Mercury 3 Hits Vercel AI Gateway at Launch Pricing

Summary: Inception’s Mercury 3 is now callable through Vercel AI Gateway as inception/mercury-3, with 50% off launch pricing. The model generates text via diffusion — parallel refinement rather than token-by-token decoding — and Inception reports generation speeds above 1,000 tokens per second. If you run latency-sensitive agents, that combination is worth a benchmark run this week.

Why a diffusion LLM changes the latency math

Most language models produce tokens sequentially: each token depends on the one before it. Mercury 3 instead refines whole spans of text in parallel, which is why Inception reports speeds above 1,000 tokens per second. Vercel’s changelog positions the model for workloads where response time is the product: browser agents, real-time search, interactive education, and customer support.

The other spec that matters is the 1M-token context window. Long documents and extended conversation histories fit without chunking, which simplifies the architecture of support agents that need to see an entire ticket history. As I covered when Muse arrived on AI Gateway, the practical question is never just “is this model good” but “how cheaply can I route to it and swap away from it” — and a gateway listing answers the second half by default.

Wiring it up takes minutes

If you already use the AI SDK, the change is a model string:

import { streamText } from 'ai';

const result = streamText({
  model: 'inception/mercury-3',
  prompt: 'Write a brief, friendly reply to a customer whose delivery is late.',
});

for await (const text of result.textStream) {
  process.stdout.write(text);
}

Set AI_GATEWAY_API_KEY in your environment and you’re done. To point coding agents like Claude Code, Codex, or Cursor at it, run:

npx vercel ai-gateway setup

Then select inception/mercury-3 in the agent.

What the gateway adds on top

Calling Mercury 3 through AI Gateway means you inherit the operational layer: request logs with latency and cost per call, budgets scoped to teams or API keys, retries, and provider failover. Vercel states the gateway passes through provider pricing with no markup and no platform fee on inference, including Bring Your Own Key requests — so the 50% launch discount flows to you directly.

Failover is the underrated piece here. If diffusion-based generation doesn’t hold up on your evals, you can fall back to an autoregressive model in routing rules without touching application code. That lowers the cost of experimenting with a genuinely different generation approach.

What to verify before committing

The changelog says Inception reports strong benchmark results, but per the supplied announcement it doesn’t name which benchmarks or scores — so treat speed and context as the confirmed specs and run your own evals for reasoning quality. Two practical tests: measure end-to-end latency on your real prompt shapes (tokens per second is a model-level number, not a p95), and check how parallel refinement behaves on long, structured outputs where coherence errors are costly. If Mercury 3 clears those bars for a latency-sensitive path, the gateway makes trying it essentially free beyond the tokens.

Sources

Original post

View the original post on X