How many tokens
does $100 buy?
The earlier $1 million comparison, divided by 10,000. Each row spends the full $100 on one model. All figures count generated output tokens; M means million.
| Model | API $/million input / output | API: output only input costs excluded | API: 8:1 input : output | Self-hosting: 8:1 input : output | GPU setup / quantization |
|---|---|---|---|---|---|
| Qwen3.8 Max | $2 / $6 | 16.67M | 4.55M | 17.09M | 24 × GB300NVFP4 · SGLang · forced acceptance 3.3$10/GPU-hour · 0.42 cluster hours |
| Qwen3.8 27B | $0.5 / $3 | 33.33M | 14.29M | 184.44M | 1 × B200NVFP4 + FP8 attention/KV · vLLM (third-party run)$7/GPU-hour · 14.29 cluster hours |
| Kimi K3 | $3 / $15 | 6.67M | 2.56M | 9.94M | 8 × B300Native MXFP4 + DSpark · SGLang$8/GPU-hour · 1.56 cluster hours |
| GLM 5.3 | $1.4 / $4.4 | 22.73M | 6.41M | 34.15M | 4 × GB300NVFP4 experts · SGLang$10/GPU-hour · 2.50 cluster hours |
| DeepSeek V4 Pro 0813 | $0.66 / $1.98 | 50.51M | 13.77M | 29.90M | 4 × GB300Native mixed FP4/FP8 · SGLang$10/GPU-hour · 2.50 cluster hours |
| DeepSeek V4.1 Flash | $0.15 / $0.6 | 166.67M | 55.56M | — | No matching serving benchmark in the original audit. |
The $100 GPU caveat. These are linear estimates at sustained benchmark throughput, not guaranteed purchasable runs. Model loading, idle time, minimum rental periods and cluster availability can dominate a small budget. For Qwen3.8 Max, $100 buys only 25 minutes on the assumed 24-GPU cluster, before any setup overhead.
What stays the same
- API assumptions: eight uncached input tokens per generated output token in the matched column. The output-only column ignores input costs. Billed output includes reasoning tokens where applicable. Qwen uses the original Alibaba International/Singapore rates.
- DeepSeek: original off-peak API rates; peak pricing in that snapshot halves the API token totals. The Pro row is retained as a historical comparison, not confirmation that its API endpoint still serves that model.
- GPU rental assumptions: $7/B200-hour, $8/B300-hour and a hypothetical $10/GB300-hour. No taxes, storage, setup, operations or idle costs; 100% productive time. These are extrapolations from reported runs, not benchmarks we ran.
- Benchmark qualifications: Qwen3.8 Max uses forced speculative acceptance of 3.3, with 20 active GPUs but 24 billed; this is not a validated ordinary-generation throughput result. Qwen3.8 27B is a third-party vLLM measurement. Kimi, GLM and DeepSeek Pro use SGLang-reported runs. Different concurrency levels mean these are throughput comparisons, not matched latency tests.
- Output, not total tokens: total-token throughput is converted to output throughput by dividing by nine for the 8:1 workloads. The Qwen 27B result already reports output tokens. Flash remains blank because the original audit had no matching verified serving benchmark.
Arithmetic and original sources
API output tokens = 100 × 1,000,000 ÷ (output price + 8 × input price)
GPU output tokens = 100 × 3,600 × output tokens/second/billed GPU ÷ dollars/GPU-hour
The linked sources are inherited from the original report and have not been rechecked for this budget conversion.
- Qwen3.8 Max: API pricing · Serving benchmark
- Qwen3.8 27B: API pricing · Serving benchmark
- Kimi K3: API pricing · Serving benchmark
- GLM 5.3: API pricing · Serving benchmark
- DeepSeek V4 Pro 0813: API pricing · Serving benchmark
- DeepSeek V4.1 Flash: API pricing · Model page; no verified benchmark