DOS AI serves high-quality open-source LLMs and frontier cloud models via an OpenAI-compatible API. Self-hosted models run on dedicated GPU clusters in Singapore (asia-southeast1-b) with NVMe storage and sub-250ms TTFT. Cloud models are served via enterprise partner backends for maximum coverage and high availability.
Use auto (or dos-auto) as the model ID to let the DOS AI Smart Router automatically select the best model for each request. The router classifies incoming prompts into 3 complexity tiers (Simple, Medium, Complex) using task classification, reasoning depth, and context length.
response = client.chat.completions.create(
model="auto", # Smart routing dynamically selects the optimal model
messages=[{"role": "user", "content": "Analyze this multi-step agent workflow..."}],
)When a request requires advanced multi-step reasoning, complex coding, or autonomous agent loops, the Smart Router escalates to the lowest-priority active tier with verified tool calling and reasoning support:
| Priority | Model | Provider / Backend | Context | Input (per 1M) | Output (per 1M) | Strengths |
|---|---|---|---|---|---|---|
| 1 (Primary) | GPT-6 Luna | OpenAI / Azure Foundry | 1.05M | $0.10 | $0.50 | Frontier efficiency, fast tool calling, agent loops |
| 2 | Gemini 3.8 Flash | Google Cloud | 1M | $0.79 | $3.94 | Sub-second TTFT, multimodal, massive context |
| 3 | Qwen 3.8 Max | Alibaba Cloud | 1M | $2.00 | $6.00 | Complex bilingual logic, deep mathematical reasoning |
Note: Simple and Medium tier requests remain on the in-house self-hosted model (dos) to optimize speed and cost.
dos (or dos-ai) is DOS.AI's curated meta-model and intelligent execution endpoint, pre-tuned for DOSClaw agent workflows, ReAct loops, and native tool-calling with zero markup. Requests sent to model: "dos" default to the self-hosted Singapore GPU cluster (qwen3.8-27b at $0.07 input / $0.50 output per 1M tokens).
The platform also provides a Curated Model Pool of pre-selected, cost-efficient engines that can be invoked directly at each engine's exact wholesale list rate:
| Engine | Provider | Context | Input Price (1M) | Output Price (1M) | Best For |
|---|---|---|---|---|---|
| Qwen 3.8 27B Dense | DOS.AI (Singapore GPU) | 256K | $0.07 | $0.50 | In-house low-latency agent loops, sub-250ms TTFT |
| GPT-6 Luna | OpenAI / Azure | 1.05M | $0.10 | $0.50 | Complex reasoning, structured outputs, 1.05M context |
| Gemini 3.8 Flash | Google Cloud | 1M | $0.79 | $3.94 | Ultra-fast multimodal processing, sub-second TTFT |
| DeepSeek V4.1 Flash | DeepSeek / Alibaba | 1M | $0.30 | $1.20 | MoE high-throughput reasoning and coding logic |
Requests to model: "dos" execute on the in-house Singapore GPU cluster with automatic multi-provider circuit breaker failover defined in the Supabase catalog, ensuring 99.99% continuity during upstream provider maintenance.
| Model | Provider | Context | Input Price (1M) | Output Price (1M) | Model ID | Key Capabilities |
|---|---|---|---|---|---|---|
| GPT-6 Luna | OpenAI / Azure | 1.05M | $0.10 | $0.50 | gpt-6-luna |
Frontier reasoning, tools, 1.05M context |
| GPT-6 Sol | OpenAI / Azure | 1.05M | $2.00 | $10.00 | gpt-6-sol |
Frontier flagship, deep architectural reasoning |
| GPT-6 Astra | OpenAI | 1.05M | $5.00 | $25.00 | gpt-6-astra |
Next-gen multistep engineering, async tools |
| Gemini 3.8 Flash | Google Cloud | 1M | $0.79 | $3.94 | gemini-3.8-flash |
Hybrid reasoning, vision, sub-second TTFT |
| DeepSeek V4.1 Flash | Alibaba / DeepSeek | 1M | $0.30 | $1.20 | deepseek-v4.1-flash |
MoE, controllable thinking, 1M context |
| DeepSeek V4 Pro | DeepSeek / Alibaba | 1M | $2.40 | $4.80 | deepseek-v4-pro |
Mathematical reasoning, complex agent workflows |
| Qwen 3.8 Flash | Alibaba Cloud | 1M | $0.15 | $0.47 | qwen3.8-flash |
Ultra-fast lightweight reasoning, high-throughput chat |
| Qwen 3.8 27B | Alibaba Cloud | 1M | $0.50 | $3.00 | qwen3.8-27b |
Dense reasoning, tool calling, explicit caching |
| Qwen 3.8 Max | Alibaba Cloud | 1M | $2.00 | $6.00 | qwen3.8-max |
Multilingual reasoning, complex logic |
| Claude Sonnet 5 | Anthropic | 1M | $3.00 | $15.00 | claude-sonnet-5 |
Enterprise agentic coding, 128K output |
| Claude Opus 5 | Anthropic | 1M | $15.00 | $75.00 | claude-opus-5 |
Flagship reasoning, long-horizon tasks |
All prices are in USD and strictly adhere to DOS AI's Zero-Markup Policy (100% of provider published list price). Check
GET /v1/catalogor the web dashboard for real-time status.
| Model | Provider | Dimensions | Max Input Tokens | Model ID |
|---|---|---|---|---|
| Qwen3-Embedding-4B | Alibaba / Self-hosted | 2560 | 32,768 | qwen3-embedding-4b |
| BGE-Large-EN-v1.5 | Cloudflare / BAAI | 1024 | 512 | baai/bge-large-en-v1.5 |
See the Embeddings API reference for request/response formats and code examples.
