OpenAI: gpt-oss-120b
openai/gpt-oss-120b
Cheapest provider
$0.04 / 1M
DeepInfra
Fastest provider (p95)
2152 tok/s
Cerebras
Intelligence (composite)
81.3
MMLU-Pro · HumanEval · math · GPQA
Per-provider performance
Latency / throughput / uptime / price measured across providers over the last 30 minutes of live traffic. This is what proves “sourced cheapest” — Atlas mode draws on these per call to serve the cheapest path that holds quality.
| Provider | Quant | Input $/1M | Output $/1M | Latency p50 / p95 | Throughput p50 / p95 | Uptime 30m | Success |
|---|---|---|---|---|---|---|---|
| DeepInfra | full· bf16 | $0.0390 | $0.1900 | 657ms / 10623ms | 82 / 405.8 tok/s | 98.66% | 59.0% |
| DekaLLM | full· bf16 | $0.0390 | $0.1800 | 509ms / 1530ms | 5 / 39.1 tok/s | 96.09% | 78.9% |
| Novita | q4· fp4 | $0.0500 | $0.2500 | 450ms / 1027ms | 97 / 198 tok/s | 100.00% | 100.0% |
| SiliconFlow | q8· fp8 | $0.0500 | $0.4500 | 1416ms / 4011ms | 17 / 38 tok/s | 92.56% | 92.2% |
| full· unknown | $0.0900 | $0.3600 | 314ms / 833ms | 221 / 658 tok/s | 100.00% | 100.0% | |
| BaseTen | q4· fp4 | $0.1000 | $0.5000 | 185ms / 891ms | 252 / 397 tok/s | 100.00% | 100.0% |
| DigitalOcean | full· unknown | $0.1000 | $0.7000 | 513ms / 1506ms | 127 / 162 tok/s | 99.72% | 99.7% |
| Parasail | q4· fp4 | $0.1000 | $0.7500 | 319ms / 655ms | 252 / 430.2 tok/s | 100.00% | 99.9% |
| SambaNova | full· unknown | $0.1400 | $0.9500 | 486ms / 4269ms | 177 / 430 tok/s | 97.83% | 85.2% |
| Amazon Bedrock | full· unknown | $0.1500 | $0.6000 | — | — | 98.53% | — |
| Amazon Bedrock | full· unknown | $0.1500 | $0.6000 | — | — | — | — |
| DeepInfra | full· bf16 | $0.1500 | $0.6000 | 657ms / 10623ms | 82 / 405.8 tok/s | 98.18% | 59.0% |
| Groq | full· unknown | $0.1500 | $0.6000 | 104ms / 649ms | 443 / 855.4 tok/s | 99.99% | 100.0% |
| Mara | full· unknown | $0.1500 | $0.7500 | 943ms / 2413ms | 89.5 / 460.9 tok/s | 99.10% | 78.6% |
| Nebius | q4· fp4 | $0.1500 | $0.6000 | 207ms / 1221ms | 308 / 489.3 tok/s | 100.00% | 100.0% |
| Phala | full· unknown | $0.1500 | $0.6000 | 1156ms / 2222ms | 91 / 122 tok/s | 100.00% | 100.0% |
| Together | full· unknown | $0.1500 | $0.6000 | 307ms / 1501ms | 62 / 112 tok/s | 99.80% | 99.9% |
| WandB | q4· fp4 | $0.1500 | $0.6000 | 274ms / 565ms | 50 / 114 tok/s | 100.00% | 100.0% |
| Cerebras | full· fp16 | $0.3500 | $0.7500 | 196ms / 652ms | 694 / 2152 tok/s | 99.97% | 99.4% |
“—” means live telemetry hasn’t accumulated enough recent traffic for that endpoint. “undisclosed” means the provider serves the model but doesn’t expose the quantization label (typically running fp8 / int8 internally).
Intelligence breakdown
Composite score is a weighted average of public benchmarks (30% MMLU-Pro, 25% code pass@1, 25% math, 20% GPQA). Numbers come from model cards and the Artificial Analysis intelligence harness; missing components are renormalised over what’s present.
MMLU-Pro
—
broad reasoning
Code
71.0
pass@1 (HumanEval / LiveCodeBench)
AIME 2025
92.5
math accuracy
GPQA Diamond
80.1
hard reasoning
Source: OpenAI gpt-oss model card (GPQA-Diamond 80.1, AIME25 92.5 no-tools, HumanEval ~71); arxiv.org/abs/2508.10925
How Atlas mode sources OpenAI: gpt-oss-120b
- Strict mode — pin OpenAI: gpt-oss-120b exactly and we pass it straight through, sourced from the cheapest provider above. The same model, no substitutions — currently DeepInfra at $0.04/1M.
- Atlas mode — the default. Each call is auto-optimized for the cheapest path that holds quality, at least 5% off going direct from call one and climbing as it ramps. You always see which model served the call and exactly what you saved — thumbs-down anything you don’t like for a full refund.