AI models, closer to your application.

We queried the public catalogue on 6 September 2026: 107 models, all marked live. The GPT-5 family is broadly covered, from nano to pro, including the codex variants and the four 5.6 versions; there are o1/o3/o4-mini, the full Gemini line (2.0, 2.5, 3, 3.1 flash and flash-lite), Gemma 3 and 4, image (gpt-image-1/1.5/2, flux-2-pro, gemini flash image, mai-image-25), video (sora-2), audio and realtime (gpt-realtime up to 2.1, transcription, elevenlabs-tts in three variants), embeddings and deepseek-v4-flash. One thing we need to say directly: in the live catalogue there is no Anthropic Claude model, although our internal README and presentation pages still promise "Anthropic Claude direct". Llama, Qwen, Mistral, Voyage and Cohere are also missing, although they appear in our document for crawlers. The public catalogue is the source of truth; those texts are out of date.

The provider’s actual cost plus 70%, rounded up to the micro-cent. The same percentage across all five plans — there is no preferential token rate on the expensive plans. The rule is an execution-time checked invariant: if any code path tried to go below 70%, it throws an error instead of issuing an invoice. The paid plan does not make your tokens cheaper; it gives you a monthly credit equal to the subscription, e-Invoice on Business and Enterprise, and support tiers.

The surface is OpenAI-compatible, served by a LiteLLM proxy at https://api.longday.io/v1. In practice: in Node you change baseURL in the OpenAI constructor, in Python base_url, and the rest of the code remains identical. The in-house documentation lists as tested OpenAI Node, OpenAI Python, LangChain JS, LangChain Python, LlamaIndex and Vercel AI SDK. It also works from no-code tools that can send a Bearer header.

The Free plan is capped at 10 requests per minute; the other four plans do not have a plan cap defined in code and are limited by the key budget. Top-ups accept between 5 and 250 dollars per transaction. Each key can have its own budget with a cycle of 1, 7, 30 or 90 days and a model allowlist. Some models have their own upstream limits noted in the catalogue — gpt-image-2 at 10 requests per minute, gpt-realtime-2.1 at a quota of 15 TPM.

A document, not a measurement — and it is important to know the difference before relying on the figure. Our SLA page publishes five tiers, from 99,0% on Free to 99,99% on Enterprise, and says that uptime is calculated monthly from status.longday.io. That subdomain does not exist: it has no DNS record and does not respond. In code there is no availability collector — neither a metrics endpoint nor external monitoring; both are unstarted tasks in our own backlog. The status page that does exist makes three real-time probes and does not keep history, so it cannot produce a monthly percentage. The 99,9% figure shown on the homepage is a written contractual commitment, not an observation, and the product self-labels as BETA in the header of the same page. Until the monitor exists, treat the percentage as a contract clause with compensation credit, not as proven performance.

Here we need to be exact, because there is a difference between what is designed and what is built. Currency conversion is real and works: the official NBM rate is read from the bank’s XML for the respective day, with a local fallback if the service does not respond within five seconds. The client that submits the invoice to the Fiscal Service is, however, still a skeleton — a wrapper over fetch, with its own comment saying that the real implementation, the one that loads the company’s PKCS#12 certificate and signs each request, is still to come. Until then the e-Invoice part is handled with accounting, not automatically from the platform.

No. It is an access and billing layer over other companies’ models. It should not be confused with Grai, which is our Romanian voice research project and has no technical connection with the gateway.