DOS.AI supports specialized text-to-image and visual generation models for autonomous agents, MCP marketing tools (generate_marketing_image), and direct visual generation:
| Model | Provider | Resolution / Tiers | Billing Unit | Model ID | Primary Use Case |
|---|---|---|---|---|---|
| Wan 2.2 T2I Flash | Alibaba Cloud | Up to 1024×1024 | Per Image ($0.08) | wan2.2-t2i-flash |
High-speed creative marketing visual generation, banners |
| Gemini Image (Nano Banana 2) | Google Cloud | 1K / 2K | Variable Tokens (Input + Output) | nano-banana-2 |
Multimodal instruction-following, bilingual marketing copy |
| Nano Banana Pro | Google Cloud | 1K / 2K | Variable Tokens (Input + Output) | nano-banana-pro |
High-fidelity photorealistic rendering and design assets |
Agent Model Priority Rule: In DOSClaw, image generation models can be included in an agent's ordered Model Priority list (
model_chain) to designate preferred visual tool routes. However, an image model cannot be assigned as the Primary Chat Model (model_chain[0]), which strictly requires conversational chat capabilities.
OpenAI's latest cost-efficient frontier reasoning model. Features state-of-the-art instruction following, coding, structured JSON outputs, and native tool-calling agent capabilities at more than 50% lower cost than previous generations ($0.10 in / $0.50 out).
- Best for: Autonomous agent loops, function calling, high-volume production chat, code assistance
- Strengths: 1.05M context window, exceptional price-to-performance, Azure Foundry low-latency hosting
- Model ID:
gpt-6-luna
DOS.AI's curated meta-model pre-tuned for DOSClaw agent workflows, ReAct loops, and low-latency execution. Served from our in-house Singapore GPU cluster (sub-250ms TTFT) with unified zero-markup billing.
- Best for: Agent turn executions, customer support, fast interactive chat, CJK bilingual tasks
- Strengths: Sub-250ms TTFT, 256K context, pre-selected 4-engine curated pool, automatic catalog circuit breakers
- Model ID:
dos,dos-ai
Google's high-speed multimodal reasoning model with hybrid thinking, 1M context window, and sub-second execution.
- Best for: Real-time document parsing, large multimodal payloads, high-throughput applications
- Strengths: 1M context, sub-second TTFT, native audio/video/image comprehension
- Model ID:
gemini-3.8-flash
| Use Case | Recommended Model | Why |
|---|---|---|
| Let DOS AI decide | auto |
Smart Router dynamically matches task complexity and cost |
| Fastest agent execution & lowest latency | dos |
Self-hosted 27B dense cluster in Singapore, sub-250ms TTFT |
| Ultra-low cost frontier reasoning | gpt-6-luna |
$0.10 / $0.50 per million tokens with 1.05M context |
| Sub-second large document analysis | gemini-3.8-flash |
1M context window, Google Cloud high-throughput backend |
| Enterprise agentic coding & architecture | claude-sonnet-5 |
Frontier coding quality with 128K max output tokens |
| Deep mathematical reasoning | deepseek-v4-pro |
SOTA chain-of-thought problem solving |
Every model detail page in the Dashboard features live, rolling operational telemetry across all active provider routes:
| Metric | Description | Source / Contract |
|---|---|---|
| Median Latency (p50) | Median round-trip execution latency | Real observed inference requests over a 7-day rolling window |
| P95 Latency | 95th-percentile response latency | Measures tail latency under peak conditions |
| TTFT (Time-to-First-Token) | Time to first streaming token | Observed on streaming requests across edge locations |
| Throughput (TPS) | Tokens generated per second (tok/s) |
Real measured generation throughput |
| Uptime & Reliability | Success ratio percentage over total attempts | Excludes client-side cancellations and user-side bad requests |
Telemetry Sample Threshold: To maintain rigorous evidence standards, a model route requires at least 20 valid samples before publishing live operational telemetry. Routes with fewer observations report
insufficient_dataorunavailablewith clean dash (-) indicators until warmed up.
DOS AI continuously synchronizes independent benchmark evidence from leading evaluation authorities, attributing each metric with capture dates and direct evaluation links:
- Artificial Analysis: Intelligence Index, Coding Index, and Agentic Index.
- Design Arena: Elo rating, win rate percentage, and leaderboard rank.
Benchmark scores are displayed directly on each model's detail page, complete with normalized visual progress bars (0–100) and direct hyperlinks to public methodology audits.
Retrieve the current list of available models programmatically:
curl https://api.dos.ai/v1/models \
-H "Authorization: Bearer $DOS_API_KEY"For the full retail catalog including pricing, context lengths, and capability flags:
curl https://api.dos.ai/v1/catalog \
-H "Authorization: Bearer $DOS_API_KEY"