Editorial standard

Methodology

compute.info is a data publication. Accuracy, provenance, and useful context take priority over catalog size or update theater.

Source hierarchy

1. Tier A: official machine-readable catalogs, APIs, model cards, specifications, or benchmark results.
2. Tier B: authoritative vendor documentation and pricing pages without stable structured data.
3. Tier C: reputable independent measurements and publications.

Forums, social posts, and comparison sites can trigger investigation but are not canonical evidence.

Normalization

AI prices retain their original decimal amount, currency and billing unit; we do not convert AI currencies. Input, output, cache reads and cache writes are distinct. Prompt brackets, regions and endpoints, service tiers and modality-specific rates remain separate. Unknown amounts are not free and unknown billing dimensions are not standard-tier assumptions.

VM prices remain in each vendor's native currency and an identified representative region or published price scope. A monthly value prefixed with “≈” is the hourly list price multiplied by 730 hours; it is not a vendor billing cap. The catalog generally favors meaningful entry, balanced, compute, memory, storage, Arm, GPU, and accelerator representatives over every near-duplicate size. For comprehensively ingested provider APIs, all listed SKUs are retained: regionally unavailable SKUs are labeled as such and receive no fabricated price. We preserve whether compute is shared, burstable, or dedicated and whether storage, outbound transfer, public IPv4, and tax are included or explicitly mark them as unspecified.

Provenance contract

Each record includes a stable identifier, source URL and type, retrieval time, effective date when known, region, currency, confidence grade, notes, and publication status. Source links are visible beside the data they support.

AI access and model identity

The AI directory keeps API pricing and downloadable model records separate. API delivery alone never establishes closed weights. Open, closed or unknown labels require explicit model-specific evidence; without reviewed evidence, status is unknown. Publisher aliases organize discovery but do not join models by name. Download availability remains subject to license and any access gates.

Developers, checkpoint records and selling API providers have distinct identities. A hosted checkpoint or base/post-training lineage link requires explicit evidence, not a similar name. A repository identifier does not prove an immutable revision. Total weights and active mixture-of-experts parameters are different counts; unknown specifications remain unknown.

Directory cards show the API seller separately from the checkpoint developer and lineage. An offer without a verified checkpoint link retains an unknown developer. Price observations and completed price-source checks come from rate evidence; model-metadata observations and weight-license reviews keep their own dates. Missing dates stay unknown. Readable pricing, creator and license sources appear on the card, with raw API evidence and detailed claims in the closed source disclosure.

Name order groups records only when they share an evidence-linked checkpoint; price order keeps individual listings in rate order. A failed refresh or a candidate held for review is labeled explicitly when recorded. Last-known observations retain their original dates and stale status; a missing result is not treated as a successful check.

Price sorts are not capability or like-for-like offer rankings. Only one unit-aligned USD token rate per direction supplies a price-sort key; foreign currency, modality-specific or multiple-tier variants remain visible but unranked. Missing/unranked values sort last in both directions. All original rate conditions remain visible. A downloadable record without an API price is not a claim that no API exists.

API availability, download status, gating, weight license and code license have independently cited observations. We do not infer a model's license from its provider. Open weights do not prove ungated access or open-source code. Retrieval, completed check, completed review and review deadline are distinct dates. An expired deadline can display as stale on rebuild, but a rebuild is not a source check and never refreshes an observation. Missing or stale evidence does not establish retirement: an explicit retirement date needs its own source. Historical rates may remain listed.

Model and hardware estimates

Open-model specifications come from creator-maintained Hugging Face model cards. “Open-weight” means weights are downloadable under the stated license; it does not automatically mean an OSI-approved open-source license.

Hardware-fit estimates use decimal gigabytes and a conservative planning allowance for quantized formats: 5 bits per weight for Q4 and 9 for Q8, followed by 20% loading overhead. This allows for scales, block metadata, and mixed-precision tensors. Published artifact sizes should replace the estimate when a specific GGUF or safetensors build is selected.

Fit labels cover estimated model weights only. KV cache, context length, activations, runtime scratch space, drivers, display reservation, backend compatibility, and CPU/GPU offload can materially change capacity and speed.

Validation and corrections

Automated checks reject duplicate identifiers, invalid URLs, negative prices, malformed currency codes and timestamps. Material changes receive review before publication.