pricing	'Pay only for what you use, down to the second. No minimum hours, no rounding up, no surprise charges', quoted; the price table on that page is script-rendered and carried no numbers to the extractor

More fields (1)

{
 "wave2": {
  "class": "gpu-marketplace",
  "free_tier_kind": "none",
  "shape": {
   "vcpu": null,
   "ram_gib": null,
   "gpu_class": "RTX 3090",
   "gpu_count": 1,
   "vram_gib": 24
  }
 }
}

Unknown or withheld values (7)

$.egress.metered egress-unknown
$.egress.terms not-stated
$.per_object_cap.notes process
$.per_object_cap.stated not-applicable
$.reliability_note none-observed
$.retention.notes process
$.tooling not-stated