This is the weekly GPU rental price report from Lantern Works for the week of Nov 2, 2026. Every number below comes from two public feeds, fetched live on 2026-11-02 at 09:14 ET:

• gpu-prices.json: feed_updated_at 2026-11-02T12:59:30Z (2026-11-02 07:59 ET), 151 rates from 25 vendors
• gpu-price-changes.json: compares the daily snapshots dated 2026-11-01 and 2026-11-02

All prices are USD per GPU-hour and cover compute only, so they leave out storage, egress, extra vCPU/RAM and taxes. "Cheapest" means cheapest among the providers we track. Check the vendor's own page before you spend money. Price data is licensed CC BY 4.0 (attribution: Lantern Works).

What moved (2026-11-01 to 2026-11-02)

2 down, 27 up, 2 new, 1 removed.

Verda moved all 24 of its tracked rates up by 0.8% to 1.0% (on-demand and spot, from V100 to GB300); the table shows its top-end on-demand rows. Vast.ai's RTX PRO 6000 Server median barely changed ($1.7344 to $1.7337). Marketplace rows (Vast.ai, Spheron) are medians of host offers, so they swing more than list prices.

New listings: Vast.ai RTX PRO 6000 Max-Q 96GB at $2.1087 (Marketplace on-demand (verified)), Vast.ai V100 32GB at $0.1603 (Marketplace on-demand (verified)). Removed: Vast.ai RTX PRO 6000 Workstation 96GB (was $2.0385).

Same GPU, 100 hours: what the spread costs

For an H100 SXM 80GB on-demand, single GPU and non-interruptible, 100 hours costs $299.00 at Civo, compared with a $399.00 median and $484.10 at Verda (formerly DataCrunch). Across the 9 tracked rates that meet those rules, that is up to $185.10 of difference in compute alone.

Cheapest on-demand per GPU (non-interruptible)

Covers all 37 GPU variants with an on-demand rate in the feed, newest and largest first. "N-GPU minimum" means the vendor only sells that instance with N GPUs; the price is still per GPU.

Lambda's B200 rate is for an 8-GPU instance; its single-GPU B200 instance is $6.99.

Cheapest spot / interruptible per GPU

These instances can be preempted by the vendor, so they sit in their own table. Don't compare them directly with the on-demand rates above.

Cheapest community cloud per GPU

RunPod's Community Cloud runs on third-party hosts. The feed marks it as not interruptible, but it is a different service level from on-demand, so it gets its own table too.

Not ranked above: marketplace medians (Vast.ai, Clore.ai, Spheron), serverless per-second rates (Modal, Koyeb), reservations and commitments (Hyperstack, CloudRift) and monthly dedicated servers (Ucartz) are separate tiers. Held out of every table and of "What moved" (held_stale): Beam, Clore.ai, CloudRift, Runcrate, SimplePod and TensorDock. For Clore.ai and CloudRift only the flagged rows are held (thin Clore.ai markets, and CloudRift commitment rates no longer on its pricing page); their other live rows stay in the feed.

Compare vendors

• RunPod vs Verda
• Lambda vs RunPod
• CoreWeave vs Nebius
• Daily price changes

Get this data yourself

• Live comparison: https://thelanternworks.com/?ref=hashnode
• Feeds: https://thelanternworks.com/api/v1/gpu-prices.json and https://thelanternworks.com/api/v1/gpu-price-changes.json (CC BY 4.0)
• Methodology: https://thelanternworks.com/methodology/?ref=hashnode
• MCP server: lantern-works-deals-mcp on PyPI. Its get_price_changes and compare_gpu_prices tools return these feeds to your agent.
• Spot a wrong price? https://thelanternworks.com/report-price/?ref=hashnode

100 hours on one H100 SXM

From the H100 100-hour cost page, on-demand, compute only (no storage, egress or tax):

Hyperstack and Thunder Compute tie at $320 for 100 hours. The median of these four in-stock rates is $359.50, so the cheapest option saves about $39.50 (about 11%). Against RunPod Secure Cloud at $399 the gap is $79. A fifth rate, from Verda (formerly DataCrunch) at $474.60, was out of stock when checked and isn't counted.

If PCIe is enough for your job, Hyperstack's H100 PCIe was $2.50/hr ($250 for 100 hours) at the same check.

