Dragon
Everyday coding
$1.00 in / $6.00 out per 1M tokens
[ MWS API ]
Frontier coding, drop-in.
One API, both Anthropic and OpenAI shapes. Point your baseURL at Dragon and your existing SDK works unchanged. Trillion-parameter models, frontier-level coding, a fraction of proprietary cost.
One-line drop-in
Already using Anthropic or OpenAI? Add a baseURL override and an MWS API key. Your existing tool calls, streaming, and message shapes work unchanged.
import Anthropic from "@anthropic-ai/sdk";
const client = new Anthropic({
baseURL: "https://api.mws.run/v1",
apiKey: process.env.MWS_API_KEY,
});
const response = await client.messages.create({
model: "claude-sonnet-4-6",
max_tokens: 1024,
messages: [{ role: "user", content: "Hello" }],
});Models
Four lanes are live today: Dragon for fast, balanced everyday coding, and Max for the hardest tasks, plus Tiny and Large for lighter and heavier everyday work (full pricing for all four below). Pass any Claude or OpenAI model name and we route it to the right lane.
Max
Hardest tasks (/max)
$2.00 in / $6.2857 out per 1M tokens
Rate limits
Rate limits apply per key. A fresh key starts at roughly 8 requests per second with a short burst ceiling, and a 429 carries a Retry-After header so your client can back off cleanly.
Parallel workloads can run on separate keys. If you need more throughput than that, email support@vellora.aiwith what you're building and we'll raise your limits.
Frequently asked
How do I pay?
Pay-as-you-go. Buy credits in dollar amounts ($10 minimum). Credits convert to tokens at a fixed published rate, so you always know the capacity you're buying. We charge via Stripe.
What latency / throughput?
Dragon streams fast enough for interactive agent loops, with sub-second time-to-first-token on typical prompts. Very large requests take longer to start, and Max trades some speed for peak capability on the hardest tasks.
Can I keep my code on Anthropic SDK?
Yes. That's the whole point. The gateway speaks Anthropic's Messages format natively (system, content blocks, tool_use, tool_result). Streaming, tool round-trips, and stop reasons round-trip correctly.
What about prompt caching?
You don't manage it. The gateway handles caching itself; cache_controlblocks are accepted but not honored as-is, so there's nothing to configure. Even setting caching aside, the cost difference vs proprietary models is large enough that most workloads come out ahead.
Do credits expire?
Credits stay valid for at least 12 months from purchase. After that they may be cleaned up. We'll email you before that happens.
Can I get a refund?
Refunds aren't automatic, but we review requests on a case-by-case basis. Email support@vellora.ai with your account details and the reason, and we'll take a look.
Is this stable?
The same gateway has been in production powering Dragon Code and our own workloads for months. Public API access is in early access. Expect occasional incidents and the odd rough edge during the first weeks. Watch live status on the status page, and email support@vellora.ai if anything breaks.
Do I need to create multiple API keys to get more throughput?
You can: limits are per key, so parallel workloads can run on separate keys. Keys are also how you separate environments (prod / staging / sandbox) and revoke access independently. If juggling keys for throughput gets awkward, email support@vellora.aiand we'll raise a single key's limit instead.