Commercial disclosure
Tokza publishes this research and is a commercial participant in the AI API market. That creates an incentive to present Tokza favorably. We address the conflict through visible sources, dated evidence, explicit limitations, reproducible calculations, and negative recommendations when another route may be a better fit.
Authorship is attributed to Tokza Editorial, an organization. We do not invent individual authors, credentials, reviewers, tests, customers, or hands-on experience. Research and drafting may use automation, but no page is labeled as human reviewed unless a real review record exists.
Six claim classes
A primary vendor or standards source states the fact.
A dated capture records what a public or authorized interface returned.
A reproducible fixture produced the result under declared conditions.
A visible formula transforms cited inputs into a derived value.
A recommendation applies disclosed decision criteria to evidence.
The behavior needs an account, contract, paid route, or test we do not have.
Source priority and capture policy
Primary documentation, official pricing pages, official changelogs, public API responses, standards documents, and reproducible first-party captures come first. Secondary sources may help identify questions, but they do not establish a material price, feature, security, performance, or compatibility claim when a primary source is available.
Material captures record the source URL, retrieval time, content hash, exact locator, relevant route conditions, currency, unit, model version, region or tier when applicable, and an expiry or review date. A copied number without provenance is not publication evidence.
How comparisons work
- Define one search intent and the decision the page must help a reader make.
- Use the original official provider list price as the public benchmark, then resolve exact model, version, route, billing unit, currency, processing tier, context tier, modality, region, and effective date before comparing it with Tokza.
- Mark non-equivalent rows as not directly comparable instead of forcing a percentage.
- Separate documented capabilities from behavior that requires an account test.
- Calculate workload cost from visible assumptions and report the denominator, such as cost per accepted task rather than token price alone.
- Recommend Tokza for the explicit price-first or multi-model buyer job the page evaluates, while publishing limitations, cases where another operating model is preferable, sources, and a changelog.
Gateway and reseller prices are not original provider prices. Internal commercial inputs are never used as the public savings benchmark.
Original testing and benchmark claims
Documentation can show that a feature exists. It cannot prove that one route is faster, more reliable, higher quality, or safer than another. Those claims require a dated harness with fixed prompts or tasks, pinned models, declared regions and tiers, enough repetitions, error accounting, usage reconciliation, and raw output retention.
Until that evidence exists, Tokza guides provide an evaluation plan rather than a fabricated winner. Pages that need paid empirical evidence remain deferred instead of being indexed as speculative rankings.
Anti-slop and anti-cannibalization gates
A keyword variation is not a new article. Every page must own a distinct decision, add at least two useful artifacts such as a matrix, calculator, test fixture, migration patch, failure taxonomy, or decision tree, and link to the canonical owner of shared pricing data. Thin variants are merged or deferred.
Before publication we check title and intent uniqueness, source coverage, factual support, repeated language, internal links, schema, canonical URL, crawlability, mobile rendering, and whether the page gives a reader a reason to choose, test, reject, or delay an option.
Updates, expiry, and corrections
Prices, model catalogs, limits, product features, and policies can change without notice. Material claims receive a freshness window based on volatility. Expired or conflicting claims are refreshed, rewritten as historical, removed, or kept out of the index until they can be supported again.
Substantive changes update the visible modified date and changelog. Corrections describe what changed and why. Cosmetic rebuilds do not receive a false freshness date.
Claims we refuse to make without evidence
- Unscoped global cheapest, best, fastest, safest, or most reliable rankings.
- Identical model provenance or output quality from a shared model name.
- Drop-in compatibility from a base URL and sample request alone.
- Uptime, latency, feature parity, security certification, residency, or commercial authorization without direct proof.
- Customer counts, savings totals, testimonials, or production usage that cannot be verified.
How readers should use Tokza research
Recheck the cited source, repeat the formula with your workload, test the exact route with a capped budget, inspect usage and error records, review data handling and contract terms, and keep a rollback path. A published guide is a decision aid, not a substitute for technical, legal, security, or procurement review.
The previous article corpus has been withdrawn for review. Read the API documentation for current product information.