A data point is useful if you can trace where it came from.

01 / Three kinds of evidence

Specs and pricing: the vendor's public documentation. The links are kept on every product page and comparison. Rates may change after the review date.

External indexes: the LLM Stats, Reasoning and Coding columns reproduce a selection from the public LLM Stats snapshot. They are that publisher's indexes — not Codifly scores, accuracy percentages or tests we ran ourselves. See their original methodology. There is no declared affiliation with LLM Stats.

Legacy illustrative models: the CloudOps editorial pages keep 0–5 scorecards labeled as illustrative. They are not mixed into this catalog and are not used to rank AI models.

02 / Comparability before a winner

Each category has its own metrics. You can compare up to four products, all from the same category. A lower price may come with fewer resources, a different architecture or more operational work.

Indexes are not normalized or averaged here. We don't publish an overall winner. Highlighting in the table marks the best numeric value on that dimension; ties are not broken artificially.

In AI, OpenAI's total window is not the same as Gemini's input limit; the product page shows the original definition. In cloud, GB and GiB are different units and vCPUs don't measure performance. In databases, a per-account quota is not a per-project quota. In embeddings, dimensions and context are specs, not quality scores. In queues, retention and delay don't describe the same capability.

03 / Missing data

“No data” means we haven't included that value in this snapshot; it doesn't mean zero or a missing capability. It's left out of the corresponding chart and sorts last in both directions. You can limit a comparison to metrics available for every product.

There are no speed, latency, availability or in-house benchmark figures, because those measurements haven't been run under a common protocol.

04 / Pricing and simulation

AI: USD per million tokens on the standard API, uncached, short context. The estimate multiplies the combined input and output volume by each rate. It excludes tools, long-context surcharges, taxes and batch discounts. DeepSeek uses its uncached peak rate; its product pages explain the time windows. Models added without verifiable indexes show those columns as “No data”.

Cloud: monthly base price of the published plan. It is not TCO; support, backups, extra disk, networking and operations can cost extra. Databases: plan price and database space, with the scope stated on each product page.

Embeddings: text rate per million tokens, excluding starter credits and index storage. Object storage: marginal rate per decimal GB-month, plus GET/PUT per million; S3 and Spaces are converted from GiB by dividing by 1.073741824. Monthly minimums and transfer are on each product page.

Queues: a calculated scenario of one million 1 KiB messages to a single consumer, with no batching, retries or free tiers. SQS and Cloudflare count three operations per message; QStash counts one HTTP attempt. It excludes execution, base plans and networking. Comparing rates doesn't make delivery guarantees equivalent.

05 / Updates and exports

The catalog is published as a versioned edition. Every update has to verify the source, record its conditions and pass the publishing rules or an independent review. CSV exports include sources, date and conditions; unknown values are never exported as zeros.

Spotted a discrepancy? Check the vendor's page first, then send us a correction with the source.