Most AI chatbots write the way you read this sentence — one word after another. A startup called Inception just launched Mercury 3, its flagship model, and it works completely differently. Instead of predicting the next word, Mercury 3 starts with a rough answer and refines the whole thing at once, using a technique called text diffusion. And it's fast: Inception claims Mercury 3 generates 1,068 tokens per second on NVIDIA GPUs — four to eight times faster than rival models — while outperforming GPT-6 Luna and Claude Haiku 5.5 across benchmarks.
How fast is "fast"?
1,068 tokens per second means a full paragraph lands in roughly a second, not a slow stream of words. The model also carries a one-million-token context window, up from 260,000 on its predecessor Mercury 2.5. Inception says Mercury 3 is more than 40 times larger than the original Mercury 1 while intelligence across standard benchmarks rose threefold — and frames the launch as proof that the scaling laws behind the last decade of AI progress apply to diffusion models too. That claim, if it holds up independently, is arguably the biggest deal here.
Cheap, open, and ready to plug in
Launch pricing is aggressive: $0.10 per million input tokens and $0.35 per million output tokens during the launch promotion (the list prices double after). Cached input tokens cost 90% less. Mercury 3 ships through the Inception API plus OpenRouter, Baseten, and Vercel, with an OpenAI-compatible API and Python/TypeScript SDKs — so teams can swap it in without rewriting their stack. New accounts get 100 million free tokens, no payment details required. One housekeeping note: Mercury 2 and 2.5 are being retired for new users, with existing customers keeping access until November 15.
Speed is becoming the real battleground
Inception's early-access partners are all latency-obsessed. In the company's production web-agent harness, Mercury 3 completed 50% of 50 live browser-navigation tasks — more than GPT-6.1-Sol managed with the same token budget, at a fraction of the cost per token. Tutoring platform RevisionDojo says students get hints and marked answers in about a second. Smart-glasses startup Octen AI measured a 0.9-second median generation time; its CEO's line sums up the whole pitch: "an answer that takes five seconds feels broken." It's the same race DeepSeek is running — its new V4.1-Flash-1017 also targets shorter responses and lower task times — and the same portfolio logic behind Mercury Decide, Inception's decision model, which it says outperforms Jev — the decision model we covered back in September — on JevBench and LegalBench.
The take
The giants are racing on capability; the upstarts are racing on latency. For AI agents — the thing everyone is building right now — a model that's fast, cheap, and good enough is often worth more than one that's brilliant and slow. Two caveats, stated plainly: the benchmark wins are Inception's own numbers, and independent verification is still pending. But the direction is unmistakable — the next frontier in AI isn't just what models know, it's how fast they can think.
Comments
Post a Comment