The daily index

The Lantern Works GPU Price Index takes each tracked vendor's lowest on-demand-class rate, one vote per vendor, and reports the median. Spot, reserved and thin marketplace rows are excluded. On 2026-10-29:

The index only has three days of history (since 2026-10-27), so I'm not drawing trend conclusions yet. Most day-to-day changes so far come from adding vendors, not price moves. The data is CC BY 4.0.

The raw feed is at https://thelanternworks.com/api/v1/gpu-prices.json. At 18:01 ET it held 100 rates from 19 vendors, each with a source URL and its own verification time.

curl -s https://thelanternworks.com/api/v1/gpu-prices.json | jq '.rates | sort_by(.hourly_equiv_usd) | .[0:5]'

Price alerts instead of checking by hand

The most useful change since the last write-up is hosted price alerts. The hosted MCP server has a watch_price tool. You give it a GPU and a max hourly price, and after the daily refresh it sends a signed POST to your webhook when a tracked rate drops to or below that price. Setup:

1. Connect any Streamable HTTP MCP client to https://lantern-works-mcp.lantern-works.workers.dev/mcp?ref=hashnode. No install and no API key.
2. Call create_watch_key with a public https webhook URL. It returns an api_key and a signing secret once, so store them. It's free, there's no payment, and email is optional.
3. Call watch_price, for example gpu: "H100", max_hourly_usd: 2.5. Use list_watches and delete_watch to manage your watches.

The ?ref=hashnode tag only feeds anonymous call counts per tool and day.

To run it locally, the same server is on PyPI as lantern-works-deals-mcp (version 0.3.24 at the 18:01 ET check, MIT):

uvx lantern-works-deals-mcp

The watch tools are hosted-only. Locally and on the hosted server you also get compare_gpu_prices, estimate_job_cost, get_price_changes, the stack optimizers and deal search. Setup for Cursor, Claude Desktop, VS Code and others is on the install page.

Caveats

• "Cheapest" means cheapest among tracked providers, not the whole market.
• The lowest hourly rate isn't always the lowest cost per job. Compare VRAM and throughput for your workload (estimate_job_cost helps with that).
• A rate that hasn't been re-checked for 7 days is held out of rankings.

Disclosure

Some links in the data and the MCP results are referral links, and Lantern Works may earn a commission if you sign up through them. That never changes the price you pay. On the H100 SXM table above, Thunder Compute and RunPod are referral links. Hyperstack and Lambda are direct. Referral links are flagged in the feed (affiliate: true), and rankings are by price only. The vendor links in the tables on the site go to the vendors' own pricing pages.

This is the weekly GPU rental price report from Lantern Works, refreshed on Oct 29. Every number below comes from two public feeds that were fetched live on 2026-10-29 at 16:45 ET:

• gpu-prices.json: feed_updated_at 2026-10-29T19:08:23Z (15:08 ET), 100 rates
• gpu-price-changes.json: compares the daily snapshots from 2026-10-28 and 2026-10-29

All prices are USD per GPU-hour. Totals cover compute only, so they leave out storage, egress, extra vCPU and taxes. "Cheapest" means cheapest among the providers we track. Check the vendor's page before you spend money. Price data is licensed CC BY 4.0 (attribution: Lantern Works).

What moved (2026-10-28 → 2026-10-29)

1 down, 11 up, 26 new, 1 removed.

Verda's other rates (A100 SXM, RTX A6000, plus spot) also rose by about 1%.

New listings: Hyperstack (H100 SXM $3.20, H100 PCIe $2.50, H100 NVLink $2.60, H200 SXM $3.99, A100 SXM $1.60, A100 PCIe $1.35, plus reservation and spot rows), CloudRift (L40S $0.63 and RTX 5090 $0.65 on-demand, plus 1- and 3-month reserved rates), Novita AI (RTX 4090 $0.67), OVHcloud US (L40S $1.80), Koyeb (L40S serverless $1.20), Spheron (L40S $0.96, RTX 5090 $0.84, aggregator), SaladCloud (RTX 4090 batch $0.16, interruptible) and Clore.ai (RTX 4090, RTX 5090 marketplace). Removed: Ucartz Radeon 780M.

Same GPU, 100 hours: what the spread costs

For an H100 SXM 80GB on-demand, single GPU and non-interruptible, 100 hours costs $320.00 at Hyperstack or Thunder Compute (tied), compared with a $399.00 median and $474.60 at Verda. Across the 5 tracked providers that is up to $154.60 of difference in compute alone.

