24Test Model Check · Live
24Test Model Check
Enter any Base URL, API Key, and model name to receive an evidence report in minutes: does the protocol conform, does behavior resemble the official endpoint, and is token usage inflated? No account is required. Your Key is used only for this run and is fully redacted in reports, progress streams, and the database.
Messages protocol, tool calls, thinking signatures, and cache behavior.
What we check
Four endpoint types. One fixed probe suite.
Probe request bodies are generated locally by the platform; callers cannot supply prompts. Each image-probe run injects random colors, counts, and other parameters into fixed templates, preventing upstream services from recognizing probes for selective routing or responding with stored images. Cache prefixes, multi-turn passphrases, visual colors, and PDF reference numbers in text probes are also randomized for each run. The remaining capability questions and protocol checks use fixed text.
Claude (Anthropic Messages)
- Model-name consistency, streaming versus non-streaming comparisons, tool calls, and structured output
- Thinking-signature replay and forged-signature rejection, Prefill boundaries, cache writes and reads
- Token-usage ledger: input usage from three minimal requests compared against an empirical 64-token budget (currently named “Prompt and cache count check” in reports; the name will be unified with the official-site launch)
- Six capability questions and fixed performance samples; optional Bedrock-specific checks
OpenAI-compatible (Chat Completions + Responses)
- Model lists, system prompts, max_tokens, stop, multi-turn conversations, parameters, n, and logprobs
- Streaming and streaming usage, tool calls and parallel tools, JSON mode, and JSON Schema
- Responses API basics, streaming, tools, structured output, and previous_id
- Error-body shape, usage fields, model consistency, and performance; basic / standard / full tiers
Native Gemini
- generateContent, SSE streaming, and countTokens
- System instructions, candidate count, thinking, tool calls, JSON mode, and vision
- Error bodies, usage, and model consistency; basic / standard / full tiers
Image generation and editing (OpenAI Images)
- Protocol checks for dimensions, formats, transparent backgrounds, n, streaming chunks, and more
- Local pixel validation of objective content: solid colors, regions, centered shapes, and counts
- Optional OpenAI Verify authenticity comparison; generated images are not stored in the database, with only thumbnails retained in the progress stream
Reports and recurring regression checks
One check provides evidence. Scheduled checks close the loop.
Shareable reports
Every report has its own page and random link, accessible to anyone who has the link and excluded from search-engine indexing. JSON and HTML downloads, printing, and PDF export are supported, including image reports.
Optional accounts
Username and password only; no email address is collected. After sign-in, “My check history” lists only checks initiated by your account, up to the latest 1,000. You can also run checks without signing in; keep the report link.
Encrypted storage for test Keys
Keys are stored on the server with AES-256-GCM encryption; the master key stays outside the database and backups. You can set an expiration date for automatic deletion. Once saved, even you can only see a masked value.
Scheduled checks and one-click reruns
Rerun your saved check options on the server at intervals from 30 minutes to 24 hours, even after closing the page. One-click reruns reuse the same options, making score drops easy to spot.
Comparison baselines
Anyone can promote a completed Claude report to a shared comparison baseline for later runs. Four reference runs against official endpoints have yet to be added (planned).
Security boundaries
Every outbound connection validates the IP after DNS resolution. Probes are never sent to private-network, loopback, or metadata addresses. No proxies or redirects are used, and each IP can run only one check at a time.
Relay sites and site owners
Add a check entry point to your site
Two integrations that need no backend: prefilled links take users directly to the check form, while badges display the latest result under your domain.
Prefilled links
Place type, base_url, and model after # in the URL fragment, without sending them to the server; users enter their own Key.
https://check.24test.com/#type=openai&base_url=https://api.example.com&model=gpt-5.6Site badges
Displays the conclusion, score, and date of the latest check for that domain, expiring after 30 days. Script data is issued only to pages under that domain and cannot be loaded by other sites. When other sites embed the SVG image, it shows only a gray “Not issued to this site” label. The badge always identifies the domain and links to the public results page.
<script async src="https://check.24test.com/embed/example.com.js"></script>
A badge describes only one check at one point in time and shows the actual score, whether favorable or unfavorable. For access, see 24test.com/get-badge; for integration instructions, see 24test.com/integrate.
Limitations
What a check can and cannot establish
How the products connect
After the check
Procurement
When a report endpoint is hosted by TBuy, it will be labeled “This channel is provided by TBuy (affiliated brand)” (planned; disclosure only, with no traffic referrals). Models and tokens are purchased through tbuy.ai .
Enterprise Access
For a unified enterprise entry point, see Token Router Enterprise: employees keep their existing tools, while the enterprise manages Keys and quotas.
Keep an evidence trail
After integration, each call can be recorded in Token Audit; audit findings can lead back here for another check.