Verified by TKBK
The models are genuine — we check them every day
In the API-key market, models are often swapped: a cheap model answers in place of Claude. So every day we give every model the same tasks with a known answer, through the same public address customers use (api.tkbk.online), and publish the results, speed and availability over 30 days.
last check 2026-10-26 00:36 UTC30-day accuracy: 98%4 tasks a day per model
| Model | Today | 30-day accuracy | Median response | 30-day uptime | Last 30 days |
|---|---|---|---|---|---|
| Claude Fable 5.1 Claude · claude-fable-5-1 | 4/4 | 97% | 2.9 s | 100.0% | |
| Claude Opus 5.5 Claude · claude-opus-5-5 | 4/4 | 97% | 3.2 s | 100.0% | |
| Claude Sonnet 5 Claude · claude-sonnet-5 | 4/4 | 100% | 2.5 s | 100.0% | |
| Claude Opus 5 Claude · claude-opus-5 | 4/4 | 98% | 2.6 s | 100.0% | |
| Claude Haiku 4.5 Claude · claude-haiku-4-5 | 4/4 | 98% | 1.4 s | 99.3% | |
| GPT-6.1 Sol GPT · gpt-6.1-sol | 4/4 | 100% | 3.5 s | 100.0% | |
| GPT-6 Astra GPT · gpt-6-astra | 4/4 | 100% | 2.8 s | 100.0% | |
| GPT-6 Sol GPT · gpt-6-sol | 4/4 | 100% | 2.2 s | 100.0% | |
| GPT-6 Luna GPT · gpt-6-luna | no answer | 100% | 2.8 s | 100.0% | |
| GPT-5.6 Sol GPT · gpt-5.6-sol | 4/4 | 100% | 2.2 s | 100.0% | |
| Gemini 3.8 Flash Gemini · gemini-3.8-flash | 4/4 | 100% | 4.4 s | 100.0% | |
| Gemini 3.1 Pro Gemini · gemini-3.1-pro | 3/3 | 100% | 6.3 s | 100.0% | |
| Grok 4.7 Grok · grok-4.7 | 3/4 | 93% | 12.9 s | 100.0% | |
| Grok 4.6 Grok · grok-4.6 | 4/4 | 100% | 16.7 s | 100.0% | |
| Grok 4.5 Grok · grok-4.5 | 4/4 | 97% | 15.8 s | 100.0% | |
| DeepSeek V4 Pro DeepSeek · GLM · deepseek-v4-pro | 4/4 | 100% | 3.7 s | 100.0% | |
| DeepSeek V4.1 Flash DeepSeek · GLM · deepseek-v4.1-flash | 4/4 | 100% | 2.5 s | 100.0% | |
| GLM-5.3 DeepSeek · GLM · glm-5.3 | 3/4 | 89% | 2.8 s | 100.0% |
How we check
- • Once a day, every model in the catalog, through the public api.tkbk.online — like any customer.
- • Tasks with a single correct answer: arithmetic, logic, strict JSON. A genuine flagship model solves them almost every time; a small substitute often gets them wrong.
- • Time is the median of a full answer to a short request (for reasoning models it includes the thinking time).
- • Uptime is the share of 5-minute probes in which the model had at least one working channel.
- • Before it is connected, every new channel goes through an extended authenticity test (hidden model fingerprints and trap tasks) and a measurement of its real cost.
Tasks in today’s check
- 1. What is 487 * 23? Answer with the number only, no explanation.
- 2. Today is Wednesday. What day of the week will it be in 10 days? Answer in one word.
- 3. What is the sum of the digits of the number 2026? Answer with the number only.
- 4. Return strictly JSON, with no explanation and no markdown: {"ok": true, "n": 7}
The models receive these prompts in Russian; they are translated here.
The data is open (CC BY 4.0): download the CSV of the last 30 days, or cite the results with a link to tkbk.online/en/verified.
The same models — on your key
Claude, GPT, Gemini, DeepSeek and Grok on one key — for Claude Code, Cursor, Codex and any SDK. Pay by a Russian bank card, YooMoney or SberPay on the site, or with crypto (USDT) in our Telegram bot. The balance never expires.