# Working account — 300 tok/s of DeepSeek-V4.1-Flash

Plug your numbers into the braces. The defaults are the ones in the note: always-on node, 730 hours, 4:1 input:output, cache misses, September 2026 list prices, $1 = €0.92.

Do not compare a pair of H100s to GPT-4o. V4.1-Flash does not fit there.

---

## Inputs

- Decode SLO (tok/s, output): {300}
- Hours the node is on this month: {730}
- Input:output ratio: {4}:1
- Cache-hit share of input: {0}
- Share of hours that fall in DeepSeek peak (weekdays 01:00–04:00 and 06:00–10:00 UTC): {0.21}
- FX: $1 = {0.92} €

Derived, with the defaults:

```
output tokens     = 300 × 3,600 × 730           = 788,400,000
input tokens      = 788,400,000 × 4             = 3,153,600,000
```

---

## Node (you pay this whether you emit 300 tok/s or 12,000)

| Config | Source | $/h | $ / 730 h | ≈ € / 730 h | Fits V4.1-Flash? |
|---|---|---|---|---|---|
| 8× H200 | CoreWeave on-demand | 50.44 | 36,821 | 33,875 | Yes (recipe) |
| 4× GB200 NVL4 | CoreWeave on-demand | 42.00 | 30,660 | 28,207 | Yes (verified) |
| 8× B200 | CoreWeave on-demand | 68.80 | 50,224 | 46,206 | Yes (recipe) |
| 8× H100 80 GB | CoreWeave on-demand | 49.24 | 35,945 | 33,069 | **No. 640 GB < 614 GB floor** |
| 2× H100 | — | — | — | — | **No** |
| 8× H200 spot | CoreWeave spot | 20.93 | 15,279 | 14,057 | Hardware yes, production no |

Node you are actually quoting: {8× H200} = {€33,875}

---

## API (you pay this on tokens)

Per 1M tokens, USD, list, September 2026:

| Model | Input (miss) | Input (hit) | Output |
|---|---|---|---|
| DeepSeek Flash, off-peak | 0.15 | 0.003 | 0.60 |
| DeepSeek Flash, peak | 0.30 | 0.006 | 1.20 |
| Claude Sonnet 5 | 2.00 | 0.20 | 10.00 |
| GPT-5.6 Terra | 2.00 | 0.20 | 12.00 |
| GPT-5.6 Sol | 4.00 | 0.40 | 20.00 |
| Claude Opus 5 | 5.00 | 0.50 | 25.00 |

OpenAI EU data-residency endpoints: **+10%** on models from 5 Mar 2026.

Default mix, all misses, DeepSeek off-peak:

```
3,153.6 × $0.15  +  788.4 × $0.60  =  $946   ≈  €870
```

Same mix, DeepSeek peak: **$1,892 ≈ €1,741**.
Blended 21% peak: **≈ $1,145 ≈ €1,053**.

| Route | $ / month | ≈ € / month | × vs DeepSeek off-peak |
|---|---|---|---|
| DeepSeek Flash, off-peak, misses | 946 | 870 | 1.0× |
| DeepSeek Flash, 21% peak | 1,145 | 1,053 | 1.2× |
| Claude Sonnet 5 | 14,191 | 13,056 | 15× |
| GPT-5.6 Terra | 15,768 | 14,507 | 17× |
| GPT-5.6 Terra, EU residency | 17,345 | 15,957 | 18× |
| GPT-5.6 Sol | 28,382 | 26,112 | 30× |
| Claude Opus 5 | 35,478 | 32,640 | 38× |
| 8× H200 on-demand | 36,821 | 33,875 | 39× |

---

## Crossover (when the node is cheaper than DeepSeek)

Off-peak, 4:1 misses, cost per 1M *output* tokens on DeepSeek = `4 × 0.15 + 0.60 = $1.20`.

```
tok/s to break even = 36,821 / (1.20 × 3,600 × 730 / 1e6)
                    = 36,821 / 3.1536
                    ≈ 11,700 tok/s sustained, all month
```

At 300 tok/s you are at **~2.6%** of that line. The GPUs are paid. They are not busy.

---

## Decision

- [ ] Tokens may leave for DeepSeek / OpenAI / Anthropic → **do not buy the node for this volume**.
- [ ] Tokens must not leave, and 300 tok/s is the product SLO → **buy a node that fits (8× H200 or 4× Blackwell). Not two H100s.**
- [ ] You need ~12k tok/s sustained on this mix → the node can win on money. Measure before you sign.

Sources: [DeepSeek pricing](https://api-docs.deepseek.com/quick_start/pricing), [vLLM recipe](https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4.1-Flash), [CoreWeave pricing](https://www.coreweave.com/pricing), [OpenAI pricing](https://developers.openai.com/api/docs/pricing), [Anthropic pricing](https://www.anthropic.com/pricing).