Skip to content

mariatomoko218/parity-test

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

8 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Token998 Parity Test

Open-source proof that token998.com does NOT silently swap models or downgrade quality.

License: MIT Uptime

Why this repo exists

The Chinese AI-API-relay industry has a documented trust problem: a few bad actors have been caught swapping gpt-4o requests for gpt-4o-mini, harvesting prompts, or routing through unreliable upstreams.

We open-source this test so anyone can audit us in 60 seconds.

How to verify us in 60 seconds

git clone https://github.com/token998/parity-test
cd parity-test
pip install -r requirements.txt

export OFFICIAL_OPENAI_KEY=sk-...     # your own official OpenAI key
export TOKEN998_KEY=sk-tk998-...      # your token998 key (get $1 free at token998.com)

python parity_test.py

You'll see a side-by-side table of identical-seed outputs from both endpoints. If we ever swap models, the test fails publicly.

Example output

============================================================================
TOKEN998 PARITY TEST RESULTS
============================================================================
| Model       | Test                  | Hash(Off) | Hash(Tk998) | Tokens | Latency       | Result |
|-------------|-----------------------|-----------|-------------|--------|---------------|--------|
| gpt-4o-mini | math_basic            | a3f9e2d1  | a3f9e2d1    | 18/18  | 0.82s/0.91s   | PASS   |
| gpt-4o-mini | code_basic            | 7c2b8e44  | 7c2b8e44    | 24/24  | 1.10s/1.18s   | PASS   |
| gpt-4o-mini | factual               | 4d5f6e7a  | 4d5f6e7a    | 12/12  | 0.65s/0.72s   | PASS   |
| gpt-4o      | math_basic            | a3f9e2d1  | a3f9e2d1    | 18/18  | 1.20s/1.30s   | PASS   |

Overall: ALL PASS
============================================================================

What we DON'T do

  • ❌ Log your prompts (we keep only billing metadata: tokens & timestamp)
  • ❌ Substitute cheaper models for premium ones
  • ❌ Re-route your traffic to questionable upstreams
  • ❌ Train on your data

What we DO

  • ✅ Direct upstream pass-through (OpenAI / Anthropic / DeepSeek / Qwen official endpoints)
  • ✅ Public status page: https://status.token998.com
  • ✅ 99.9% monthly uptime SLA (with automatic credit refund if breached — see SLA.md)
  • ✅ Refund unused credits within 7 days, no questions
  • ✅ USDT settlement (TRC20 / BEP20)

Pricing

3 RMB ≈ $1 official-equivalent credit (≈ 42% off official prices).

Model Official Token998 Savings
GPT-4o (1M input) $2.50 $1.05 58%
Claude 3.5 Sonnet (1M input) $3.00 $1.26 58%
DeepSeek V3 (1M input) $0.27 $0.11 59%
Qwen 2.5 Max (1M input) $0.45 $0.19 58%

See full pricing for all 500+ models.

Contributing

Found a test case where we fail parity? Please open an issue with the failing output. We commit to investigating within 24 hours and posting a public post-mortem within 48 hours.

Want to add a new test case? PRs welcome — see CONTRIBUTING.md.

License

MIT — fork, modify, run against any other AI relay you suspect.

Contact

About

Open-source proof that token998.com does not swap models or harvest prompts.

Topics

Resources

License

Contributing

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages