September 2026 Paid API Test Report

September evidence snapshot

Paid test providers

6

Successful / total records

177 / 243

Exact model IDs / protocols

8 / 4

Latest recorded request

2026-09-29

These are rolling records from dedicated test accounts, not a uniform random availability sample. Each benchmark row represents one recorded run; a completed ten-request session is identified separately below.

Token totals are the fields returned by recorded runs and may include cache-read tokens. They are shown as evidence volume, not as a reconstructed bill or cross-provider efficiency ranking.

Latest recorded failures

A failure applies only to the exact model, protocol, key and time shown. It does not mean the provider's website or every model was unavailable.

Completed ten-request sessions

The current evidence set contains 7 completed sessions. An unavailable meter is not treated as zero charge. Same-route price reconciliation remains in the dedicated billing report.

Claude Opus 5 · API samples and published premium rates

28 requests across six test accounts: 28 scheduled, 0 manual, 0 other. These are not uniform random samples.

The published input-price range spans 5.6x across provider-labeled premium groups. The test log does not reliably identify the key's upstream group, so these request outcomes cannot be attributed to the displayed price group. A completed request does not prove model identity.

GPT-5.6 Sol · API samples and published premium rates

16 requests across six test accounts: 16 scheduled, 0 manual, 0 other. These are not uniform random samples.

The published input-price range spans 6.0x across provider-labeled premium groups. The test log does not reliably identify the key's upstream group, so these request outcomes cannot be attributed to the displayed price group. A completed request does not prove model identity.

Three findings from our own records

Coverage is model-specific. A provider with no GPT-5.6 Sol sample above cannot be described as GPT-verified for that model. The site-wide count of passing models is not a substitute for this row.

A reachable homepage and a completed API request answer different questions. The request table includes failures; website availability is tracked separately and is not used to fill missing model samples.

Rates are input and output prices for provider-selected premium groups, not observed wallet charges. Subscription full-use equivalents, cache discounts and promotional credit are excluded because plans and actual credit burn have not been reconciled across all six providers.

Method and limitations

Window: September 1–30, 2026 (China time), updated as new records arrive. Each row counts all recorded requests for that exact model, including manual and scheduled runs. P50 is the median end-to-end response time of successful requests; it is not time to first token. Samples differ by provider and time of day, and test-key groups are not fully recorded. This is an evidence ledger, not a controlled winner ranking.

Raw model-level logs are accessible from each provider profile and its evidence endpoint. We have not completed model-identity, cache, and before/after balance reconciliation for every route. Affiliate links and test credit, where present, are disclosed on provider pages; corrections: airelaycheck@163.com

Download privacy-reviewed request results (CSV) · Dated evidence changes (CSV) · Provider summary (JSON)