Synthetic set of five request classes, no external incident window. One day is a measurement of that day, not a promise. Method, limits and the Super Deal (economical tariff, depends on the Gonka network, no insurance outside Gonka) figures: benchmarks.
SAGG is a gateway that sits between your code and several independent LLM providers, automatically switching to a backup the instant one stumbles — so your requests keep working, not just the API you happen to be pointed at. Get a key in one click, drop it into your existing OpenAI client, start sending requests. No code changes beyond the URL.
Measured on version R1 over 24 hours (2026-10-23 to 2026-10-24, synthetic set of five request classes): Standard served 99.94% of requests (4,769 of 4,772), median time to first text 1.4 s, 95th percentile 7.7 s. Super Deal (economical tariff, depends on the Gonka network, no insurance outside Gonka) served 98.82%, 95th percentile 6.9 s. Billing matched delivery with no discrepancy. One day is a measurement of that day, not a promise. Method, data and limits: benchmarks.
Methodology and complete raw data (request-level JSONL, reproduction scripts, every failure explained): github.com/privatedeskai/sagg-benchmark-data. Full breakdown by line and workload: /benchmarks.
DeepSeek-V4-Flash-0731 — 400,000 token context window, 16,000 token max output. Open-weight (MIT-licensed) — part of why we can offer this at a fraction of centralized-provider pricing. We make no claim here about the model's own quality; what we measured is on the benchmarks page.
SAGG is a gateway for accessing language models through a familiar OpenAI-compatible API. Instead of a direct contract with a single cloud provider, you go through us — and we decide which of several independent inference backends handles your request, based on their live availability.
If one backend is unreachable or errors out, your request automatically retries on the next one — no action needed on your end, no dropped connection. You pay for tokens actually used, not a subscription.
Failover, not a single point of failure. Requests go through up to three independent inference providers in a fixed priority order — not random distribution: while the first provider is healthy, it gets 100% of traffic; on failure, an immediate switch to the next one in line.
One time budget for the whole request, not per attempt. If the first provider is slow or unreachable, the next attempt still fits within a single wait limit — you don't get doubled latency from retries.
Prepaid, no billing surprises. Top up your balance in advance, spend as you go. Running low triggers an advance warning, not a surprise invoice.
Input: $0.03971 / 1M · Output: $0.07941 / 1M
Short answers. “Served” means the answer was delivered, not that it was clean. Providers of the Gonka network require a minimum of 64 generated tokens on every request (a known behaviour of the network, gonka-ai/gonka issue 1955), so on very short requests (a one-word answer, a short JSON) they append repeated text. In our test of 2026-10-27, 17% to 23% of a mixed set of 60 requests were affected when the Gonka providers were called directly (15% and 18% through our two lines), up to 70% of one-word requests and up to 50% of short JSON requests; in the R1 24-hour test, 3.8% to 5.4% of the short answers from the Gonka providers ended at the output limit. The provider outside the Gonka network did not show this (0 of 243 in R1, 0 of 60 on 2026-10-27). The network has published a fix (pull request 1972); when it will be switched on is not known. On Standard, the provider outside the Gonka network is used only when the earlier providers fail, so the repeated text is not avoided on this line. Details: benchmarks.
Example request:
curl https://api.privatedeskai.com/v1/chat/completions \
-H "Authorization: Bearer YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"deepseek-ai/DeepSeek-V4-Flash-0731","messages":[{"role":"user","content":"hi"}]}'Already have a Super Deal key on a funded account? Request an additional key for Standard directly via POST /internal/issue-key (omit line, or {"line": ""}) with your existing Bearer token — same balance, no separate signup.
Gonka-network-only routing, no non-Gonka backup — the lowest price we offer, with a real tradeoff explained below.
Short answers. “Served” means the answer was delivered, not that it was clean. Providers of the Gonka network require a minimum of 64 generated tokens on every request (a known behaviour of the network, gonka-ai/gonka issue 1955), so on very short requests (a one-word answer, a short JSON) they append repeated text. In our test of 2026-10-27, 17% to 23% of a mixed set of 60 requests were affected when the Gonka providers were called directly (15% and 18% through our two lines), up to 70% of one-word requests and up to 50% of short JSON requests; in the R1 24-hour test, 3.8% to 5.4% of the short answers from the Gonka providers ended at the output limit. The provider outside the Gonka network did not show this (0 of 243 in R1, 0 of 60 on 2026-10-27). The network has published a fix (pull request 1972); when it will be switched on is not known. Super Deal has no provider outside the Gonka network, so all three of its providers are affected. Details: benchmarks.
Honest tradeoff, not hidden: all three of this line's providers are Gonka-network brokers. If the underlying Gonka network itself has an outage (not just one broker), all three can fail at the same time — this line does not have Standard's independent non-Gonka fallback. That's the real reason the price is lower.
Example request — same shape as Standard, your key alone decides the line/routing/price, not the request:
curl https://api.privatedeskai.com/v1/chat/completions \
-H "Authorization: Bearer YOUR_SUPER_DEAL_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"deepseek-ai/DeepSeek-V4-Flash-0731","messages":[{"role":"user","content":"hi"}]}'Already have a Standard key on a funded account? Request an additional key for this line directly via POST /internal/issue-key ({"line": "Super Deal"}) with your existing Bearer token — same balance, no separate signup.
api.privatedeskai.com/v1 — nothing else in your code changes.GET /internal/balance.Model: DeepSeek-V4-Flash-0731 — a 284B-parameter Mixture-of-Experts model (13B active per token), 400,000 token context window, 16,000 token max output. Open-weight under the MIT license. The quality of its answers is the model's own and is not a claim of this service.
Failover, in detail: up to 3 independent inference providers in strict priority order (not round-robin). The first healthy provider handles 100% of traffic; on failure, the request moves to the next one within a single shared 45-second time budget for the entire non-streaming request, not a fresh timeout per attempt. On both lines the first two providers are on the Gonka network; the third is a provider outside the Gonka network on Standard (insurance), and a Gonka-network provider on Super Deal. Streaming requests use a different, per-attempt timeout instead, so a legitimately long stream is never cut short by a fixed total budget.
Shared infrastructure, disclosed: Standard and Super Deal share the same first two providers, both on the Gonka network — an outage there can affect both lines at the same time, even though the two lines are measured separately (see Status and Benchmarks). Standard has a third provider outside the Gonka network; Super Deal has none, and its pricing-card details cover that tradeoff.
API compatibility: drop-in OpenAI-compatible endpoint — change one line in your existing client.
curl https://api.privatedeskai.com/v1/chat/completions \
-H "Authorization: Bearer YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"deepseek-ai/DeepSeek-V4-Flash-0731","messages":[{"role":"user","content":"hi"}]}'Privacy: we do not store the text of your requests or of the answers. See the Privacy Policy for what we keep and who processes your requests; we give no zero-retention or “no training” promise for the Gonka network path.