ClaimedLast updated: 2026-08-18
What is the claim?
Every model price shown on harpd.com is a public list price verified against its official provider pricing page (last verified 2026-08-18).
What data supports it?
/data/llm-pricing.json — one record per model with source (official pricing page URL) and last_updated (verification date). CSV mirror at /data/llm-pricing.csv.
What methodology was used?
/methodology/model-pricing/ — source only official provider pages, attach a source URL and verification date per record, never guess unknown prices.
ClaimedLast updated: 2026-08-18
What is the claim?
The current JSON-extraction benchmark dataset (8 models × 7 metrics) was last updated 2026-08-18 and is a MODELED ESTIMATE, not a measured result.
What data supports it?
/data/benchmarks.json and /data/benchmarks.csv — long-format records (model, task, metric, score, cost, timestamp) exported from the benchmark data, which carries isModeled: true and its methodology note. Mirrored at github.com/harpd-dev/llm-cost-benchmark.
What methodology was used?
/methodology/benchmarks/ — real tasks, same prompt and seed per model, cost per successful task as the headline metric, open data and runner.
The numbers show how the live benchmark will be scored. We do not cite them as measured results.
ClaimedLast updated: Rolling — each record carries its own updatedAt timestamp; the dataset envelope exposes the most recent one as lastUpdated.
What is the claim?
A product's Rank Points equal the Rank Credit spend it completed (1 Credit = 1 permanent Rank Point), and Rank Points are promotional placement, not an editorial recommendation.
What data supports it?
/data/rank.json — rank and rankPoints per approved product, updatedAt per record; /data/methodology.json exports the exact rule.
What methodology was used?
/rank/methodology/ — Rank Points rule, boards, entry and disclosure, mirrored machine-readably in /data/methodology.json.
Not claimedLast updated: n/a
What is the claim?
That any model is "safe to switch to" for a production workload.
What data supports it?
No dataset yet — a switch decision requires a real-task quality gate on the specific workload.
What methodology was used?
/methodology/model-replacement/ defines the gate; until a workload is measured, no claim is made.
We publish the method, not a result.