A - Private
SOVAPI STRUCTURE
Live measured from this run.
S
SOVINFRA | PRIVATE AI
Private modeless inference · EU-hosted.
Sovereign AI stack
August 2026
Compare frontier AI. Measured live.
Run the same prompt. Compare latency, throughput, quality and real request cost.
Qwen3-Coder-Next FP8
Code and technical generation
Gemma 4 31B FP8
Text, vision and extraction
Whisper Large V3
FR / EN audio coverage
BGE-M3
Embeddings and semantic search
Kokoro
FR / EN speech synthesis
B - Challenger
Anthropic — Claude Sonnet 5
Live measured from this run.
The live verdict appears after both measured runs complete.
Structure accuracy
Neutral - run the reference prompt to score accuracy.
Reference target: 6x faster TTFT
SovAPI TTFT
0.12 sTarget path for code and technical generation.
Code throughput
184 tok/sReference throughput on the SovAPI Code lane.
Reasoning throughput
234 tok/sReference throughput on the SovAPI Text lane (analysis, reasoning, structure).
Latency reference table
Reference / expected values. The Arena verdict above is recalculated from each live duel.
| Scenario | SovAPI route | SovAPI reference | Challenger reference | Signal |
|---|---|---|---|---|
| TTFT | All routes | 0.12 – 0.18 s | 0.74 – 1.11 s | ~6x faster |
| Code throughput | SovAPI Code | ~184 tok/s | ~84 – 96 tok/s | ~2x more tokens/s |
| Analysis throughput | SovAPI Text | ~234 tok/s | ~87 – 100 tok/s | ~2.4x more tokens/s |
| Reasoning throughput | SovAPI Text | ~234 tok/s | ~74 – 96 tok/s | ~2.5x more tokens/s |
| Structure / Vision throughput | SovAPI Text | ~234 tok/s | ~87 – 100 tok/s | ~2.4x more tokens/s |
| Tool calling / Function call | SovAPI Code | ~184 tok/s | ~84 – 96 tok/s | ~2x more tokens/s |
| Speech-to-text | SovAPI Voice | ~391 ms | ~1.0 – 1.4 s | ~3x faster |
| Embeddings + vector search | SovAPI Memory | ~189 ms | ~350 – 500 ms | ~2x faster |
| End-to-end streaming | Modeless SovAPI | SSE direct | Official provider API | Measured live only in Arena |
Engine. Four SovAPI products routed automatically to the model best suited to your task: SovAPI Code (code + tool calling), SovAPI Text (analysis, reasoning, structure, vision), SovAPI Voice (speech-to-text), SovAPI Memory (embeddings + long-term memory). EU sovereign hosting, open-weight Apache 2.0. Powered by Qwen3-Coder-Next, Gemma 4 31B, Whisper Large V3, BGE-M3.
ANTHROPIC-CLASS QUALITY · 60x cheaper than Claude Sonnet 5 · IN EUROPE
SovAPI vs Claude Sonnet 5
Verified benchmark values. Fast-inference mode (Non-Think) for developer and interactive workloads. Live Arena results remain measured separately for each prompt.
| Benchmark | SovAPI | Sonnet 5 | Gap |
|---|---|---|---|
| GPQA Diamond | 87% | ~83% | +4 |
| AIME 2025 | 91% | ~78% | +13 |
| MMLU Pro | 86% | ~86% | = |
| LiveCodeBench Non-Think | 77.1% | 64.8% | +12.3 pts |
| SWE-bench Verified | ~70% | ~72% | -2 |
| BFCL v2 | ~89% | ~88% | +1 |
SovAPI vs DeepSeek V4 (Flash & Pro)
Verified benchmark values. Fast-inference mode (Non-Think) for both V4-Flash (13B activated) and V4-Pro (49B activated). Gap measured vs V4-Pro.
| Benchmark | Qwen3 | Gemma 4 | SovAPI | V4-Flash | V4-Pro | Gap |
|---|---|---|---|---|---|---|
| LiveCodeBench Non-Think | 71.4% | 80.0% | 77.1% | 55.2% | 56.8% | +20.3 pts |
Usage-based pricing in US dollars, with no commitment
Gemma 4 31B
$0.12 · $0.38Per million input and output tokens. No cache-read tariff.
