When we published our own concurrency-1 numbers for Nemotron 3 Ultra in May, we made a point of measuring exactly the way Artificial Analysis measures — same settings, same tiktoken counting basis — so the results would be comparable and checkable. Now the check has happened. Artificial Analysis benchmarks every API provider serving NVIDIA Nemotron 3 Ultra on identical methodology, and in the July 2026 snapshot, BLACKBOX AI holds the #1 spot for output speed: 454.4 tokens per second, 47% ahead of Nebius at #2 — while charging $0.44 per 1M tokens to Nebius's $1.20.
We built a page that puts the result in motion — the same prompt streamed side by side against Nebius, plus the full provider leaderboard: blackbox.ai/nemotron.
The leaderboard
Artificial Analysis runs the same benchmark against every provider: identical prompts, identical sampling settings, and tokens counted in the same tiktoken encoding across the board — so a tokens-per-second figure from one provider is directly comparable to any other. No provider supplies its own numbers.
Nemotron 3 Ultra, tokens per second. Higher is better.
| Model | t/s |
|---|---|
| BLACKBOX AI | 454.4 t/s |
| Nebius | 309.3 t/s |
| CoreWeave | 249.6 t/s |
| Together AI | 159.5 t/s |
| DeepInfra | 99.2 t/s |
| Lightning AI | 72.8 t/s |
Source: Artificial Analysis, July 2026.
Speed is only half the story. The same table shows the price column: at $0.44 per 1M blended tokens, BLACKBOX AI is 2.7× cheaper than the #2 provider it outruns by 47%. Nemotron 3 Ultra is a reasoning model, so time to first answer includes thinking tokens — among the high-throughput providers, BLACKBOX AI answers soonest as well.
Why independent measurement matters
Every provider claims to be fast. Self-reported benchmarks — including ours — deserve skepticism, because the vendor controls the prompt mix, the tokenizer, and which runs make the chart. A third-party leaderboard removes all of that: one methodology, applied identically to everyone, published in the open. That's why we benchmarked against the Artificial Analysis protocol from day one, and why this ranking is the one we care about.
Same weights, different engine
Every provider on that table serves the identical open-weights model. The gap is entirely in the serving stack:
We generate tokens, not resell them. Direct access to open weights lets us self-host on our own infrastructure, with custom CUDA kernels tuned for exactly this architecture.
FP4 on NVIDIA B300. Advanced quantization on latest-generation hardware — on the same bench, FP4 leads BF16 on every axis with no measurable quality loss on evals.
Multi-token prediction. Nemotron ships a built-in MTP head; our engine exploits it for speculative decoding, emitting multiple tokens per forward pass instead of one.
Encrypted end to end. Speed without the trade-off — prompts and completions are encrypted in transit and at the edge, never logged, retained, or used for training.
The full engineering story — the architecture, the benchmark harness, and why a 550B-parameter model can decode this fast — is in Nemotron on BLACKBOX.
Run it
The BLACKBOX API is OpenAI-compatible: point your SDK at https://api.blackbox.ai/v1, set the model to blackboxai/nvidia/nemotron-3-ultra, and every token arrives sooner. Streaming, function calling, and JSON mode work unchanged.
Or skip straight to the API and measure it on your own prompts — the leaderboard is only 47% of the fun.