Flops skill

FLOPS Compute Prices

Look up GPU / compute rental reference rates and cite them with a verifiable source, using the public FLOPS Index API. Everything here is key-free.

When to use this skill

• The user asks what an accelerator costs to rent — "what's an H100 going for?", "cheapest A100 right now?", "on-demand H200 price?".
• The user wants to compare rates across GPUs or across market types (spot / on-demand / DePIN).
• The user states a compute price and wants it fact-checked, or wants a citable price with a link a reader can verify.

What FLOPS is (and is NOT)

FLOPS publishes indicative reference levels for GPU compute, snapped to the most recent 6-hour UTC mark and refreshed every 6 hours. Values are source-opaque (no provider identities) and delayed on the public surface. A value is not a live provider quote, not a settlement mark, and not investment advice. Always carry that caveat when you cite one. Never fabricate a number — if a call fails or returns null, say so.

Index slugs

FLOPS-<ACCELERATOR>-<MARKET> where <MARKET> ∈ SPOT, OD (on-demand), DEPIN (decentralized). E.g. FLOPS-H100-SPOT, FLOPS-A100-OD, FLOPS-A100-DEPIN. Don't guess the accelerator list — use the catalog.

Which endpoint returns what

• /v1/price/{slug} — the authoritative single-index record: full envelope + citation/verify URLs. Use this for any value you will cite.
• /v2/catalog/public — every index with its value in one call; best for scanning/comparing a family. ⚠️ For a given slug, catalog may report a slightly different confidence / change_24h than /v1/price (a known backend inconsistency). When they disagree, /v1/price wins — always confirm a citable value with /v1/price.
• /v1/search?q= — returns matching slugs + citation URLs only, NO value. Use it to discover slugs, then call price.

How to use it — price → verify → cite

1. Get a price (authoritative)

GET https://app.flopsindex.com/v1/price/{slug} — helper: python scripts/flops.py price FLOPS-H100-SPOT Envelope: value, unit ("USD/GPU-hr"), as_of, delayed, confidence ("HIGH" | "MED" | "LOW" — an ordinal label, never a number), change_24h ("UP" | "FLAT" | "DOWN" | null), data_tier (e.g. "LIVE" = coverage tier, not a freshness claim — the value is still delayed), verify_url, citation_url, methodology_url, disclaimer. (Other fields like permalink, upgrade may appear; list is not exhaustive.)

as_of is a 6-hour UTC bucket label (00/06/12/18 UTC). The current bucket can read slightly ahead of your wall clock — that's normal bucketing, not stale/future data. Present it as UTC.

2. Cheapest for a given chip

Fastest path: python scripts/flops.py cheapest a100 — fetches /v1/price for each of that chip's SPOT/OD/DEPIN slugs and returns the lowest (authoritative), with market, an all_markets breakdown, and the verify link. Or read /v2/catalog/public, filter FLOPS-<CHIP>-*, take the min value, then confirm the winner with /v1/price.

Two things to get right when reporting the cheapest:

• Flag the market class. The family-min is very often a DEPIN (decentralized) rate — a different reliability/risk class than on-demand or spot. Do not present a DePIN number as an apples-to-apples cloud rate; say it's decentralized, and offer the cheapest conventional (SPOT/OD) rate alongside. (cheapest emits a market_class_note when the winner is DePIN.)
• Cite from a fresh /v1/price. The 6-hour as_of bucket can roll between calls, so the envelope inside cheapest may carry an older bucket than a /v1/price you make seconds later. Re-pull /v1/price/{winner} at cite time for the timestamp you print.

3. Verify / fact-check a claimed value

GET https://app.flopsindex.com/v1/verify?index_id={slug}&value={n} — helper: python scripts/flops.py verify FLOPS-H100-SPOT 1.50 Returns verified, actual_value, delta_pct, tolerance_pct, tolerance_abs.

Interpreting the result — read this before you narrate a verdict:

• delta_pct = (submitted − actual) / actual × 100. Negative ⇒ the user's number is below the reference; positive ⇒ above. State the direction explicitly.
• tolerance_pct is tiny (±0.5%). Verify is an exact-match check against the published mark, NOT a plausibility/fairness test. So verified:false means "doesn't match the reference to ~2 decimals," not "the user is wrong." Never tell a user their estimate is "wrong" — say it "differs from the FLOPS reference by X% (below/above)".
• If a value fails for the market the user named, check the sibling markets (SPOT/OD/DEPIN) before concluding — the gap is often a market mismatch (e.g. a DePIN listing vs a spot rate), and saying so is the fair answer.

4. Cite it

Always present the value WITH its as_of, the delayed-reference caveat, and the verify + citation URLs. The API's raw verify_url has no value attached — append &value={value} to make it checkable (the price and cheapest helpers already emit a ready verify_url_checkable). Canonical one-liner:

$2.33/GPU-hr — H100 spot reference (FLOPS-H100-SPOT) as of 2026-07-16 12:00 UTC; delayed indicative level, not a live quote. Verify: https://app.flopsindex.com/v1/verify?index_id=FLOPS-H100-SPOT&value=2.33 · Source: https://app.flopsindex.com/i/FLOPS-H100-SPOT

Notes

• Key-free. A FLOPS_API_KEY (header X-FLOPS-Api-Key) upgrades the same calls to real-time full precision, but is never required.
• change_24h is often null — handle it; not an error.
• Same data is available as MCP tools (https://app.flopsindex.com/mcp) and SDKs: pip install flopsindex, npm i @flopsindex/sdk.