Qwen3-Coder-Next
$0.12 · $0.90Per million input and output tokens.
Qwen cache read
$0.04 cachedPer million cached input tokens.
Trial grant
1 BTrial tokens for developers starting from API ACCESS.
Economy reference table
Reference / expected values. SovAPI cost is recalculated from live token usage. Challenger cost remains unavailable when the official API response does not expose a monetary amount.
| Metric | SovAPI | Challenger | How it is shown |
|---|---|---|---|
| Gemma 4 31B | $0.12 / M input · $0.38 / M output | -- when unavailable | No cache-read tariff |
| Qwen3-Coder-Next | $0.12 / M input · $0.90 / M output | -- when unavailable | Model-specific live cost |
| Qwen cache read | $0.04 cached | -- when unavailable | Never inferred from uncached input |
| Per-request cost | Calculated from real tokens | -- when unavailable | Only a server-reported amount is displayed |
| Trial grant | 1 B tokens | None | API ACCESS onboarding |
| Rate limit | 1,000 req/min free; 5,000 req/min PAYG; 1,200 tok/s/key | Provider-specific | Applied server-side per key |
| Billing | Exclusively token usage | Provider-specific | No SovAPI commitment |
Step 1
Create your key
Use GitHub or Google OAuth, then generate your trial SovAPI key.
Create my API keyStep 2
Change two lines
Keep your OpenAI SDK, replace only base_url and api_key.
Step 3
Call sovapi
One public model, automatic routing by modality.
Plans and pricing
| Tier | Model / service | Rate limit | Input | Output | Cache read | Support | Retention |
|---|---|---|---|---|---|---|---|
| Free | 1 B tokens · all models | 1,000 req/min · 1,200 tok/s/key | Included | Included | Included | Community | None |
| Pay-as-you-go | Gemma 4 31B | 5,000 req/min · 1,200 tok/s/key | $0.12 / M tokens | $0.38 / M tokens | Not billed separately | Email within 24h | None |
| Pay-as-you-go | Qwen3-Coder-Next | 5,000 req/min · 1,200 tok/s/key | $0.12 / M tokens | $0.90 / M tokens | $0.04 cached | Email within 24h | None |
Whisper, BGE-M3 and Kokoro are included services and are not token-billed.
curl https://api.sovinfra.ai/v1/chat/completions \
-H "Authorization: Bearer $SOVINFRA_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"sovapi","messages":[{"role":"user","content":"Hello"}]}'
Use sovapi for automatic routing, or pin a generation model directly:
gemma-4-31b for text and vision, and qwen3-coder-next for code
and reasoning. Whisper, BGE-M3 and Kokoro remain included through SovAPI.
Direct model call
Gemma 4 31B
Text, vision and structured extraction.
curl https://api.sovinfra.ai/v1/chat/completions \
-H "Authorization: Bearer $SOVINFRA_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"gemma-4-31b","messages":[{"role":"user","content":"Summarize this request."}]}'Direct model call
Qwen3-Coder-Next
Code, reasoning and technical generation.
curl https://api.sovinfra.ai/v1/chat/completions \
-H "Authorization: Bearer $SOVINFRA_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"qwen3-coder-next","messages":[{"role":"user","content":"Write a Python health check."}]}'from openai import OpenAI
client = OpenAI(
base_url="https://api.sovinfra.ai/v1",
api_key="sk-sov-...",
)
response = client.chat.completions.create(
model="sovapi",
messages=[{"role": "user", "content": "Explain SovAPI in one sentence."}],
)import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.sovinfra.ai/v1",
apiKey: process.env.SOVINFRA_API_KEY,
});Compliance and control
Private inference, models hosted in Europe, server-side secrets, and logs without prompts or keys. Legal notices and terms are available from the footer.