DeepSeek · open weights
DeepSeek V4 Flash API
The cheapest model we serve and the one plan sizes are quoted in — MIT-licensed, 304B parameters, within a couple of points of Opus 4.8 on agentic benchmarks.
~40% below OpenRouter list for the same model. Prices verified August 14, 2026. Full market table on /pricing.
Call it
Model id: deepseek-ai/DeepSeek-V4-Flash-0731 (case-sensitive). Any OpenAI SDK works.
from openai import OpenAI
client = OpenAI(
base_url="https://router.mingles.ai/v1",
api_key="sk-your-free-key",
)
resp = client.chat.completions.create(
model="deepseek-ai/DeepSeek-V4-Flash-0731",
messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)Published benchmarks
| Benchmark | DeepSeek-V4-Flash | Claude Opus 4.8 | GLM-5.2 |
|---|---|---|---|
| Terminal-Bench 2.1 | 82.7 | 85 | 81 |
| NL2Repo | 54.2 | 69.7 | 48.9 |
| Cybergym | 76.7 | 83.1 | — |
| DeepSWE | 54.4 | 58 | 46.2 |
| Toolathlon-Verified | 70.3 | 76.2 | 59.9 |
| Agents' Last Exam | 25.2 | 25.7 | 23.8 |
| AutomationBench (public) | 25.1 | 27.2 | 12.9 |
| DSBench-FullStack | 68.7 | 71.6 | 61.8 |
| DSBench-Hard | 59.6 | 71.7 | 54.5 |
Source: DeepSeek — DeepSeek-V4-Flash-0731 model card. Full comparison with sources on /compare/claude-api.
Specs
| Architecture | 304B total parameters with a speculative-decoding module; the card states a "far smaller activated parameter count" than V4-Pro but does not publish the number |
| Reasoning effort | Three levels — low, high, max — selectable per request (per model card) |
| Strengths | Agentic and tool-use workloads: Terminal-Bench 2.1 82.7, Toolathlon-Verified 70.3 |
| License | MIT |
| Availability | Free tier and every paid plan — served by the free Gonka primary |
On a plan
Plans hold a priced allowance, not a token count, so what it buys depends on the model. On DeepSeek V4 Flash it works out to roughly:
V4-Flash is the volume model the rest of the catalogue is now measured against. On DeepSeek's own like-for-like table it lands within about two points of Claude Opus 4.8 on Terminal-Bench 2.1 (82.7 vs 85.0), Agents' Last Exam (25.2 vs 25.7) and AutomationBench (25.1 vs 27.2), and ahead of GLM-5.2 on every row both models ran. It is further behind on the two hardest repo-scale rows — NL2Repo 54.2 against 69.7, DSBench-Hard 59.6 against 71.7 — so the gap is real where a task spans a whole codebase. At a blended $0.10 per 1M tokens it costs about a third of MiniMax and a fifth of Kimi, which is what makes "a couple of points behind Opus" the interesting sentence rather than an academic one.
FAQ
▸How do I call DeepSeek V4 Flash through the API?
POST /v1/chat/completions with "model": "deepseek-ai/DeepSeek-V4-Flash-0731" (case-sensitive) against https://router.mingles.ai/v1 — any OpenAI SDK works after changing base_url. Streaming, tool-calling and reasoning_effort are all supported.
▸What does DeepSeek V4 Flash cost?
$0.084 per 1M input tokens and $0.168 per 1M output (~40% below OpenRouter list) — the cheapest model we bill, and the one every plan's "up to N tokens" figure is counted in. On the $9 Mini plan the allowance covers roughly 220M tokens a month.
▸Is it on the free tier?
Yes. It is served by the free Gonka primary, so a free key can call it with no card. That is what separates it from GLM-5.2, which is cheaper per token on paper but only reachable through a paid upstream.
▸DeepSeek, MiniMax or Kimi?
DeepSeek V4 Flash spends the allowance at ×1 — it is the baseline. MiniMax-M2.7 runs about ×2.9 and Kimi-K2.6 about ×5.7, so reach for them when a task needs the depth and your evals show it. Most workloads end up mixed: cheap first pass here, escalate where quality shows.
▸What does reasoning_effort do?
The card documents three levels — low, high and max — controlling how much the model deliberates before answering. Pass it as a normal OpenAI-style request field; higher effort costs more output tokens, which come out of the same allowance.