Cheapest on-demand per GPU (non-interruptible)

Cheapest spot / community per GPU (interruptible or shared hosts)

These rows are in their own table because the instance can be preempted or runs on a community host. Don't compare them directly with the on-demand rates above.

Marketplace medians (Vast.ai, Clore.ai, Spheron), serverless per-second rates (Modal, Koyeb), reservations (Hyperstack, CloudRift) and monthly dedicated servers (Ucartz) each count as a separate tier and are not included in either table. Held out of all rankings (held_stale): TensorDock, because its published table is stale, and Clore.ai's thin-marketplace rows.

Compare vendors

• RunPod vs Verda
• Lambda vs RunPod
• CoreWeave vs Nebius
• Daily price changes

Get this data yourself

• Live comparison: https://thelanternworks.com
• Feeds: https://thelanternworks.com/api/v1/gpu-prices.json and https://thelanternworks.com/api/v1/gpu-price-changes.json (CC BY 4.0)
• MCP server: lantern-works-deals-mcp on PyPI. Its get_price_changes and compare_gpu_prices tools return these feeds to your agent.

Affiliate link: Lantern Works may earn a commission if you sign up through this link. This does not change the price or ranking you see.

If you're building an agent, at some point you have to answer a boring question: what will the infrastructure cost each month? Scraping, proxies, a search API, an LLM, maybe some GPU hours. optimize_agent_stack in the Lantern Works MCP server answers it from a plain-text description of the agent.

It parses the description with keywords and regex (no LLM), then picks the cheapest option with a cited price for each component among tracked providers. Every line shows the math, the vendor's pricing URL and when the price was checked. If a component has no cited price, it says so instead of guessing.

Try it

uvx lantern-works-deals-mcp   # lantern-works-deals-mcp 0.3.11: stdio MCP server, 7 read-only tools, no API key

No install: connect any Streamable HTTP MCP client to https://lantern-works-mcp.lantern-works.workers.dev/mcp?ref=hashnode (it keeps only anonymous call counts per tool, day and ref tag).

Browser version: https://thelanternworks.com/stack/?utm_source=hashnode&utm_medium=post&utm_campaign=stack-optimizer · OpenAI Agents SDK example: https://github.com/connormatthewdouglas/agent-work-stack-deals/tree/main/examples/stack-optimizer

Real output (PyPI 0.3.11, run Oct 29, 2026, 10:15 AM ET)

tools: compare_deals, compare_gpu_prices, get_deal, get_price_changes, optimize_agent_stack, optimize_stack, search_deals
job: scrapes 20k JS pages/month via 5 GB residential proxies, 3k search queries, summarizes with an LLM, budget $300/mo
  scraping: ZenRows Launch, $69.0/mo, 20,000 x 5 = 100,000 credits; within 250,000 included at $69/mo, https://www.zenrows.com/pricing
  proxy: 9Proxy Residential (per GB), $3.4/mo, 5 x $0.68 per GB, https://9proxy.com/pricing  (referral link)
  llm_api: Groq free tier, $0.0/mo [entry_price], https://console.groq.com/docs/rate-limits
  search: Exa Starter credits + pay as you go, $4/mo, $0/mo incl. 2,500 + 1 x $4 per 1,000 extra, https://exa.ai/pricing  (referral link)
    alternatives: Firecrawl Hobby $24/mo (referral link), SearchApi Developer $40/mo (referral link)
$69.00 (scraping) + $3.40 (proxy) + $0.00 (llm_api) + $4.00 (search) = $76.40/mo (complete), within budget

job: serve a 13B model for 40 GPU hours, send 5k emails, budget $150/mo
  gpu: RunPod RTX A6000 (Community Cloud), $13.2/mo, $0.33/hr x 40 h, https://www.runpod.io/pricing  (referral link; manual signup only)
    alternatives: Thunder Compute $14.0, Verda spot $16.4, RunPod Secure Cloud $23.6
  email: SendPulse Basic (free), $0.0/mo, within 12,000 included, https://sendpulse.com/prices/smtp  (referral link)
$13.20 (gpu) + $0.00 (email) = $13.20/mo (complete), within budget

(Output condensed from the tool's JSON. Use the example script's --json flag for the full result.)

A few things to notice:

• The numbers show their work, for example "20,000 x 5 = 100,000 credits; within 250,000 included at $69/mo".
• The LLM line is entry_price: the cheapest cited entry plan (a free tier). Real usage may need a bigger plan.
• The 13B line uses a stated rule: 13 x 2 bytes (FP16) x 1.2 = 31.2 GB of VRAM on one GPU for serving. Training needs far more.
• RunPod is flagged manual_signup_only, because the vendor's terms ban automated signups.
• GPU rates are re-verified daily, and the response includes data_checked_at and next_refresh_utc. get_price_changes(since=...) shows what moved.
• "Cheapest" means cheapest among tracked providers. Prices are monthly list prices before tax. Check the vendor page before you pay.

Disclosure

Disclosure: some result links returned by the server are referral links. They're marked affiliate: true with a referral_disclosure (tagged "(referral link)" above), and affiliate status isn't a ranking factor. source_url is always the vendor's own pricing page.

Repo: https://github.com/connormatthewdouglas/agent-work-stack-deals

"Cheapest" in this post always means cheapest among tracked providers, not the whole market.

Cheapest tracked on-demand rate per GPU (verified 2026-10-29)

88 rates from 14 providers: 99Stack Cloud, CoreWeave, DigitalOcean GPU Droplets, Hyperstack, Lambda, Modal, Nebius, Novita AI, RunPod, SaladCloud, Thunder Compute, Ucartz, Vast.ai and Verda (formerly DataCrunch). Rates are per GPU-hour, from each vendor's own pricing page or API. This table leaves out spot, interruptible and reserved tiers.

Each rate has the vendor quote it came from, a source URL and a verification time. Live tables: GPU hub · full price comparison · H100 page.

What the table doesn't show:

• Community vs. secure tiers: several of the cheapest rows are RunPod Community Cloud. Secure Cloud costs more (for example, H100 PCIe is $2.89/hr there).
• Marketplace prices move: Vast.ai figures are the median of live offers from verified hosts, not a fixed list price.
• Stale prices are held out: a rate that hasn't been re-verified for 7 days isn't ranked.
• Lowest hourly ≠ lowest cost per job: compare VRAM and speed for your workload. Confirm on the vendor page before you pay.

Use it from code

curl -s https://thelanternworks.com/api/v1/gpu-prices.json | jq '.rates | sort_by(.hourly_equiv_usd) | .[0:5]'

Daily history: /api/v1/gpu-prices/history.json. To see what moved since a date, use the cron-ready gpu-price-watcher example.

Use it from your agent

The MCP server (lantern-works-deals-mcp 0.3.11, MIT, no API key) is read-only and has 7 tools: compare_gpu_prices, get_price_changes, optimize_stack, optimize_agent_stack, search_deals, compare_deals, get_deal.

uvx lantern-works-deals-mcp
claude mcp add lantern-works-deals -- uvx lantern-works-deals-mcp

No install: connect any Streamable HTTP MCP client to https://lantern-works-mcp.lantern-works.workers.dev/mcp?ref=hashnode. It keeps only anonymous call counts per tool, day and ref tag.

Setup for Cursor, Cline, VS Code, Claude Desktop and others: thelanternworks.com/install.

Disclosure

Disclosure: some result links returned by the server are referral links. RunPod and Thunder Compute, which hold several of the cheapest rows above, are among them. Referral links are flagged in the data (affiliate: true), and ranking is by price only, never by affiliate status. The source links above go to the vendors' own pricing pages.

What's missing

Tell us which providers or GPUs to add: https://github.com/connormatthewdouglas/agent-work-stack-deals/issues

Disclosure: I run Lantern Works. Many listings (105 of 137) are affiliate/referral links; the rest are non-affiliate free tiers and vendor coupons. Affiliate status doesn't affect ordering.

Agents keep needing two boring things: search results and a picture of a web page. I compared the free options from vendor pricing pages. Anything not verified is marked as such.

Search / SERP APIs

Rule of thumb: raw SERP APIs (SerpApi, Bright Data) for rank tracking; AI-search APIs (Tavily, Exa) for agent retrieval.

Screenshot / headless browser APIs

Dedicated screenshot APIs are simpler; hosted browsers handle logins and multi-step flows.

For agents: the JSON feed

curl -s https://thelanternworks.com/deals.json | jq '.deals[] | select(.category=="search") | .title'
python3 query_deals.py --category unblock --max-price 20

No auth. Full tables: https://thelanternworks.com/compare/serp-api-comparison/ and https://thelanternworks.com/compare/screenshot-api-comparison/. Prices change, so check them at checkout.